Documenti di Didattica
Documenti di Professioni
Documenti di Cultura
Workshop organizers make last-minute changes to their schedule. West 217 - 219, AI for Humanitarian Assistance and Disaster
Download this document again to get the lastest changes, or use the Response Gupta, Murphy, Darrell, Heim, Wang, Goodman, Bili■ski
NeurIPS mobile application.
West 220 - 222, Shared Visual Representations in Human and
Schedule Highlights Machine Intelligence Deza, Peterson, Murty, Griffiths
Page 1 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
West 220 - 222, Robot Learning: Control and Interaction in the Real
World Calandra, Rakelly, Kamthe, Kragic, Schaal, Wulfmeier
West Ballroom C, “Do the right thing”: machine learning and causal
inference for improved decision making Santacatterina,
Joachims, Kallus, Swaminathan, Sontag, Zhou
Page 2 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
The last decade has seen both machine learning and biology
transformed: the former by the ability to train complex predictors on
massive labelled data sets; the latter by the ability to perturb and
measure biological systems with staggering throughput, breadth, and
resolution. However, fundamentally new ideas in machine learning are
needed to translate biomedical data at scale into a mechanistic
understanding of biology and disease at a level of abstraction beyond
Page 3 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
single genes. This challenge has the potential to drive the next decade of Abstracts (10):
creativity in machine learning as the field grapples with how to move
beyond prediction to a regime that broadly catalyzes and accelerates Abstract 1: Opening Remarks in Learning Meaningful
scientific discovery. Representations of Life, Yeshwant 08:45 AM
To seize this opportunity, we will bring together current and future Opening remarks by Francis Collins (Director, National Institutes of
leaders within each field to introduce the next generation of machine Health) via video and Krishna Yeshwant, General Partner at Google
learning specialists to the next generation of biological problems. Our Ventures.
full-day workshop will start a deeper dialogue with the goal of Learning
Abstract 2: Keynote - Bio in Learning Meaningful Representations of
Meaningful Representations of Life (LMRL), emphasizing interpretable
Life, Regev 09:00 AM
representation learning of structure and principles. The workshop will
address this challenge at five layers of biological abstraction (genome,
Aviv Regev. Professor of Biology; Core Member, Broad Institute;
molecule, cell, system, phenome) through interactive breakout sessions
Investigator, Howard Hughes Medical Institute. Aviv Regev pioneers the
led by a diverse team of experimentalists and computational scientists to
use of single-cell genomics and other techniques to dissect the
facilitate substantive discussion.
molecular networks that regulate genes, define cells and tissues, and
influence health and disease.
We are calling for short abstracts from computer scientists and biological
scientists. Submission deadline is Friday, September 20. Significant Abstract 3: Keynote - ML in Learning Meaningful Representations of
travel support is also available. Details here: Life, Welling 09:30 AM
12:00 PM Synthetic Systems Silver, Marks, Liu Abstract 7: Synthetic Systems in Learning Meaningful
Representations of Life, Silver, Marks, Liu 12:00 PM
12:30 PM GWAS Discussion Wang, D'Amour
Pamela Silver, Debora Marks, and Chang Liu in conversation.
HaCohen, Reshef, Johnson,
Morris, Nagy, Eraslan, Singer, Abstract 8: GWAS Discussion in Learning Meaningful
Van Allen, Krishnaswamy, Representations of Life, Wang, D'Amour 12:30 PM
01:30 PM Phenotype
Greene, Linderman,
Bloemendal, Wiltschko, Yixin Wang and Alex D'Amour in conversation.
Kotliar, Zou, Bulik-Sullivan
Abstract 9: Phenotype in Learning Meaningful Representations of
03:15 PM Coffee Break
Life, HaCohen, Reshef, Johnson, Morris, Nagy, Eraslan, Singer, Van
Carpenter, Zhou, Chikina, Allen, Krishnaswamy, Greene, Linderman, Bloemendal, Wiltschko,
Tong, Lengerich, Kotliar, Zou, Bulik-Sullivan 01:30 PM
Abdelkareem, Eraslan,
Blumberg, Ra, Burkhardt, Nir Hacohen, David Reshef, Matt Johnson, Samantha Morris, Aurel
03:30 PM Cell Nagy, Gokcen Eraslan, Meromit Singer, Eli van Allen, Smita
Matsen IV, Moses, Chen,
Haghighi, Lu, Schau, Nivala, Krishnaswamy, Casey Greene, Scott Linderman, Alex Bloemendal, Alex
Shiffman, Harbrecht, Masengo Wiltschko, Dylan Kotliar, James Zou, and Brendan Bulik-Sullivan
Wa Umba participate.
05:00 PM Closing Remarks Sander, Fiete, Peer Abstract 11: Cell in Learning Meaningful Representations of Life,
Carpenter, Zhou, Chikina, Tong, Lengerich, Abdelkareem, Eraslan,
06:00 PM Posters and Social Hour
Blumberg, Ra, Burkhardt, Matsen IV, Moses, Chen, Haghighi, Lu, Schau,
Nivala, Shiffman, Harbrecht, Masengo Wa Umba 03:30 PM
Page 4 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Anne Carpenter, Hui Ting Grace Yeo, Jian Zhou, Maria Chikina,
, Kroshnin, Delalande, Carr,
Alexander Tong, Benjamin Lengerich, Aly O. Abdelkareem, Gokcen
Tompkins, Pooladian, Robert,
Eraslan, Andrew Blumberg, Stephen Ra, Daniel Burkhardt, Emanuel
Makkuva, Genevay, Liu, Zeng,
Flores Bautista, Frederick Matsen, Alan Moses, Zhenghao Chen,
Frogner, Cazelles, Tabak,
Marzieh Haghighi, Alex Lu, Geoffrey Schau, Jeff Nivala, Luke O'Connor,
Ramos, PATY, Balikas,
Miriam Shiffman, Hannes Harbrecht and Shimbi Masengo Wa Umba
05:20 PM Poster Session Trigila, Wang, Mahler,
Papa Levi present in a lightning round.
Nielsen, Lounici, Swanson,
Bhutani, Bréchet, Indyk,
Abstract 12: Closing Remarks in Learning Meaningful
cohen, Jegelka, Wu, Sejourne,
Representations of Life, Sander, Fiete, Peer 05:00 PM
Manole, zhao, Wang, Wang,
Chris Sander, Ila Fiete, and Dana Pe'er present. Dukler, Wang, Dong
10:30 AM Stefanie Jegelka Jegelka This workshop hopes to bring together researchers from different
disciplines, identify common grounds, and spur discussion on how
11:10 AM SPOTLIGHTS 5 x 10 information theory can apply to and benefit from modern machine
12:00 PM Poster Session learning setups.
Page 5 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Our plan is to run this workshop annually co-located with one ML venue
02:00 PM Invited Talk: Varun Jog Jog
and one Systems venue, to help build a strong community which we
02:30 PM Invited Talk: Jelani Nelson think will complement newer conferences like SysML targeting research
at the intersection of systems and machine learning. We believe this dual
Invited Talk: Irena
03:30 PM approach will help to create a low barrier to participation for both
Fischer-Hwang
communities.
Flamich, Ubaru, Zheng,
Djolonga, Wickstrøm, This workshop is part two of a two-part series with one day focusing on
Granziol, Pitas, Li, Williamson, ML for Systems and the other on Systems for ML. Although the two
Yoon, Lee, Zilly, Petrini, workshops are being led by different organizers, we are coordinating our
Fischer, Dong, Alemi, Nguyen, call for papers to ensure that the workshops complement each other and
Brekelmans, Wu, Mahajan, Li, that submitted papers are routed to the appropriate venue.
Shiragur, Carmon, Adilova,
LIU, An, Dash, Gunluk,
Mazumdar, Motani,
Rosenzweig, Kamp, Havasi, Schedule
Barnes, Zhou, Hao, Foster,
Benjamini, Srebro,
08:30 AM Welcome
04:10 PM Poster Session Tschannen, Rubenstein,
Gelly, Duchi, Sidford, Ru, Keynote 1: Machine
Zohren, Dalal, Osborne, Learning Reproducibility:
Roberts, Charikar, An update from the NeurIPS
Subramanian, Fan, 08:40 AM 2019 Reproducibility
Schwarzer, Roberts, Co-Chairs, Joelle Pineau,
Lacoste-Julien, Prabhu, McGill University and
Galstyan, Ver Steeg, Sankar, Facebook
Noh, Dasarathy, Park,
Contributed Talk: SLIDE :
Cheung, Tran, Yang, Poole,
Training Deep Neural
Censi, Sylvain, Hjelm, Liu,
09:10 AM Networks with Large
Gallego, Sypherd, Yang,
Outputs on a CPU faster
Morshuis
than a V100-GPU
This workshop will follow the successful model we have previously run at 12:00 PM Lunch
ICML, NeurIPS and SOSP.
Page 6 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
05:20 PM Closing
Perception as generative reasoning: structure, causality, Igor Rubinov, Risi Kondor, Jack Poulson, Manfred K. Warmuth,
probability Emanuel Moss, Alexa Hagerty
Dan Rosenbaum, Marta Garnelo, Peter Battaglia, Kelsey Allen, Ilker East Meeting Rooms 8 + 15, Fri Dec 13, 08:00 AM
Yildirim
When researchers and practitioners, as well as policy makers and the
East Meeting Rooms 1 - 3, Fri Dec 13, 08:00 AM public, discuss the impacts of deep learning systems, they draw upon
multiple conceptual frames that do not sit easily beside each other.
Many perception tasks can be cast as ‘inverse problems’ where the input Questions of algorithmic fairness arise from a set of concerns that are
signal is the outcome of a causal process and perception is to invert that similar, but not identical, to those that circulate around AI safety, which in
process. For example in visual object perception, the image is caused by turn overlap with, but are distinct from, the questions that motivate work
an object and perception is to infer which object gave rise to that image. on AI ethics, and so on. Robust bodies of research on privacy, security,
Following an analysis-by-synthesis approach, modelling the forward and transparency, accountability, interpretability, explainability, and opacity
causal direction of the data generation process is a natural way to are also incorporated into each of these frames and conversations in
capture the underlying scene structure, which typically leads to broader variable ways. These frames reveal gaps that persist across both highly
generalisation and better sample efficiency. Such a forward model can technical and socially embedded approaches, and yet collaboration
be applied to solve the inverse problem (inferring the scene structure across these gaps has proven challenging.
from an input image) using Bayes rule, for example. This workflow
stands in contrast to common approaches in deep learning, where Fairness, Ethics, and Safety in AI each draw upon different disciplinary
typically one first defines a task, and then optimises a deep model prerogatives, variously centering applied mathematics, analytic
end-to-end to solve it. In this workshop we propose to revisit ideas from philosophy, behavioral sciences, legal studies, and the social sciences in
the generative approach and advocate for learning-based ways that make conversation between these frames fraught with
analysis-by-synthesis methods for perception and inference. In addition, misunderstandings. These misunderstandings arise from a high degree
we pose the question of how ideas from these research areas can be of linguistic slippage between different frames, and reveal the epistemic
combined with and complement modern deep learning practices. fractures that undermine valuable synergy and productive collaboration.
This workshop focuses on ways to translate between these ongoing
Schedule
efforts and bring them into necessary conversation in order to
understand the profound impacts of algorithmic systems in society.
Rosenbaum, Garnelo,
08:50 AM Opening Remarks Schedule
Battaglia, Allen, Yildirim
Page 7 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Christian, Hu, Kondor, The risks and benefits of AI are unevenly distributed within societies and
04:30 PM Solutions
Marshall, Rogers, Schuur across the globe. Governance regimes are drastically different in various
regions of the world, as are the political and ethical implications of AI
technologies.
Abstracts (5): • How do we better understand how AI technologies operate around the
world and the range of risks they carry for different societies?
Abstract 3: Approaches to Understanding AI in Minding the Gap: • Are there global claims about the implications of AI that can apply
Between Fairness and Ethics, Bengio, Dobbe, Elish, Kroll, Metcalf everywhere around the globe? If so, what are they?
08:45 AM • What can we learn from AI’s impacts on labor, environment, public
health and agriculture in diverse settings?
The stakes of AI certainly alter how we relate to each other as humans -
how we know what we know about reality, how we communicate, how Abstract 12: Solutions in Minding the Gap: Between Fairness and
we work and earn money, and about how we think of ourselves as Ethics, Christian, Hu, Kondor, Marshall, Rogers, Schuur 04:30 PM
human. But in grappling with these changing relations, three fairly
concrete approaches have dominated the conversation: ethics, fairness, With the recognition that there are no fully sufficient steps that can be
and safety. These approaches come from very different academic taken to addressing all AI impacts, there are concrete things that ought
backgrounds, draw attention to very different aspects of AI, and imagine to be done, ranging across technical, socio-technical, and legal or
very different problems and solutions as relevant, leading us to ask: regulatory possibilities.
• What are the commonalities and differences between ethics, fairness, • What are the technical, social, and/or regulatory solutions that are
and safety as approaches to addressing the challenges of AI? necessary to address the riskiest aspects of AI?
• How do these approaches imagine different problems and solutions for • What are key approaches to minimize the risks of AI technologies?
the challenges posed by AI?
• How can these approaches work together, or are there some areas
where they are mutually incompatible? KR2ML - Knowledge Representation and Reasoning Meets
Machine Learning
Abstract 6: Detecting and Documenting AI Impacts in Minding the
Gap: Between Fairness and Ethics, Christian, Hagerty, Rogers,
Veronika Thost, Christian Muise, Kartik Talamadupula, Sameer
Schuur, Snow 10:30 AM
Singh, Chris Ré
Algorithmic systems are being widely used in key social institutions and
West 109 + 110, Fri Dec 13, 08:00 AM
while they promise radical improvements in fields from public health to
energy allocation, they also raises troubling issues of bias, Machine learning (ML) has seen a tremendous amount of recent success
discrimination, and “automated inequality.” They also present and has been applied in a variety of applications. However, it comes with
irresolvable challenges related to the dual-use nature of these several drawbacks, such as the need for large amounts of training data
technologies, secondary effects that are difficult to anticipate, and alter and the lack of explainability and verifiability of the results. In many
power relations between individuals, companies, and governments. domains, there is structured knowledge (e.g., from electronic health
• How should we delimit the scope of AI impacts? What can properly be records, laws, clinical guidelines, or common sense knowledge) which
considered an AI impact, as opposed to an impact arising from some can be leveraged for reasoning in an informed way (i.e., including the
other cause? information encoded in the knowledge representation itself) in order to
• How do we detect and document the social impacts of AI? obtain high quality answers. Symbolic approaches for knowledge
• What tools, processes, and institutions ought to be involved in representation and reasoning (KRR) are less prominent today - mainly
addressing these questions? due to their lack of scalability - but their strength lies in the verifiable and
interpretable reasoning that can be accomplished. The KR2ML workshop
Abstract 7: Responsibilities in Minding the Gap: Between Fairness
aims at the intersection of these two subfields of AI. It will shine a light on
and Ethics, Chowdhury, Kim, O'Sullivan, Schuur, Smart 11:30 AM
the synergies that (could/should) exist between KRR and ML, and will
initiate a discussion about the key challenges in the field.
While there is a great deal of AI research happening in academic
settings, much of that work is operationalized within corporate contexts.
Schedule
Page 8 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 9 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
10:30 AM Invited talk #4 West 116 + 117, Fri Dec 13, 08:00 AM
Panel discussing how to
https://nips.cc/Conferences/2019/CallForCompetitions
increase transparency and
10:50 AM
dissemination of ‘soft
Schedule
knowledge’ in ML
Page 10 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
* Prize ceremony.
An approach that has the potential to address a number of problems in
Abstract 8: MicroNet Challenge in Competition Track Day 1, Gale, this space is federated learning (FL). FL is an ML setting where many
Wang, Leng, Cheng, Wang, Elsen, Yan 02:15 PM clients (e.g., mobile devices or whole organizations) collaboratively train
a model under the orchestration of a central server (e.g., service
Trevor Gale and Erich Elsen. Introduction to the competition and provider), while keeping the training data decentralized. Organizations
overview of results. and mobile devices have access to increasing amounts of sensitive data,
with scrutiny of ML privacy and data handling practices increasing
correspondingly. These trends have produced significant interest in FL,
Peisong Wang, Cong Leng, and Jian Cheng. An Empirical Study of since it provides a viable path to state-of-the-art ML without the need for
Network Compression for Image Classification. the centralized collection of training data – and the risks and
responsibilities that come with such centralization. Nevertheless,
significant challenges remain open in the FL setting, the solution of which
Trevor Gale and Erich Elsen. Highlights of other notable entries. will require novel techniques from multiple fields, as well as improved
open-source tooling for both FL research and real-world deployment
Zhongxia Yan and Hanrui Wang. Efficient Memory-Augmented This workshop aims to bring together academic researchers and industry
Language Models with Network Compression practitioners with common interests in this domain. For industry
participants, we intend to create a forum to communicate what kind of
problems are practically relevant. For academic participants, we hope to
Trevor Gale and Erich Elsen. Updates and improvements for the 2020 make it easier to become productive in this area. Overall, the workshop
MicroNet Challenge. will provide an opportunity to share the most recent and innovative work
in FL, and discuss open problems and relevant approaches. The
Abstract 9: Reconnaissance Blind Chess competition in Competition
technical issues encouraged to be submitted include general
Track Day 1, Llorens, Gardner, Perrotta, Highley, Clark, Perrotta,
computation based on decentralized data (i.e., not only machine
Bernardoni, Jordan, Wang 03:15 PM
learning), and how such computations can be combined with other
research areas, such as differential privacy, secure multi-party
* Chair: I-Jeng Wang
computation, computational efficiency, coding theory, etc. Contributions
* Competition and Game Overview (Ashley Llorens)
in theory as well as applications are welcome, including proposals for
* Challenges of the Game (Ryan Gardner)
novel system design. Work on fully-decentralized (peer-to-peer) learning
* Competition Results (Casey Richardson)
will also be considered, as there is significant overlap in both interest and
* Overview of the StrangeFish Bot (Gino Perrotta and Robert Perrotta)
techniques with federated learning.
* Overview of the LaSalle Bot (T.J. Highley)
* Overview of the penumbra Bot (Gregory Clark)
Call for Contributions
* Overview of the wbernar5 Bot (William Bernardoni)
We welcome high quality submissions in the broad area of federated
* Overview of the MBot Bot (Mark Jordan)
learning (FL). A few (non-exhaustive) topics of interest include:
. Optimization algorithms for FL, particularly communication-efficient
algorithms tolerant of non-IID data
Workshop on Federated Learning for Data Privacy and
. Approaches that scale FL to larger models, including model and
Confidentiality
gradient compression techniques
. Novel applications of FL
Lixin Fan, Jakub Kone■ný, Yang Liu, Brendan McMahan, Virginia
. Theory for FL
Smith, Han Yu
. Approaches to enhancing the security and privacy of FL, including
cryptographic techniques and differential privacy
West 118 - 120, Fri Dec 13, 08:00 AM
. Bias and fairness in the FL setting
Overview . Attacks on FL including model poisoning, and corresponding defenses
. Incentive mechanisms for FL
Privacy and security have become critical concerns in recent years, . Software and systems for FL
particularly as companies and organizations increasingly collect detailed . Novel applications of techniques from other fields to the FL setting:
information about their products and users. This information can enable information theory, multi-task learning, model-agnostic meta-learning,
machine learning methods that produce better products. However, it also and etc.
has the potential to allow for misuse, especially when private data about . Work on fully-decentralized (peer-to-peer) learning will also be
individuals is involved. Recent research shows that privacy and utility do considered, as there is significant overlap in both interest and techniques
not necessarily need to be at odds, but can be addressed by careful with FL.
design and analysis. The need for such research is reinforced by the
recent introduction of new legal constraints, led by the European Union’s Submissions in the form of extended abstracts must be at most 4 pages
General Data Protection Regulation (GDPR), which is already inspiring long (not including references), be anonymized, and adhere to the
novel legislative approaches around the world such as Cyber-security NeurIPS 2019 format. Submissions will be accepted as contributed talks
Law of the People’s Republic of China and The California Consumer or poster presentations. The workshop will not have formal proceedings,
Privacy Act of 2018. but accepted papers will be posted on the workshop website.
Page 11 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
We support reproducible research and will sponsor a prize to be given to 2. Xin Yao, Tianchi Huang, Rui-Xiao Zhang, Ruiyu Li and Lifeng Sun.
the best contribution that provides code to reproduce their results. Federated Learning with Unbiased Gradient Aggregation and
Controllable Meta Updating
Submission link: https://easychair.org/conferences/?conf=flneurips2019
3. Daniel Peterson, Pallika Kanani and Virendra Marathe. Private
Important Dates (2019) Federated Learning with Domain Adaptation
Submission deadline: Sep 9
Author notification: Sep 30 4. Daliang Li and Junpu Wang.FedMD: Heterogenous Federated
Camera-Ready Papers Due: TBD Learning via Model Distillation
Workshop: Dec 13
5. Sebastian Caldas, Jakub Kone■ný, H. Brendan Mcmahan and Ameet
Organizers: Talwalkar.Mitigating the Impact of Federated Learning on Client
Lixin Fan, WeBank Resources
Jakub Kone■ný, Google
Yang Liu, WeBank 6. Jianyu Wang, Anit Sahu, Zhouyi Yang, Gauri Joshi and Soummya
Brendan McMahan, Google Kar.MATCHA: Speeding Up Decentralized SGD via Matching
Virginia Smith, CMU Decomposition Sampling
Han Yu, NTU
7. Sebastian Caldas, Sai Meher Karthik Duddu, Peter Wu, Tian Li, Jakub
Invited Speakers: Kone■ný, H. Brendan Mcmahan, Virginia Smith and Ameet
Francoise Beaufays, Principal Researcher, Google Talwalkar.Leaf: A Benchmark for Federated Settings
Shahrokh Daijavad, Distinguished Research, IBM
Dawn Song, Professor, University of California, Berkeley 8. Yihan Jiang, Jakub Kone■ný, Keith Rush and Sreeram
Ameet Talwalkar, Assistant Professor, CMU; Chief Scientist, Determined Kannan.Improving Federated Learning Personalization via Model
AI Agnostic Meta Learning
Max Welling, Professor, University of Amsterdam; VP Technologies,
Qualcomm 9. Zhicong Liang, Bao Wang, Stanley Osher and Yuan Yao.Exploring
Qiang Yang, Hong Kong University of Science and Technology, Hong Private Federated Learning with Laplacian Smoothing
Kong; Chief AI Officer, WeBank
10. Tribhuvanesh Orekondy, Seong Joon Oh, Yang Zhang, Bernt Schiele
FAQ and Mario Fritz.Gradient-Leaks: Understanding Deanonymization in
Can supplementary material be added beyond the 4-page limit and are Federated Learning
there any restrictions on it?
Yes, you may include additional supplementary material, but you should 11. Yang Liu, Yan Kang, Xinwei Zhang, Liping Li and Mingyi Hong.A
ensure that the main paper is self-contained, since looking at Communication Efficient Vertical Federated Learning Framework
supplementary material is at the discretion of the reviewers. The
supplementary material should also follow the same NeurIPS format as 12. Ahmed Khaled, Konstantin Mishchenko and Peter Richtárik.Better
the paper and be limited to a reasonable amount (max 10 pages in Communication Complexity for Local SGD
addition to the main submission).
Can a submission to this workshop be submitted to another NeurIPS 13. Yang Liu, Xiong Zhang, Shuqi Qin and Xiaoping Lei.Differentially
workshop in parallel? Private Linear Regression over Fully Decentralized Datasets
We discourage this, as it leads to more work for reviewers across
multiple workshops. Our suggestion is to pick one workshop to submit to. 14. Florian Hartmann, Sunah Suh, Arkadiusz Komarzewski, Tim D. Smith
Can a paper be submitted to the workshop that has already appeared at and Ilana Segall. Federated Learning for Ranking Browser History
a previous conference with published proceedings? Suggestions
We won’t be accepting such submissions unless they have been
adapted to contain significantly new results (where novelty is one of the 15. Aleksei Triastcyn and Boi Faltings.Federated Learning with Bayesian
qualities reviewers will be asked to evaluate). Differential Privacy
Can a paper be submitted to the workshop that is currently under review
or will be under review at a conference during the review phase? 16. Jack Goetz, Kshitiz Malik, Duc Bui, Seungwhan Moon, Honglei Liu
It is fine to submit a condensed version (i.e., 4 pages) of a parallel and Anuj Kumar.Active Federated Learning
conference submission, if it also fine for the conference in question. Our
workshop does not have archival proceedings, and therefore parallel 17. Kartikeya Bhardwaj, Wei Chen and Radu Marculescu.FedMAX:
submissions of extended versions to other conferences are acceptable. Activation Entropy Maximization Targeting Effective Non-IID Federated
Learning
=====================================================
Accepted papers: 18. Mingshu Cong, Zhongming Ou, Yanxin Zhang, Han Yu, Xi Weng,
1. Paul Pu Liang, Terrance Liu, Liu Ziyin, Russ Salakhutdinov and Jiabao Qu, Siu Ming Yiu, Yang Liu and Qiang Yang.Neural Network
Louis-Philippe Morency. Think Locally, Act Globally: Federated Learning Optimization for a VCG-based Federated Learning Incentive Mechanism
with Local and Global Representations
19. Kai Yang, Tao Fan, Tianjian Chen, Yuanming Shi and Qiang Yang.A
Page 12 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
23. Zhaorui Li, Zhicong Huang, Chaochao Chen and Cheng 11:40 AM Dawn Song Talk Song
Hong.Quantification of the Leakage in Federated Learning Sattler, El Mekkaoui, Shoham,
Hong, Hartmann, Li, Li,
24. Tzu-Ming Harry Hsu, Hang Qi and Matthew Brown.Measuring the Caldas Rivera, Wang,
Effects of Non-Identical Data Distribution for Federated Visual Bhardwaj, Orekondy, KANG,
Classification Gao, Cong, Yao, Lu, LUO,
Cen, Kairouz, Jiang, Hsu,
25. Boyue Li, Shicong Cen, Yuxin Chen and Yuejie 12:10 PM Lunch break and poster Triastcyn, Liu, Khaled Ragab
Chi.Communication-Efficient Distributed Optimization in Networks with Bayoumi, Liang, Faltings,
Gradient Tracking Moon, Li, Fan, Huang, Miao,
Qi, Brown, Glass, Wang,
26. Khaoula El Mekkaoui, Paul Blomstedt, Diego Mesquita and Samuel Chen, Marculescu, avidor,
Kaski.Towards federated stochastic gradient Langevin dynamics Wu, Hong, Ju, Rush, Zhang,
ZHOU, Beaufays, Zhu, Xia
27. Felix Sattler, Klaus-Robert Müller and Wojciech Samek.Clustered
Federated Learning 01:30 PM Dan Ramage Talk Ramage
29. Neta Shoham, Tomer Avidor, Aviv Keren, Nadav Israel, Daniel 02:50 PM Contributed talk #7
Benditkis, Liron Mor-Yosef and Itai Zeitak.Overcoming Forgetting in
03:00 PM Contributed talk #8
Federated Learning on Non-IID Data
03:10 PM Raluca Popa Talk Popa
30. Ahmed Khaled and Peter Richtárik.Gradient Descent with
03:40 PM Coffee break and poster
Compressed Iterates
04:15 PM Contributed talk #9
31. Jiahuan Luo, Xueyang Wu, Yun Luo, Anbu Huang, Yunfeng Huang,
04:25 PM Contributed talk #10
Yang Liu and Qiang Yang.Real-World Image Datasets for Federated
Learning FOCUS: Federate
04:35 PM Opportunity Computing for Chen
32. Ahmed Khaled, Konstantin Mishchenko and Peter Richtárik.First Ubiquitous System
Analysis of Local GD on Heterogeneous Data
05:00 PM Panel disucssion
33. Dashan Gao, Ce Ju, Xiguang Wei, Yang Liu, Tianjian Chen and 06:00 PM Closing Remark
Qiang Yang. HHHFL: Hierarchical Heterogeneous Horizontal Federated
Learning for Electroencephalography
=====================================================
Machine Learning for the Developing World (ML4D):
The workshop schedule (tentative):
Challenges and Risks
Schedule
Maria De-Arteaga, Amanda Coston, Tejumade Afonja
08:45 AM Opening remarks Fan West 121 + 122, Fri Dec 13, 08:00 AM
08:50 AM Contributed talk #0 As the use of machine learning becomes ubiquitous, there is growing
interest in understanding how machine learning can be used to tackle
Page 13 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 14 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
alignment of Pidgin English word vectors with English word vectors approach is founded on a distributional notion of semantics, i.e. that the
which achieves a Nearest Neighbor accuracy of 0.1282. This aligned ''meaning'' of a word is based only on its relationship to other words.
vector will be helpful in the performance of various downstream tasks While effective for many applications, this approach suffers from limited
and transfer of models from English to Pidgin. Finally, the creation of an semantic understanding -- symbols learned this way lack any concrete
Unsupervised Neural Machine Translation model between Pidgin English groundings into the multimodal, interactive environment in which
and English which achieves a BLEU score of 20.82 from English to communication takes place. The symbol grounding problem first
Pidgin and 21.59 from Pidgin to English. In all, this work greatly reduces highlighted this limitation, that ``meaningless symbols (i.e. words) cannot
the barrier of entry for future works on Pidgin English. be grounded in anything but other meaningless symbols''.
Abstract 11: A Noxious Market for Personal Data in Machine On the other hand, humans acquire language by communicating about
Learning for the Developing World (ML4D): Challenges and Risks, and interacting within a rich, perceptual environment -- providing
Abdulrahim 02:30 PM concrete groundings, e.g. to objects or concepts either physical or
psychological. Thus, recent works have aimed to bridge computer vision,
Many policymakers, academics and governments have advocated for interactive learning, and natural language understanding through
exchangeable property rights over information as it presents a market language learning tasks based on natural images or through embodied
solution to what could be considered a market failure. Particularly in agents performing interactive tasks in physically simulated environments,
jurisdictions such as Africa, Asia or South America, where weaker legal often drawing on the recent successes of deep learning and
protections and fleeting regulatory enforcement leaves data subjects reinforcement learning. We believe these lines of research pose a
vulnerable or exploited regardless of the outcome. We argue that promising approach for building models that do grasp the world's
whether we could achieve this personal data economy in which underlying complexity.
individuals have ownership rights akin to property rights over their data
should be approached with caution as a solution to ensuring individuals The goal of this third ViGIL workshop is to bring together scientists from
have agency over their data across different legal landscapes. various backgrounds - machine learning, computer vision, natural
language processing, neuroscience, cognitive science, psychology, and
We present an objection to the use of property rights, a market solution, philosophy - to share their perspectives on grounding, embodiment, and
due to the noxious nature of personal data - which is founded on Satz interaction. By providing this opportunity for cross-discipline discussion,
and Sandell's objection to markets. Ultimately, our rights over personal we hope to foster new ideas about how to learn and leverage grounding
data and privacy are borne out of our basic human rights and are a in machines as well as build new bridges between the science of human
precondition for the self-development, personal fulfilment and the free cognition and machine learning.
enjoyment of other fundamental human rights - and putting it up for sale
risks corrupting its essence and value. Schedule
Computer vision systems for wood identification have the potential to From Human Language to
09:10 AM Thomason
empower both producer and consumer countries to combat illegal Agent Action
logging if they can be deployed effectively in the field. In this work,
09:50 AM Coffee Break
carried out as part of an active international partnership with the support
of UNIDO, we constructed and curated a field-relevant image data set to 10:30 AM Spotlight
train a classifier for wood identification of $15$ commercial Ghanaian
woods using the XyloTron system. We tested model performance in the Why language
laboratory, and then collected real-world field performance data across 10:50 AM understanding is not a McClelland
multiple sites using multiple XyloTron devices. We present efficacies of solved problem
the trained model in the laboratory and in the field, discuss practical 11:30 AM Louis-Philippe Morency Morency
implications and challenges of deploying machine learning wood
identification models, and conclude that field testing is a necessary step - Ross, Mrabet, Subramanian,
and should be considered the gold-standard - for validating computer Cideron, Mu, Bhooshan, Okur
vision wood identification systems. Kavil, Delbrouck, Kuo, Lair,
Ilharco, Jayram, Herrera
Palacio, Fujiyama, Tieleman,
Potapenko, Chao, Sutter,
Visually Grounded Interaction and Language
12:10 PM Poster session Kovaleva, Lai, Wang, Sharma,
Florian Strub, Abhishek Das, Erik Wijmans, Harm de Vries, Stefan Cangea, Krishnaswamy,
Lee, Alane Suhr, Drew Arad Hudson Tsuboi, Kuhnle, Nguyen, Yu,
Saha, Xiang, Venkataraman,
West 202 - 204, Fri Dec 13, 08:00 AM Kalra, Xie, Doran, Goodwin,
Kadav, Daghaghi, Baldridge,
The dominant paradigm in modern natural language understanding is Wu
learning statistical language models from text-only corpora. This
Page 15 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 16 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Oral highlight presentations Machine learning is about computational methods that enable machines
02:00 PM for selected contributed to learn concepts and improve performance from experience. Here,
papers (10 min x 6) experience can take diverse forms, including data examples, abstract
knowledge, interactions and feedback from the environment, other
02:30 PM Discussion Panel Veloso models, and so forth. Depending on different assumptions on the types
Putting Ethical AI to the and amount of experience available there are different learning
03:00 PM Procaccia paradigms, such as supervised learning, active learning, reinforcement
Vote
learning, knowledge distillation, adversarial learning, and combinations
thereof. On the other hand, a hallmark of human intelligence is the ability
to learn from all sources of information. In this workshop, we aim to
explore various aspects of learning paradigms, particularly theoretical
properties and formal connections between them, and new algorithms
combining multiple modes of supervisions, etc.
Schedule
Page 17 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
West 211 - 214, Fri Dec 13, 08:00 AM 11:15 AM Spotlight talks Scieur, Mishchenko, Anil
Page 18 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abstract 1: Opening Remarks in Beyond first order methods in Abstract 5: Adaptive gradient methods: efficient implementation and
machine learning systems, Kyrillidis, Berahas, Roosta, Mahoney 08:00 generalization in Beyond first order methods in machine learning
AM systems, 10:30 AM
Opening remarks for the workshop by the organizers Adaptive gradient methods have had a transformative impact in deep
learning. We will describe recent theoretical and experimental advances
Abstract 2: Economical use of second-order information in training in their understanding, including low-memory adaptive preconditioning,
machine learning models in Beyond first order methods in machine and insights into their generalizaton ability.
learning systems, Goldfarb 08:30 AM
Abstract 6: Spotlight talks in Beyond first order methods in machine
Stochastic gradient descent (SGD) and variants such as Adagrad and learning systems, Scieur, Mishchenko, Anil 11:15 AM
Adam, are extensively used today to train modern machine learning
models. In this talk we will discuss ways to economically use Symmetric Multisecant quasi-Newton methods. Damien Scieur
second-order information to modify both the step size (learning rate) (Samsung AI Research Montreal); Thomas Pumir (Princeton University);
used in SGD and the direction taken by SGD. Our methods adaptively Nicolas Boumal (Princeton University)
control the batch sizes used to compute gradient and Hessian
approximations and and ensure that the steps that are taken decrease Stochastic Newton Method and its Cubic Regularization via
the loss function with high probability assuming that the latter is Majorization-Minimization. Konstantin Mishchenko (King Abdullah
self-concordant, as is true for many problems in empirical risk University of Science & Technology (KAUST)); Peter Richtarik (KAUST);
minimization. For such cases we prove that our basic algorithm is Dmitry Koralev (KAUST)
globally linearly convergent. A slightly modified version of our method is
presented for training deep learning models. Numerical results will be Full Matrix Preconditioning Made Practical. Rohan Anil (Google); Vineet
presented that show that it exhibits excellent performance without the Gupta (Google); Tomer Koren (Google); Kevin Regan (Google); Yoram
need for learning rate tuning. If there is time, additional ways to efficiently Singer (Princeton)
make use of second-order information will be presented.
Abstract 8: K-FAC: Extensions, improvements, and applications in
Abstract 3: Spotlight talks in Beyond first order methods in machine Beyond first order methods in machine learning systems, Martens
learning systems, Granziol, Pedregosa, Asi 09:00 AM 02:00 PM
How does mini-batching affect Curvature information for second order Second order optimization methods have the potential to be much faster
deep learning optimization? Diego Granziol (Oxford); Stephen Roberts than first order methods in the deterministic case, or pre-asymptotically
(Oxford); Xingchen Wan (Oxford University); Stefan Zohren (University of in the stochastic case. However, traditional second order methods have
Oxford); Binxin Ru (University of Oxford); Michael A. Osborne (University proven ineffective or impractical for neural network training, due in part to
of Oxford); Andrew Wilson (NYU); sebastien ehrhardt (Oxford); Dmitry P the extremely high dimension of the parameter space.
Vetrov (Higher School of Economics); Timur Garipov (Samsung AI Kronecker-factored Approximate Curvature (K-FAC) is second-order
Center in Moscow) optimization method based on a tractable approximation to the
Gauss-Newton/Fisher matrix that exploits the special structure present in
Acceleration through Spectral Modeling. Fabian Pedregosa (Google); neural network training objectives. This approximation is neither low-rank
Damien Scieur (Princeton University) nor diagonal, but instead involves Kronecker-products, which allows for
efficient estimation, storage and inversion of the curvature matrix. In this
Using better models in stochastic optimization. Hilal Asi (Stanford talk I will introduce the basic K-FAC method for standard MLPs and then
University); John Duchi (Stanford University) present some more recent work in this direction, including extensions to
CNNs and RNNs, both of which requires new approximations to the
Ellipsoidal Trust Region Methods for Neural Nets. Leonard Adolphs Fisher. For these I will provide mathematical intuitions and empirical
(ETHZ); Jonas Kohler (ETHZ) results which speak to their efficacy in neural network optimization. Time
permitting, I will also discuss some recent results on large-batch
Sub-sampled Newton Methods Under Interpolation. Si Yi Meng optimization with K-FAC, and the use of adaptive adjustment methods
(University of British Columbia); Sharan Vaswani (Mila, Université de that can eliminate the need for costly hyperparameter tuning.
Montréal); Issam Laradji (University of British Columbia); Mark Schmidt
(University of British Columbia); Simon Lacoste-Julien (Mila, Université Abstract 9: Spotlight talks in Beyond first order methods in machine
de Montréal) learning systems, Grigas, Yao, Adolphs, Meng 02:45 PM
Abstract 4: Poster Session in Beyond first order methods in machine Hessian-Aware trace-Weighted Quantization. Zhen Dong (UC Berkeley);
learning systems, Gorbunov, d'Aspremont, Wang, Wang, Ginsburg, Zhewei Yao (University of California, Berkeley); Amir Gholami (UC
Quaglino, Castera, Adya, Granziol, Das, Bollapragada, Pedregosa, Berkeley); Yaohui Cai (Peking University); Daiyaan Arfeen (UC
Takac, Jahani, Karimireddy, Asi, Daroczy, Adolphs, Rawal, Brandt, Li, Berkeley); Michael Mahoney ("University of California, Berkeley"); Kurt
Ughi, Romero, Skorokhodov, Scieur, Bae, Mishchenko, Anil, Sharan, Keutzer (UC Berkeley)
Balu, Chen, Yao, Ergen, Grigas, Li, Ba, Roberts, Vaswani, Eftekhari,
Sharma 09:45 AM New Methods for Regularization Path Optimization via Differential
Equations. Paul Grigas (UC Berkeley); Heyuan Liu (University of
Page 19 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
California, Berkeley) (Oxford); Xingchen Wan (Oxford University); Stefan Zohren (University of
Oxford); Binxin Ru (University of Oxford); Michael A. Osborne (University
Ellipsoidal Trust Region Methods for Neural Nets. Leonard Adolphs of Oxford); Andrew Wilson (NYU); sebastien ehrhardt (Oxford); Dmitry P
(ETHZ); Jonas Kohler (ETHZ) Vetrov (Higher School of Economics); Timur Garipov (Samsung AI
Center in Moscow)
Sub-sampled Newton Methods Under Interpolation. Si Yi Meng
(University of British Columbia); Sharan Vaswani (Mila, Université de On the Convergence of a Biased Version of Stochastic Gradient
Montréal); Issam Laradji (University of British Columbia); Mark Schmidt Descent. Rudrajit Das (University of Texas at Austin); Jiong Zhang
(University of British Columbia); Simon Lacoste-Julien (Mila, Université (UT-Austin); Inderjit S. Dhillon (UT Austin & Amazon)
de Montréal)
Adaptive Sampling Quasi-Newton Methods for Derivative-Free
Abstract 10: Poster Session (same as above) in Beyond first order Stochastic Optimization. Raghu Bollapragada (Argonne National
methods in machine learning systems, 03:30 PM Laboratory); Stefan Wild (Argonne National Laboratory)
An Accelerated Method for Derivative-Free Smooth Stochastic Convex * Acceleration through Spectral Modeling. Fabian Pedregosa (Google);
Optimization. Eduard Gorbunov (Moscow Institute of Physics and Damien Scieur (Princeton University)
Technology); Pavel Dvurechenskii (WIAS Germany); Alexander
Gasnikov (Moscow Institute of Physics and Technology) Accelerating Distributed Stochastic L-BFGS by sampled 2nd-Order
Information. Jie Liu (Lehigh University); Yu Rong (Tencent AI Lab);
Fast Bregman Gradient Methods for Low-Rank Minimization Problems. Martin Takac (Lehigh University); Junzhou Huang (Tencent AI Lab)
Radu-Alexandru Dragomir (Université Toulouse 1); Jérôme Bolte
(Université Toulouse 1); Alexandre d'Aspremont (Ecole Normale Grow Your Samples and Optimize Better via Distributed Newton CG and
Superieure) Accumulating Strategy. Majid Jahani (Lehigh University); Xi He (Lehigh
University); Chenxin Ma (Lehigh University); Aryan Mokhtari (UT Austin);
Gluster: Variance Reduced Mini-Batch SGD with Gradient Clustering. Dheevatsa Mudigere (Intel Labs); Alejandro Ribeiro (University of
Fartash Faghri (University of Toronto); David Duvenaud (University of Pennsylvania); Martin Takac (Lehigh University)
Toronto); David Fleet (University of Toronto); Jimmy Ba (University of
Toronto) Global linear convergence of trust-region Newton's method without
strong-convexity or smoothness. Sai Praneeth Karimireddy (EPFL);
Neural Policy Gradient Methods: Global Optimality and Rates of Sebastian Stich (EPFL); Martin Jaggi (EPFL)
Convergence. Lingxiao Wang (Northwestern University); Qi Cai
(Northwestern University); Zhuoran Yang (Princeton University); Zhaoran FD-Net with Auxiliary Time Steps: Fast Prediction of PDEs using
Wang (Northwestern University) Hessian-Free Trust-Region Methods. Nur Sila Gulgec (Lehigh
University); Zheng Shi (Lehigh University); Neil Deshmukh (MIT
A Gram-Gauss-Newton Method Learning Overparameterized Deep BeaverWorks - Medlytics); Shamim Pakzad (Lehigh University); Martin
Neural Networks for Regression Problems. Tianle Cai (Peking Takac (Lehigh University)
University); Ruiqi Gao (Peking University); Jikai Hou (Peking University);
Siyu Chen (Peking University); Dong Wang (Peking University); Di He * Using better models in stochastic optimization. Hilal Asi (Stanford
(Peking University); Zhihua Zhang (Peking University); Liwei Wang University); John Duchi (Stanford University)
(Peking University)
Tangent space separability in feedforward neural networks. Bálint
Stochastic Gradient Methods with Layerwise Adaptive Moments for Daróczy (Institute for Computer Science and Control, Hungarian
Training of Deep Networks. Boris Ginsburg (NVIDIA); Oleksii Hrinchuk Academy of Sciences); Rita Aleksziev (Institute for Computer Science
(NVIDIA); Jason Li (NVIDIA); Vitaly Lavrukhin (NVIDIA); Ryan Leary and Control, Hungarian Academy of Sciences); Andras Benczur
(NVIDIA); Oleksii Kuchaiev (NVIDIA); Jonathan Cohen (NVIDIA); Huyen (Hungarian Academy of Sciences)
Nguyen (NVIDIA); Yang Zhang (NVIDIA)
* Ellipsoidal Trust Region Methods for Neural Nets. Leonard Adolphs
Accelerating Neural ODEs with Spectral Elements. Alessio Quaglino (ETHZ); Jonas Kohler (ETHZ)
(NNAISENSE SA); Marco Gallieri (NNAISENSE); Jonathan Masci
(NNAISENSE); Jan Koutnik (NNAISENSE) Closing the K-FAC Generalisation Gap Using Stochastic Weight
Averaging. Xingchen Wan (University of Oxford); Diego Granziol
An Inertial Newton Algorithm for Deep Learning. Camille Castera (CNRS, (Oxford); Stefan Zohren (University of Oxford); Stephen Roberts (Oxford)
IRIT); Jérôme Bolte (Université Toulouse 1); Cédric Févotte (CNRS,
IRIT); Edouard Pauwels (Toulouse 3 University) * Sub-sampled Newton Methods Under Interpolation. Si Yi Meng
(University of British Columbia); Sharan Vaswani (Mila, Université de
Nonlinear Conjugate Gradients for Scaling Synchronous Distributed DNN Montréal); Issam Laradji (University of British Columbia); Mark Schmidt
Training. Saurabh Adya (Apple); Vinay Palakkode (Apple Inc.); Oncel (University of British Columbia); Simon Lacoste-Julien (Mila, Université
Tuzel (Apple Inc.) de Montréal)
* How does mini-batching affect Curvature information for second order Learned First-Order Preconditioning. Aditya Rawal (Uber AI Labs); Rui
deep learning optimization? Diego Granziol (Oxford); Stephen Roberts Wang (Uber AI); Theodore Moskovitz (Gatsby Computational
Page 20 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Neuroscience Unit); Sanyam Kapoor (Uber); Janice Lan (Uber AI); Jason Berkeley); Zhewei Yao (University of California, Berkeley); Amir Gholami
Yosinski (Uber AI Labs); Thomas Miconi (Uber AI Labs) (UC Berkeley); Yaohui Cai (Peking University); Daiyaan Arfeen (UC
Berkeley); Michael Mahoney ("University of California, Berkeley"); Kurt
Iterative Hessian Sketch in Input Sparsity Time. Charlie Dickens Keutzer (UC Berkeley)
(University of Warwick); Graham Cormode (University of Warwick)
Random Projections for Learning Non-convex Models. Tolga Ergen
Nonlinear matrix recovery. Florentin Goyens (University of Oxford); (Stanford University); Emmanuel Candes (Stanford University); Mert
Coralia Cartis (Oxford University); Armin Eftekhari (EPFL) Pilanci (Stanford)
Making Variance Reduction more Effective for Deep Networks. Nicolas * New Methods for Regularization Path Optimization via Differential
Brandt (EPFL); Farnood Salehi (EPFL); Patrick Thiran (EPFL) Equations. Paul Grigas (UC Berkeley); Heyuan Liu (University of
California, Berkeley)
Novel and Efficient Approximations for Zero-One Loss of Linear
Classifiers. Hiva Ghanbari (Lehigh University); Minhan Li (Lehigh Hessian-Aware Zeroth-Order Optimization. Haishan Ye (HKUST);
University); Katya Scheinberg (Lehigh) Zhichao Huang (HKUST); Cong Fang (Peking University); Chris Junchi Li
(Tencent); Tong Zhang (HKUST)
A Model-Based Derivative-Free Approach to Black-Box Adversarial
Examples: BOBYQA. Giuseppe Ughi (University of Oxford) Higher-Order Accelerated Methods for Faster Non-Smooth Optimization.
Brian Bullins (TTIC)
Distributed Accelerated Inexact Proximal Gradient Method via System of
Coupled Ordinary Differential Equations. Chhavi Sharma (IIT Bombay); Abstract 11: Analysis of linear search methods for various gradient
Vishnu Narayanan (IIT Bombay); Balamurugan Palaniappan (IIT approximation schemes for noisy derivative free optimization. in
Bombay) Beyond first order methods in machine learning systems,
Scheinberg 04:15 PM
Finite-Time Convergence of Continuous-Time Optimization Algorithms
via Differential Inclusions. Orlando Romero (Rensselaer Polytechnic We develop convergence analysis of a modified line search method for
Institute); Mouhacine Benosman (MERL) objective functions whose value is computed with noise and whose
gradient estimates are not directly available. The noise is assumed to be
Loss Landscape Sightseeing by Multi-Point Optimization. Ivan bounded in absolute value without any additional assumptions. In this
Skorokhodov (MIPT); Mikhail Burtsev (NI) case, gradient approximation can be constructed via interpolation or
sample average approximation of smoothing gradients and thus they are
* Symmetric Multisecant quasi-Newton methods. Damien Scieur always inexact and possibly random. We extend the framework based on
(Samsung AI Research Montreal); Thomas Pumir (Princeton University); stochastic methods which was developed to provide analysis of a
Nicolas Boumal (Princeton University) standard line-search method with exact function values and random
gradients to the case of noisy function. We introduce a condition on the
Does Adam optimizer keep close to the optimal point? Kiwook Bae gradient which when satisfied with some sufficiently large probability at
(KAIST)*; Heechang Ryu (KAIST); Hayong Shin (KAIST) each iteration, guarantees convergence properties of the line search
method. We derive expected complexity bounds for convex, strongly
* Stochastic Newton Method and its Cubic Regularization via convex and nonconvex functions. We motivate these results with several
Majorization-Minimization. Konstantin Mishchenko (King Abdullah recent papers related to policy optimization.
University of Science & Technology (KAUST)); Peter Richtarik (KAUST);
Abstract 12: Second-order methods for nonconvex optimization with
Dmitry Koralev (KAUST)
complexity guarantees in Beyond first order methods in machine
learning systems, Wright 05:00 PM
* Full Matrix Preconditioning Made Practical. Rohan Anil (Google); Vineet
Gupta (Google); Tomer Koren (Google); Kevin Regan (Google); Yoram
We consider problems of smooth nonconvex optimization:
Singer (Princeton)
unconstrained, bound-constrained, and with general equality constraints.
We show that algorithms for these problems that are widely used in
Memory-Sample Tradeoffs for Linear Regression with Small Error. Vatsal
practice can be modified slightly in ways that guarantees convergence to
Sharan (Stanford University); Aaron Sidford (Stanford); Gregory Valiant
approximate first- and second-order optimal points with complexity
(Stanford University)
guarantees that depend on the desired accuracy. The methods we
discuss are constructed from Newton's method, the conjugate gradient
On the Higher-order Moments in Adam. Zhanhong Jiang (Johnson
method, log-barrier method, and augmented Lagrangians. (In some
Controls International); Aditya Balu (Iowa State University); Sin Yong
cases, special structure of the objective function makes for only a weak
Tan (Iowa State University); Young M Lee (Johnson Controls
dependence on the accuracy parameter.) Our methods require Hessian
International); Chinmay Hegde (Iowa State University); Soumik Sarkar
information only in the form of Hessian-vector products, so do not require
(Iowa State University)
the Hessian to be evaluated and stored explicitly. This talk describes
joint work with Clement Royer, Yue Xie, and Michael O'Neill.
h-matrix approximation for Gauss-Newton Hessian. Chao Chen (UT
Austin) Abstract 13: Final remarks in Beyond first order methods in machine
learning systems, Kyrillidis, Berahas, Roosta, Mahoney 05:45 PM
* Hessian-Aware trace-Weighted Quantization. Zhen Dong (UC
Page 21 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
The model-to-data
The workshop provides room for discussion on these topics and aims to
paradigm: overcoming data
bring together potential partners to organize such challenges and 03:00 PM Guinney
access barriers in
stimulate "machine learning for good", i.e. the organization of challenges
biomedical competitions
for the benefit of society. We have invited prominent speakers that have
experience in this domain. The Deep Learning Epilepsy
Detection Challenge:
Schedule Design, Implementation, and
03:15 PM Kiral
Test of a New
Crowd-Sourced AI
Welcome and Opening Mendrik, Tu, Guyon, Viegas,
08:00 AM Challenge Ecosystem
Remarks LI
03:30 PM Coffee Break
Amir Banifatemi (XPrize) "AI
08:15 AM for Good via Machine Banifatemi Frank Hutter (University of
Learning Challenges" Freiburg) "A Proposal for a
04:15 PM New Competition Design Hutter
Emily Bender (University of
Emphasizing Scientific
Washington) "Making
Insights"
Stakeholder Impacts Visible
in the Evaluation Cycle: Open Space Topic “The
09:00 AM Bender
Towards Organization of Challenges Mendrik, Guyon, Tu, Viegas,
Fairness-Integrated Shared 05:00 PM
for the Benefit of More LI
Tasks and Evaluation Diverse Communities”
Metrics"
Page 22 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
In a typical machine learning competition or shared task, success is Abstract 8: The AI Driving Olympics: An Accessible Robot Learning
measured in terms of systems' ability to reproduce gold-standard labels. Benchmark in CiML 2019: Machine Learning Competitions for All,
The potential impact of the systems being developed on stakeholder Walter 11:45 AM
populations, if considered at all, is studied separately from system
`performance'. Given the tight train-eval cycle of both shared tasks and Despite recent breakthroughs, the ability of deep learning and
system development in general, we argue that making disparate impact reinforcement learning to outperform traditional approaches to control
on vulnerable populations visible in dataset and metric design will be key physically embodied robotic agents remains largely unproven. To help
to making the potential for such impact present and salient to bridge this gap, we have developed the “AI Driving Olympics” (AI-DO), a
developers. We see this as an effective way to promote the development competition with the objective of evaluating the state-of-the-art in
of machine learning technology that is helpful for people, especially machine learning and artificial intelligence for mobile robotics. Based on
those who have been subject to marginalization. This talk will explore the simple and well specified autonomous driving and navigation
how to develop such shared tasks, considering task choice, stakeholder environment called “Duckietown,” AI-DO includes a series of tasks of
community input, and annotation and metric design desiderata. increasing complexity—from simple lane-following to fleet management.
For each task, we provide tools for competitors to use in the form of
Joint work with Hal Daumé III, University of Maryland, Bernease Herman, simulators, data logs, code templates, baseline implementations, and
University of Washington, and Brandeis Marshall, Spelman College. low-cost access to robotic hardware. We evaluate submissions in
simulation online, on standardized hardware environments, and finally at
Abstract 5: Dina Machuve (Nelson Mandela African Institution of the competition events. We have held successful AI-DO competitions at
Science and Technology) “Machine Learning Competitions: The NeurIPS 2018 and ICRA 2019, and will be holding AI-DO 3 at NeurIPS
Outlook from Africa” in CiML 2019: Machine Learning Competitions 2020. Together, these competitions highlight the need for better
for All, Machuve 10:30 AM benchmarks, which are lacking in robotics, as well as improved
mechanisms to bridge the gap between simulation and reality.
The current AI landscape in Africa mainly focuses on capacity building.
The ongoing efforts to strengthen the AI capacity in Africa are organized Abstract 10: Catered Lunch and Poster Viewing (in Workshop Room)
in summer schools, workshops, meetups, competitions and one in CiML 2019: Machine Learning Competitions for All, Stolovitzky,
long-term program at the Masters level. The main AI initiatives driving Pradhan, Duboue, Tang, Natekin, Bondi, Bouthillier, Milani, Müller,
the AI capacity building agenda in Africa include a) Deep Learning Holzinger, Harrer, Day, Ustyuzhanin, Guss, Mirmomeni 12:15 PM
Indaba, b) Data Science Africa, c) Data Science Nigeria, d) Nairobi
Women in Machine Learning and Data Science, e) Zindi and f) The Accepted Posters
African Master's in Machine Intelligence (AMMI) at AIMS. The talk will
summarize our experience on low participation of African AI developers Kandinsky Patterns: An open toolbox for creating explainable machine
at machine learning competitions and our recommendations to address learning challenges
the current challenges. Heimo Muller · Andreas Holzinger
Page 23 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
FDL: Mission Support Challenge Abstract 13: The model-to-data paradigm: overcoming data access
Luís F. Simões · Ben Day · Vinutha M. Shreenath · Callum Wilson barriers in biomedical competitions in CiML 2019: Machine Learning
Competitions for All, Guinney 03:00 PM
From data challenges to collaborative gig science. Coopetitive research
process and platform Data competitions often rely on the physical distribution of data to
Andrey Ustyuzhanin · Mikhail Belous · Leyla Khatbullina · Giles Strong challenge participants, a significant limitation given that much data is
proprietary, sensitive, and often non-shareable. To address this, the
Smart(er) Machine Learning for Practitioners DREAM Challenges has advanced a challenge framework called
Prabhu Pradhan modelto-data (MTD), requiring participants to submit re-runnable
algorithms instead of model predictions. The DREAM organization has
Improving Reproducibility of Benchmarks successfully completed multiple MTD-based challenges, and is
Xavier Bouthillier expanding this approach to unlock highly sensitive and non-distributable
human data for use in biomedical data challenges.
Guaranteeing Reproducibility in Deep Learning Competitions
Abstract 16: Frank Hutter (University of Freiburg) "A Proposal for a
Brandon Houghton
New Competition Design Emphasizing Scientific Insights" in CiML
2019: Machine Learning Competitions for All, Hutter 04:15 PM
Organizing crowd-sourced AI challenges in enterprise environments:
opportunities and challenges
The typical setup in machine learning competitions is to provide one or
Mahtab Mirmomeni · Isabell Kiral · Subhrajit Roy · Todd Mummert · Alan
more datasets and a performance metric, leaving it entirely up to
Braz · Jason Tsay · Jianbin Tang · Umar Asif · Thomas Schaffter · Eren
participants which approach to use, how to engineer better features,
Mehmet · Bruno De Assis Marques · Stefan Maetschke · Rania Khalaf ·
whether and how to pretrain models on related data, how to tune
Michal Rosen-Zvi · John Cohn · Gustavo Stolovitzky · Stefan Harrer
hyperparameters, how to combine multiple models in an ensemble, etc.
The fact that work on each of these components often leads to
WikiCities: a Feature Engineering Educational Resource
substantial improvements has several consequences: (1) amongst
Pablo Duboue
several skilled teams, the one with the most manpower and engineering
drive often wins; (2) it is often unclear *why* one entry performs better
Reinforcement Learning Meets Information Seeking: Dynamic Search
than another one; and (3) scientific insights remain limited.
Challenge
Zhiwen Tang · Grace Hui Yang
Based on my experience in both participating in several challenges and
also organizing some, I will propose a new competition design that
AI Journey 2019: School Tests Solving Competition
instead emphasizes scientific insight by dividing the various ways in
Alexey Natekin · Peter Romov · Valentin Malykh
which teams could improve performance into (largely orthogonal)
modular components, each of which defines its own competition. E.g.,
A BIRDSAI View for Conservation
one could run a competition focussing only on effective hyperparameter
Elizabeth Bondi · Milind Tambe · Raghav Jain · Palash Aggrawal · Saket
tuning of a given pipeline (across private datasets). With the same code
Anand · Robert Hannaford · Ashish Kapoor · Jim Piavis · Shital Shah ·
base and datasets, one could likewise run a competition focussing only
Lucas Joppa · Bistra Dilkina
on finding better neural architectures, or only better preprocessing
methods, or only a better training pipeline, or only better pre-training
Abstract 12: Design and Analysis of Experiments: A Challenge
methods, etc. One could also run multiple of these competitions in
Approach in Teaching in CiML 2019: Machine Learning
parallel, hot-swapping better components found in one competition into
Competitions for All, Pavao 02:45 PM
the other competitions. I will argue that the result would likely be
Over the past few years, we have explored the benefits of involving substantially more valuable in terms of scientific insights than traditional
students both in organizing and in participating in challenges as a competitions and may even lead to better final performance.
pedagogical tool, as part of an international collaboration. Engaging in
Abstract 17: Open Space Topic “The Organization of Challenges for
the design and resolution of a competition can be seen as a hands-on
the Benefit of More Diverse Communities” in CiML 2019: Machine
means of learning proper design and analysis of experiments and
Learning Competitions for All, Mendrik, Guyon, Tu, Viegas, LI 05:00
gaining a deeper understanding other aspects of Machine Learning.
PM
Graduate students of University Paris-Sud (Paris, France) are involved in
class projects in creating a challenge end-to-end, from defining the
“Open Space” is a technique for running meetings where the participants
research problem, collecting or formatting data, creating a starting kit, to
create and manage the agenda themselves. Participants can propose
implementing and testing the website. The application domains and
ideas that address the open space topic, these will be divided into
types of data are extremely diverse: medicine, ecology, marketing,
various sessions that all other participants can join and brainstorm about.
computer vision, recommendation, text processing, etc. The challenges
After the open space we will collect all the ideas and see whether we
thus created are then used as class projects of undergraduate students
could write a whitepaper on this topic with all participants.
who have to solve them, both at University Paris-Sud, and at Rensselaer
Page 24 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
* Yang Cai
AI for Humanitarian Assistance and Disaster Response Abstract 3: Spotlight Talks (x6) in AI for Humanitarian Assistance
and Disaster Response, Kruspe, Dalmasso, Schrempf, Oh, Doshi, Lu
Ritwik Gupta, Robin Murphy, Trevor Darrell, Eric Heim, Zhangyang 10:15 AM
Wang, Bryce Goodman, Piotr Bili■ski
TBD based on accepted papers
West 217 - 219, Fri Dec 13, 08:00 AM
Abstract 5: Invited Talks (x4) in AI for Humanitarian Assistance and
Natural disasters are one of the oldest threats to not just individuals but Disaster Response, Rasmussen, Stromberg, Darrell 01:00 PM
to the societies they co-exist in. As a result, humanity has ceaselessly
sought way to provide assistance to people in need after disasters have * Eric Rasmussen
struck. Further, natural disasters are but a single, extreme example of * Maj Megan Stromberg
the many possible humanitarian crises. Disease outbreak, famine, and * TBD
oppression against disadvantaged groups can pose even greater * TBD
dangers to people that have less obvious solutions.
Abstract 6: Spotlight Talks (x6) in AI for Humanitarian Assistance
In this proposed workshop, we seek to bring together the Artificial
and Disaster Response, Wang, Seo, Veitch-Michaelis, Sidrane,
Intelligence (AI) and Humanitarian Assistance and Disaster Response
Kapadia, Nevo, Dubey 03:00 PM
(HADR) communities in order to bring AI to bear on real-world
humanitarian crises.
TBD based on accepted papers
Through this workshop, we intend to establish meaningful dialogue
between the communities.
Abstract 7: Convergence: Two-Way Limitations in Taking Theory to
Applications in AI for Humanitarian Assistance and Disaster
By the end of the workshop, the NeurIPS research community can come
Response, Dzombak, Yang, 04:30 PM
to understand the practical challenges of in aiding those in crisis, while
the HADR can understand the landscape that is the state of art and Speakers from Berkeley, Oak Ridge National Lab, Red Cross, and more.
practice in AI.
Through this, we seek to begin establishing a pipeline of transitioning the
research created by the NeurIPS community to real-world humanitarian Shared Visual Representations in Human and Machine
issues.
Intelligence
Schedule
Arturo Deza, Joshua Peterson, Apurva Ratan Murty, Tom Griffiths
08:00 AM Introduction and Welcome Gupta, Sajeev West 220 - 222, Fri Dec 13, 08:00 AM
08:15 AM Invited Talks (x4) Matias, Adole, Brown The goal of the Shared Visual Representations in Human and Machine
Intelligence workshop is to disseminate relevant, parallel findings in the
Kruspe, Dalmasso, Schrempf,
10:15 AM Spotlight Talks (x6) fields of computational neuroscience, psychology, and cognitive science
Oh, Doshi, Lu
that may inform modern machine learning methods. In the past few
11:30 AM Lunch years, machine learning methods---especially deep neural
networks---have widely permeated the vision science, cognitive science,
Rasmussen, Stromberg,
01:00 PM Invited Talks (x4) and neuroscience communities.
Darrell
As a result, scientific modeling in these fields has greatly benefited,
Wang, Seo, Veitch-Michaelis, producing a swath of potentially critical new insights into human learning
03:00 PM Spotlight Talks (x6) Sidrane, Kapadia, Nevo, and intelligence, which remains the gold standard for many tasks.
Dubey However,
the machine learning community has been largely unaware of these
Convergence: Two-Way
cross-disciplinary insights and analytical tools, which may help to solve
04:30 PM Limitations in Taking TheoryDzombak, Yang,
many of the current problems that ML theorists and engineers face today
to Applications
(\textit{e.g.,} adversarial attacks, compression, continual learning,
05:15 PM Poster Session and unsupervised learning).
Thus we propose to invite leading cognitive scientists with strong
computational backgrounds to disseminate their findings to the machine
Abstracts (5): learning community with the hope of closing the loop by nourishing new
ideas and creating cross-disciplinary collaborations.
Abstract 2: Invited Talks (x4) in AI for Humanitarian Assistance and
Disaster Response, Matias, Adole, Brown 08:15 AM Schedule
* Yossi Matias
Deza, Peterson, Murty,
* Tracy Adole 08:50 AM Opening Remarks
Griffiths
* Col Jason Brown
Page 25 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Nikolaus Kriegeskorte Plamen P Angelov, Nuria Oliver, Adrian Weller, Manuel Rodriguez,
03:55 PM Kriegeskorte
(Columbia) Isabel Valera, Silvia Chiappa, Hoda Heidari, Niki Kilbertus
Matthias Bethge (Universität
04:20 PM Bethge West 223 + 224, Fri Dec 13, 08:00 AM
Tübingen)
04:45 PM Eero Simoncelli (NYU) Simoncelli The growing field of Human-centric ML seeks to minimize the potential
harms, risks, and burdens of big data technologies on the public, and at
Panel Discussion: What the same time, maximize their societal benefits. In this workshop, we
sorts of cognitive or address a wide range of challenges from diverse, multi-disciplinary
biological (architectural) viewpoints. We bring together experts from a diverse set of backgrounds.
05:10 PM inductive biases will be Higgins, Konkle, Bethge Our speakers are leading experts in ML, human-computer interaction,
crucial for developing ethics, and law. Each of our speakers will focus on one core
effective artificial human-centred challenge (namely, fairness, accountability,
intelligence? interpretability, transparency, security, and privacy) in specific application
Concluding Remarks & Deza, Peterson, Murty, domains (such as medicine, welfare programs, governance, and
06:00 PM regulation). One of the main goals of this workshop is to help the
Prizes Ceremony Griffiths
community understand where it stands after a few years of rapid
06:10 PM Evening Reception technical development and identify promising research directions to
pursue in the years to come. Our speakers identify in their presentations
3-5 research directions that they consider to be of crucial importance.
Abstracts (5):
Page 26 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
These directions are further debated in one of our panel discussions. well as community detection in networks. Until recently, most algorithms
for solving inverse problems in the imaging and network sciences were
Schedule based on static signal models derived from physics or intuition, such as
wavelets or sparse representations.
04:00 PM Invited talk #5 Kim This workshop aims at bringing together theoreticians and practitioners
in order to chart out recent advances and discuss new directions in deep
Panel #2: Future research
neural network based approaches for solving inverse problems in the
directions and
04:30 PM imaging and network sciences.
interdisciplinary
collaborations in HCML
Schedule
Gu, Xiang, Kasirzadeh, Han,
Florez, Harder, Nguyen,
Heckel, Hand, Dimakis, Bruna,
Akhavan Rahnama, Donini, 08:30 AM Opening Remarks
Needell, Baraniuk
Slack, Ali, Koley, Bakker,
Hilgard, James-Sorenson, The spiked matrix model
08:40 AM Zdeborová
Ramos, Lu, Yang, with generative priors
05:00 PM Poster session Boyarskaya, Pawelczyk,
Robust One-Bit Recovery
Sokol, Jaiswal, Bhatt, Alvarez
via ReLU Generative
Melis, Grover, Marx, Yang,
09:10 AM Networks: Improved Qiu, Wei, Yang
Liang, Wang, Çapan, Wang,
Statistical Rate and Global
Grünewälder, Khajehnejad,
Landscape Analysis
Patro, Kunes, Deng, Liu,
Oneto, Li, Weber, Matthes, Tu 09:40 AM Coffee Break
Basis Decomposition of
11:00 AM Sapiro
Deep Learning
Solving inverse problems with deep networks: New
architectures, theoretical foundations, and applications Neural Reparameterization
Hoyer, Sohl-Dickstein,
11:30 AM Improves Structural
Greydanus
Reinhard Heckel, Paul Hand, Richard Baraniuk, Joan Bruna, Alex Optimization
Dimakis, Deanna Needell
12:00 PM Lunch Break
West 301 - 305, Fri Dec 13, 08:00 AM Learning-Based Low-Rank
02:00 PM Indyk
Approximations
There is a long history of algorithmic development for solving inverse
problems arising in sensing and imaging systems and beyond. Examples Blind Denoising,
include medical and computational imaging, compressive sensing, as 02:30 PM Self-Supervision, and Batson
Implicit Inverse Problems
Page 27 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Scarlett, Indyk, Vakilian, We study the robust one-bit compressed sensing problem whose goal is
Weller, Mitra, Aubin, Loureiro, to design an algorithm that faithfully recovers any sparse target vector
Krzakala, Zdeborová, $\theta_0\in\mathbb{R}^d$ \emph{uniformly} from $m$ quantized noisy
Monakhova, Yurtsever, measurements. Under the assumption that the measurements are
Waller, Sommerhoff, Moeller, sub-Gaussian, to recover any $k$-sparse $\theta_0$ ($k\ll d$)
Anirudh, Qiu, Wei, Yang, J. \emph{uniformly} up to an error $\varepsilon$ with high probability, the
Thiagarajan, Asif, Gillhofer, best known computationally tractable algorithm requires\footnote{Here,
Brandstetter, Hochreiter, an algorithm is ``computationally tractable'' if it has provable
Petersen, Patel, Oberai, convergence guarantees. The notation $\tilde{\mathcal{O}}(\cdot)$ omits
Kamath, Karmalkar, Price, a logarithm factor of $\varepsilon^{-1}$.} $m\geq\tilde{\mathcal{O}}(k\log
Ahmed, Kadkhodaie, Mohan, d/\varepsilon^4)$. In this paper, we consider a new framework for the
Simoncelli, one-bit sensing problem where the sparsity is implicitly enforced via
04:15 PM Poster Session Fernandez-Granda, Leong, mapping a low dimensional representation $x_0$ through a known
Sakla, Willett, Hoyer, $n$-layer ReLU generative network
Sohl-Dickstein, Greydanus, $G:\mathbb{R}^k\rightarrow\mathbb{R}^d$. Such a framework poses
Jagatap, Hegde, Kellman, low-dimensional priors on $\theta_0$ without a known basis. We propose
Tamir, Laanait, Dia, Ravanelli, to recover the target $G(x_0)$ via an unconstrained empirical risk
Binas, Rostamzadeh, Jalali, minimization (ERM) problem under a much weaker
Fang, Schwing, Lachapelle, \emph{sub-exponential measurement assumption}. For such a problem,
Brouillard, Deleu, we establish a joint statistical and computational analysis. In particular,
Lacoste-Julien, Yu, we prove that the ERM estimator in this new framework achieves an
Mazumdar, Rawat, Zhao, improved statistical rate of $m=\tilde{\mathcal{O}} (kn\log d /\epsilon^2)$
Chen, Li, Ramsauer, Rizzuti, recovering any $G(x_0)$ uniformly up to an error $\varepsilon$.
Mitsakos, Cao, Strohmer, Li, Moreover, from the lens of computation, despite non-convexity, we prove
Peng, Ongie that the objective of our ERM problem has no spurious stationary point,
that is, any stationary point is equally good for recovering the true target
up to scaling with a certain accuracy. Our analysis sheds some light on
Abstracts (7): the possibility of inverting a deep generative model under partial and
quantized measurements, complementing the recent success of using
Abstract 2: The spiked matrix model with generative priors in deep generative models for inverse problems.
Solving inverse problems with deep networks: New architectures,
theoretical foundations, and applications, Zdeborová 08:40 AM Abstract 5: Computational microscopy in scattering media in Solving
inverse problems with deep networks: New architectures,
Using a low-dimensional parametrization of signals is a generic and theoretical foundations, and applications, Waller 10:30 AM
powerful way to enhance performance in signal processing and statistical
inference. A very popular and widely explored type of dimensionality Computational imaging involves the joint design of imaging system
reduction is sparsity; another type is generative modelling of signal hardware and software, optimizing across the entire pipeline from
distributions. Generative models based on neural networks, such as acquisition to reconstruction. Computers can replace bulky and
GANs or variational auto-encoders, are particularly performant and are expensive optics by solving computational inverse problems. This talk
gaining on applicability. In this paper we study spiked matrix models, will describe new microscopes that use computational imaging to enable
where a low-rank matrix is observed through a noisy channel. This 3D fluorescence and phase measurement using image reconstruction
problem with sparse structure of the spikes has attracted broad attention algorithms that are based on large-scale nonlinear non-convex
in the past literature. Here, we replace the sparsity assumption by optimization combined with unrolled neural networks. We further discuss
generative modelling, and investigate the consequences on statistical engineering of data capture for computational microscopes by
and algorithmic properties. We analyze the Bayes-optimal performance end-to-end learned design.
under specific generative models for the spike. In contrast with the
sparsity assumption, we do not observe regions of parameters where Abstract 6: Basis Decomposition of Deep Learning in Solving
statistical performance is superior to the best known algorithmic inverse problems with deep networks: New architectures,
performance. We show that in the analyzed cases the approximate theoretical foundations, and applications, Sapiro 11:00 AM
message passing algorithm is able to reach optimal performance. We
Ordinary convolutional neural networks (CNNs) learn non-parametric
also design enhanced spectral algorithms and analyze their performance
filters, applied in multiple leyers, leading to to need to learn tens of
and thresholds using random matrix theory, showing their superiority to
millions
the classical principal component analysis. We complement our
of variables with large training data. In this talk we show how such filters
theoretical results by illustrating the performance of the spectral
can be replaced by basis, not only reducing the number of parameters
algorithms when the spikes come from real datasets.
and needed training samples by orders of magnitudes but also
Abstract 3: Robust One-Bit Recovery via ReLU Generative Networks: intrinsically and naturally achieving invariance, domain adaptation, and
Improved Statistical Rate and Global Landscape Analysis in Solving stochasticity.
Page 28 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abstract 7: Neural Reparameterization Improves Structural West 306, Fri Dec 13, 08:00 AM
Optimization in Solving inverse problems with deep networks: New
architectures, theoretical foundations, and applications, Hoyer, A new wave of intelligent computing, driven by recent advances in
Sohl-Dickstein, Greydanus 11:30 AM machine learning and cognitive algorithms coupled with process
technology and new design methodologies, has the potential to usher
Structural optimization is a popular method for designing objects such as unprecedented disruption in the way modern computing systems are
bridge trusses, airplane wings, and optical devices. Unfortunately, the designed and deployed. These new and innovative approaches often
quality of solutions depends heavily on how the problem is provide an attractive and efficient alternative not only in terms of
parameterized. In this paper, we propose using the implicit bias over performance but also power, energy, and area. This disruption is easily
functions induced by neural networks to improve the parameterization of visible
structural optimization. Rather than directly optimizing densities on a across the whole spectrum of computing systems -- ranging from low end
grid, we instead optimize the parameters of a neural network which mobile devices to large scale data centers and servers including
outputs those densities. This reparameterization leads to different and intelligent infrastructures.
often better solutions. On a selection of 116 structural optimization tasks,
our approach produces an optimal design 50% more often than the best A key class of these intelligent solutions is providing real-time, on-device
baseline method. cognition at the edge to enable many novel applications including
computer vision and image processing, language understanding, speech
Abstract 10: Blind Denoising, Self-Supervision, and Implicit Inverse and gesture recognition, malware detection and autonomous driving.
Problems in Solving inverse problems with deep networks: New Naturally, these applications have diverse requirements for performance,
architectures, theoretical foundations, and applications, Batson energy, reliability, accuracy, and security that demand a holistic
02:30 PM approach to designing the hardware, software, and
intelligence algorithms to achieve the best power, performance, and area
We will discuss a self-supervised approach to the foundational inverse (PPA).
problem of denoising (Noise2Self). By taking advantage of statistical
independence in the noise, we can estimate the mean-square error for a Topics:
large class of deep architectures without access to ground truth. This - Architectures for the edge: IoT, automotive, and mobile
allows us to train a neural network to denoise from noisy data alone, and - Approximation, quantization reduced precision computing
also to compare between architectures, selecting one which will produce - Hardware/software techniques for sparsity
images with the lowest MSE. However, architectures with the same MSE - Neural network architectures for resource constrained devices
performance can produce qualitatively different results, i.e., the - Neural network pruning, tuning and and automatic architecture search
hypersurface of images with fixed MSE is very heterogeneous. We will - Novel memory architectures for machine learning
discuss ongoing work in understanding the types of artifacts which - Communication/computation scheduling for better performance and
different denoising architectures give rise to. energy
- Load balancing and efficient task distribution techniques
Abstract 11: Learning Regularizers from Data in Solving inverse
- Exploring the interplay between precision, performance, power and
problems with deep networks: New architectures, theoretical
energy
foundations, and applications, Chandrasekaran 03:00 PM
- Exploration of new and efficient applications for machine learning
- Characterization of machine learning benchmarks and workloads
Regularization techniques are widely employed in the solution of
- Performance profiling and synthesis of workloads
inverse problems in data analysis and scientific computing due to
- Simulation and emulation techniques, frameworks and platforms for
their effectiveness in addressing difficulties due to ill-posedness.
machine learning
In their most common manifestation, these methods take the form of
- Power, performance and area (PPA) based comparison of neural
penalty functions added to the objective in variational approaches for
networks
solving inverse problems. The purpose of the penalty function is to
- Verification, validation and determinism in neural networks
induce a desired structure in the solution, and these functions are
- Efficient on-device learning techniques
specified based on prior domain-specific expertise. We consider the
- Security, safety and privacy challenges and building secure AI systems
problem of learning suitable regularization functions from data in
settings in which precise domain knowledge is not directly available;
Schedule
the objective is to identify a regularizer to promote the type of
structure contained in the data. The regularizers obtained using our
framework are specified as convex functions that can be computed 08:00 AM TBD LeCun
efficiently via semidefinite programming. Our approach for learning
Efficient Computing for AI
such semidefinite regularizers combines recent techniques for rank 08:45 AM Sze
and Robotics
minimization problems along with the Operator Sinkhorn procedure.
(Joint work with Yong Sheng Soh)
Page 29 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
will describe how joint algorithm and hardware design can be used to
Abandoning the Dark Arts:
reduce energy consumption while delivering real-time and robust
09:30 AM New Directions in Efficient Keutzer
performance for applications including deep learning, computer vision,
DNN Design
autonomous navigation/exploration and video/image processing. We will
Spasov, Nayak, Diego Andilla, show how energy-efficient techniques that exploit correlation and
10:00 AM Poster Session 1
Zhang, Trivedi sparsity to reduce compute, data movement and storage costs can be
applied to various tasks including image classification, depth estimation,
Putting the “Machine” Back
super-resolution, localization and mapping.
in Machine Learning: The
10:30 AM Marculescu
Case for Hardware-ML Abstract 3: Abandoning the Dark Arts: New Directions in Efficient
Model Co-design DNN Design in EMC2: Energy Efficient Machine Learning and
Adaptive Multi-Task Neural Cognitive Computing (5th edition), Keutzer 09:30 AM
11:00 AM Networks for Efficient Feris
Deep Neural Net models have provided the most accurate solutions to a
Inference
very wide variety of problems in vision, language, and speech; however,
Yu, Hartmann, Li, Shafiee, the design, training, and optimization of efficient DNNs typically requires
11:30 AM Oral Session 1
Yang, Zafrir resorting to the “dark arts” of ad hoc methods and extensive
hyperparameter tuning. In this talk we present our progress on
12:00 PM Qualcomm Industry Talk Lee
abandoning these dark arts by using Differential Neural Architecture
12:30 PM Lunch Search to guide the design of efficient DNNs and by using
Hessian-based methods to guide the processes of training and
Cheap, Fast, and Low Power
quantizing those DNNs.
02:00 PM Deep Learning: I need it Delp
now!
Abstract 5: Putting the “Machine” Back in Machine Learning: The
Advances and Prospects for Case for Hardware-ML Model Co-design in EMC2: Energy Efficient
02:45 PM Verma Machine Learning and Cognitive Computing (5th edition),
In-memory Computing
Marculescu 10:30 AM
Algorithm-Accelerator
03:15 PM Co-Design for Neural Zhang Machine learning (ML) applications have entered and impacted our lives
Network Specialization unlike any other technology advance from the recent past. Indeed,
almost every aspect of how we live or interact with others relies on or
Prato, Thakker, Galindez
uses ML for applications ranging from image classification and object
03:45 PM Poster Session 2 Olascoaga, Zhang, Partovi
detection, to processing multi■modal and heterogeneous datasets.
Nia, Adamczewski
While the holy grail for judging the quality of a ML model has largely
Efficient Algorithms to been serving accuracy, and only recently its resource usage, neither of
04:15 PM Accelerate Deep Learning Han these metrics translate directly to energy efficiency, runtime, or mobile
on Edge Devices device battery lifetime. This talk will uncover the need for building
accurate, platform■specific power and latency models for convolutional
Liao, McKinstry, Izsak, Li,
04:45 PM Oral Session 2 neural networks (CNNs) and efficient hardware-aware CNN design
Huang, Mordido
methodologies, thus allowing machine learners and hardware designers
05:30 PM Microsoft Industry Talk Darvish Rouhani to identify not just the best accuracy NN configuration, but also those that
satisfy given hardware constraints. Our proposed modeling framework is
06:00 PM LPCVC Results
applicable to both high■end and mobile platforms and achieves 88.24%
accuracy for latency, 88.34% for power, and 97.21% for energy
prediction. Using similar predictive models, we demonstrate a novel
Abstracts (9):
differentiable neural architecture search (NAS) framework, dubbed
Abstract 1: TBD in EMC2: Energy Efficient Machine Learning and Single-Path NAS, that uses one single-path over-parameterized CNN to
Cognitive Computing (5th edition), LeCun 08:00 AM encode all architectural decisions based on shared convolutional kernel
parameters. Single-Path NAS achieves state-of-the-art top-1 ImageNet
TBD accuracy (75.62%), outperforming existing mobile NAS methods for
similar latency constraints (∼80ms) and finds the final configuration up to
Abstract 2: Efficient Computing for AI and Robotics in EMC2: Energy 5,000× faster compared to prior work. Combined with our quantized
Efficient Machine Learning and Cognitive Computing (5th edition), CNNs (Flexible Lightweight CNNs or FLightNNs) that customize
Sze 08:45 AM precision level in a layer-wise fashion and achieve almost iso-accuracy
at 5-10x energy reduction, such a modeling, analysis, and optimization
Computing near the sensor is preferred over the cloud due to privacy framework is poised to lead to true co-design of hardware and ML model,
and/or latency concerns for a wide range of applications including orders of magnitude faster than state of the art, while satisfying both
robotics/drones, self-driving cars, smart Internet of Things, and accuracy and latency or energy constraints.
portable/wearable electronics. However, at the sensor there are often
stringent constraints on energy consumption and cost in addition to the Abstract 6: Adaptive Multi-Task Neural Networks for Efficient
throughput and accuracy requirements of the application. In this talk, we Inference in EMC2: Energy Efficient Machine Learning and
Page 30 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Cognitive Computing (5th edition), Feris 11:00 AM hardware accelerators in both academic and commercial settings. In line
with this trend, there has been an active body of research on both
Very deep convolutional neural networks have shown remarkable algorithms and hardware architectures for neural network specialization.
success in many computer vision tasks, yet their computational expense
limits their impact in domains where fast inference is essential. While This talk presents our recent investigation into DNN optimization and
there has been significant progress on model compression and low-precision quantization, using a co-design approach featuring
acceleration, most methods rely on a one-size-fits-all network, where the contributions to both algorithms and hardware accelerators. First, we
same set of features is extracted for all images or tasks, no matter their review static network pruning techniques and show a fundamental link
complexity. In this talk, I will first describe an approach called BlockDrop, between group convolutions and circulant matrices – two previously
which learns to dynamically choose which layers of a deep network to disparate lines of research in DNN compression. Then we discuss
execute during inference, depending on the image complexity, so as to channel gating, a dynamic, fine-grained, and trainable technique for DNN
best reduce total computation without degrading prediction accuracy. acceleration. Unlike static approaches, channel gating exploits
Then, I will show how this approach can be extended to design compact input-dependent dynamic sparsity at run time. This results in a significant
multi-task networks, where a different set of layers is executed reduction in compute cost with a minimal impact on accuracy. Finally, we
depending on the task complexity, and the level of feature sharing across present outlier channel splitting, a technique to improve DNN weight
tasks is automatically determined to maximize both the accuracy and quantization by removing outliers from the weight distribution without
efficiency of the model. Finally, I will conclude the talk presenting an retraining.
efficient multi-scale neural network model, which achieves state-of-the
art results in terms of accuracy and FLOPS reduction on standard Abstract 14: Efficient Algorithms to Accelerate Deep Learning on
benchmarks such as the ImageNet dataset. Edge Devices in EMC2: Energy Efficient Machine Learning and
Cognitive Computing (5th edition), Han 04:15 PM
Abstract 10: Cheap, Fast, and Low Power Deep Learning: I need it
now! in EMC2: Energy Efficient Machine Learning and Cognitive Efficient deep learning computing requires algorithm and hardware
Computing (5th edition), Delp 02:00 PM co-design to enable specialization. However, the extra degree of
freedom creates a much larger design space. We propose AutoML
In this talk I will describe the need for low power machine learning techniques to architect efficient neural networks. We investigate
systems. I will motivate this by describing several current projects at automatically designing small and fast models (ProxylessNAS), auto
Purdue University that have a need for energy efficient deep learning channel pruning (AMC), and auto mixed-precision quantization (HAQ).
and in some cases the real deployment of these methods will not be We demonstrate such learning-based, automated design achieves
possible without lower power solutions. The applications include superior performance and efficiency than rule-based human design.
precision farming, health care monitoring, and edge-based surveillance. Moreover, we shorten the design cycle by 200× than previous work to
efficiently search efficient models, so that we can afford to design
Abstract 11: Advances and Prospects for In-memory Computing in specialized neural network models for different hardware platforms. We
EMC2: Energy Efficient Machine Learning and Cognitive Computing accelerate computation-intensive AI applications including (TSM) for
(5th edition), Verma 02:45 PM efficient video recognition and PVCNN for efficient 3D recognition on
point clouds. Finally, we’ll describe scalable distributed training and the
Edge AI applications retain the need for high-performing inference
potential security issues of efficient deep learning.
models, while driving platforms beyond their limits of energy efficiency
and throughput. Digital hardware acceleration, enabling 10-100x gains
over general-purpose architectures, is already widely deployed, but is
Machine Learning for Health (ML4H): What makes machine
ultimately restricted by data-movement and memory accessing that
dominates deep-learning computations. In-memory computing, based on
learning in medicine different?
both SRAM and emerging memory, offers fundamentally new tradeoffs
Andrew Beam, Tristan Naumann, Brett Beaulieu-Jones, Irene Y
for overcoming these barriers, with the potential for 10x higher energy
Chen, Sam Finlayson, Emily Alsentzer, Adrian Dalca, Matthew
efficiency and area-normalized throughput demonstrated in recent
McDermott
designs. But, those tradeoffs instate new challenges, especially affecting
scaling to the level of computations required, integration in practical
West Ballroom A, Fri Dec 13, 08:00 AM
heterogeneous architectures, and mapping of diverse software. This talk
examines those tradeoffs to characterize the challenges. It then explores The goal of the NeurIPS 2019 Machine Learning for Health Workshop
recent research that provides promising paths forward, making (ML4H) is to foster collaborations that meaningfully impact medicine by
in-memory computing more of a practical reality than ever before. bringing together clinicians, health data experts, and machine learning
researchers. Attendees at this workshop can also expect to broaden their
Abstract 12: Algorithm-Accelerator Co-Design for Neural Network
network of collaborators to include clinicians and machine learning
Specialization in EMC2: Energy Efficient Machine Learning and
researchers who are focused on solving some of the most import
Cognitive Computing (5th edition), Zhang 03:15 PM
problems in medicine and healthcare. The organizers of this proposal
have successfully run NeurIPS workshops in the past and are
In recent years, machine learning (ML) with deep neural networks
well-equipped to run this year’s workshop should this proposal be
(DNNs) has been widely deployed in diverse application domains.
accepted.
However, the growing complexity of DNN models, the slowdown of
technology scaling, and the proliferation of edge devices are driving a
This year’s theme of “What makes machine learning in medicine
demand for higher DNN performance and energy efficiency. ML
different?” aims to elucidate the obstacles that make the development of
applications have shifted from general-purpose processors to dedicated
Page 31 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Attendees at the workshop will gain an appreciation for problems that are
unique to the application of machine learning for healthcare and a better Meta-Learning
understanding of how machine learning techniques may be leveraged to
solve important clinical problems. This year’s workshop builds on the last Roberto Calandra, Ignasi Clavera Gilaberte, Frank Hutter, Joaquin
two NeurIPS ML4H workshops, which were both attended by more than Vanschoren, Jane Wang
500 people each year, and helped form the foundations of an emerging
research community. West Ballroom B, Fri Dec 13, 08:00 AM
Please see the attached document for the full program. Recent years have seen rapid progress in metalearning methods, which
learn (and optimize) the performance of learning methods based on data,
Schedule generate new learning methods from scratch, and learn to transfer
knowledge across tasks and domains. Metalearning can be seen as the
logical conclusion of the arc that machine learning has undergone in the
08:45 AM Daphne Koller Talk
last decade, from learning classifiers, to learning representations, and
09:15 AM Emily Fox Talk Fox finally to learning algorithms that themselves acquire representations and
classifiers. The ability to improve one’s own learning capabilities through
10:15 AM Luke Oakden-Rayner Talk Oakden-Rayner
experience can also be viewed as a hallmark of intelligent beings, and
10:45 AM Paper spotlight talks there are strong connections with work on human learning in
neuroscience. The goal of this workshop is to bring together researchers
Zheng, Kapur, Asif,
from all the different communities and topics that fall under the umbrella
Rozenberg, Gilet, Sidorov,
of metalearning. We expect that the presence of these different
Kumar, Van Steenkiste, Boag,
communities will result in a fruitful exchange of ideas and stimulate an
Ouyang, Jaeger, Liu,
open discussion about the current challenges in metalearning, as well as
Balagopalan, Rajan, Skreta,
possible solutions.
Pattisapu, Goschenhofer,
Prabhu, Jin, Gardiner, Li, Schedule
kumar, Hu, Motani, Lovelace,
Roshan, Wang, Valmianski,
Lee, Mallya, Chaibub Neto, 09:10 AM Invited Talk 1 Abbeel
Kemp, Charpignon, Nigam,
09:40 AM Invited Talk 2 Clune
Weng, Boughorbel, Bellot,
Gondara, Zhang, Bahadori, 10:10 AM Poster Spotlights 1
Zech, Shao, Choi,
Takagi, Javed, Sommer,
Seyyed-Kalantari, Aiken, Bica,
Sharaf, D'Oro, Wei, Doveh,
11:15 AM Poster Session I Shen, Chin-Cheong, Roy,
White, Gonzalez, Nguyen, li,
Baldini, Min, Deschrijver,
Yu, Ramalho, Nomura, Alvi,
Marttinen, Pascual Ortiz,
Ton, Huang, Lee, Flennerhag,
Nagesh, Rindtorff, Mulyar, 10:30 AM Coffee/Poster session 1
Zhang, Friesen, Blomstedt,
Hoebel, Shaka, Machart,
Dubatovka, Bartunov, Yi,
Gatys, Ng, Hüser, Taylor,
Shcherbatyi, Simon, Shang,
Barbour, Martinez, McCreery,
MacLeod, Liu, Fowl, Parente
Eyre, Natarajan, Yi, Ma,
Paiva Mesquita, Quillen
Nagpal, Du, Gao, Tuladhar,
Shleifer, Ren, Mashouri, Lu, 11:30 AM Invited Talk 3 Grant
Bagherzadeh-Khiabani,
12:00 PM Discussion 1
Choudhury, Raghu, Fleming,
Jain, YANG, Harley, Pfohl, 02:00 PM Invited Talk 4 Abel
Rumetshofer, Fedorov, Dash,
02:30 PM Invited Talk 5 Hadsell
Pfau, Tomkins, Targonski,
Brudno, Li, Yu, Patel 03:00 PM Poster Spotlights 2
Page 32 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Song, Mangla, Salinas, 09:15 AM Invited Talk #1: Jane Wang Wang
Zhuang, Feng, Hu, Puri,
Mohinta, Agostinelli,
Maddox, Raghu, Tossou, Yin,
Moringen, Lee, Lo, Maass,
Dasgupta, Lee, Alet, Xu,
Sheffer, Bredenberg,
03:20 PM Coffee/Poster session 2 Franke, Harrison, Warrell,
Eysenbach, Xia, Markou,
Dhillon, Zela, Qiu, Siems,
Lichtenberg, Richemond,
Mendonca, Schlessinger, Li, Coffee Break & Poster
09:45 AM Zhang, Lanier, Lin, Fedus,
Manolache, Dutta, Glass, Session
Berseth, Sarrico, Crosby,
Singh, Koehler
McAleer, Ghiassian, Scherr,
04:30 PM Contributed Talk 1 Bellec, Salaj, Kolbeinsson,
Rosenberg, Shin, Lee, Cecchi,
04:45 PM Contributed Talk 2
Rish, Hajek
05:00 PM Invited Talk 6 Lake
Contributed Talk #1:
05:30 PM Discussion 2 Humans flexibly transfer
10:30 AM Xia
options at multiple levels of
abstractions
Page 33 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abstract 4: Contributed Talk #1: Humans flexibly transfer options at find that, in humans and monkeys, information sampling is partially
multiple levels of abstractions in Biological and Artificial sensitive to uncertainty but is also biased by Pavlovian tendencies, which
Reinforcement Learning, Xia 10:30 AM push agents to engage with signals predicting positive outcomes and
avoid those predicting negative outcomes in ways that interfere with a
Humans are great at using prior knowledge to solve novel tasks, but how reduction of uncertainty. In a second paradigm, agents are given several
they do so is not well understood. Recent work showed that in contextual tasks of different difficulty and can freely organize their exploration in
multi-armed bandits environments, humans create simple one-step order to learn. In these contexts, uncertainty-based heuristics become
policies that they can transfer to new contexts by inferring context ineffective, and optimal strategies are instead based on learning
clusters. However, the daily tasks humans face are often temporally progress – the ability to first engage with and later reduce uncertainty. I
extended, and demand more complex, hierarchically structured skills. will show evidence that humans are motivated to select difficult tasks
The options framework provides a potential solution for representing consistent with learning maximization, but they guide their task selection
such transferable skills. Options are abstract multi-step policies, according to success rates rather than learning progress per se, which
assembled from simple actions or other options, that can represent risks trapping them in tasks with too high levels of difficulty (e.g., random
meaningful reusable skills. We developed a novel two-stage decision unlearnable tasks). Together, the results show that information demand
making protocol to test if humans learn and transfer multi-step options. has consistent features that can be quantitatively measured at various
We found transfer effects at multiple levels of policy complexity that levels of complexity, and a research agenda exploring these features will
could not be explained by flat reinforcement learning models. We also greatly expand our understanding of complex decision strategies.
devised an option model that can qualitatively replicate the transfer
effects in human participants. Our results provide evidence that humans Abstract 14: Contributed Talk #3 MEMENTO: Further Progress
create options, and use them to explore in novel contexts, consequently Through Forgetting in Biological and Artificial Reinforcement
transferring past knowledge and speeding up learning. Learning, Fedus 04:15 PM
Abstract 5: Contributed Talk #2: Slow processes of neurons enable a Modern Reinforcement Learning (RL) algorithms, even those with
biologically plausible approximation to policy gradient in Biological intrinsic reward bonuses, suffer performance plateaus in hard-exploration
and Artificial Reinforcement Learning, Maass 10:45 AM domains suggesting these algorithms have reached their ceiling.
However, in what we describe as the MEMENTO observation, we find
Recurrent neural networks underlie the astounding information that new agents launched from the position where the previous agent
processing capabilities of the brain, and play a key role in many saturated, can reliably make further progress. We show that this is not an
state-of-the-art algorithms in deep reinforcement learning. But it has artifact of limited model capacity or training duration, but rather indicative
remained an open question how such networks could learn from rewards of interference in learning dynamics between various stages of the
in a biologically plausible manner, with synaptic plasticity that is both domain [Schaul et al., 2019], signatures of multi-task and continual
local and online. We describe such an algorithm that approximates learning. To mitigate interference we design an end-to-end learning
actor-critic policy gradient in recurrent neural networks. Building on an agent which partitions the environment into various segments, and
approximation of backpropagation through time (BPTT): e-prop, and models the value function separately in each score context per Jain et al.
using the equivalence between forward and backward view in [2019]. We demonstrate increased learning performance by this
reinforcement learning (RL), we formulate a novel learning rule for RL ensemble of agents on Montezuma’s Revenge and further show how this
that is both online and local, called reward-based e-prop. This learning ensemble can be distilled into a single agent with the same model
rule uses neuroscience inspired slow processes and top-down signals, capacity as the original learner. Since the solution is empirically
while still being rigorously derived as an approximation to actor-critic expressible by the original network, this provides evidence of
policy gradient. To empirically evaluate this algorithm, we consider a interference and our approach validates an avenue to circumvent it.
delayed reaching task, where an arm is controlled using a recurrent
network of spiking neurons. In this task, we show that reward-based
e-prop performs as well as an agent trained with actor-critic policy Graph Representation Learning
gradient with biologically implausible BPTT.
Will Hamilton, Rianne van den Berg, Michael Bronstein, Stefanie
Abstract 6: Invited Talk 2: Understanding information demand at
Jegelka, Thomas Kipf, Jure Leskovec, Renjie Liao, Yizhou Sun,
different levels of complexity in Biological and Artificial
Petar Veli■kovi■
Reinforcement Learning, Gottlieb 11:00 AM
West Exhibition Hall A, Fri Dec 13, 08:00 AM
In the 1950s, Daniel Berlyne wrote extensively about the importance of
curiosity – our intrinsic desire to know. To understand curiosity, Berlyne Graph-structured data is ubiquitous throughout the natural and social
argued, we must explain why humans exert so much effort to obtain sciences, from telecommunication networks to quantum chemistry.
knowledge, and how they decide which questions to explore, given that Building relational inductive biases into deep learning architectures is
exploration is difficult and its long-term benefits are impossible to crucial if we want systems that can learn, reason, and generalize from
ascertain. I propose that these questions, although relatively neglected in this kind of data. Furthermore, graphs can be seen as a natural
neuroscience research, are key to understanding cognition and complex generalization of simpler kinds of structured data (such as images), and
decision making of the type that humans routinely engage in and therefore, they represent a natural avenue for the next breakthroughs in
autonomous agents only aspire to. I will describe our investigations of machine learning.
these questions in two types of paradigms. In one paradigm, agents are
placed in contexts with different levels of uncertainty and reward Recent years have seen a surge in research on graph representation
probability and can sample information about the eventual outcome. We learning, including techniques for deep graph embeddings,
Page 34 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Jamadandi, Sanborn, Yao, Neural architectures and many learning environments can conveniently
Cai, Chen, Andreoli, Stoehr, be expressed by
Su, Duan, Ferreira, Belli, graphs. Interestingly, it has been recently shown that the notion of
Boyarski, Ye, Ghalebi, Sarkar, receptive field and the
KHADEMI, Faerman, Bose, correspondent convolutional computation can nicely be extended to
11:30 AM Poster Session #1 Ma, Meng, Kazemi, Wang, graph-based data domains
Wu, Wu, Joshi, Brockschmidt, with successful results. On the other hand, graph neural networks (GNN)
Zambon, Graber, Van Belle, were introduced by
Malik, Glorot, Krenn, extending the notion of time-unfolding, which ended up into a
Cameron, Huang, Stoica, state-based representation along
Toumpa with a learning process that requires state relaxation to a fixed-point. It
turns out that
12:30 PM Lunch
algorithms based on this approach applied to learning tasks on
Outstanding Contribution collections of graphs are more
01:30 PM Talk: Pre-training Graph Hu computationally expensive than recent graph convolutional nets.
Neural Networks
In this talk we advocate the importance of refreshing state-based graph
Page 35 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 36 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 37 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Real Neurons & Hidden Units: future directions at the 11:45 AM Contributed Talk #3
intersection of neuroscience and AI 12:00 PM Lunch Break
Guillaume Lajoie, Eli Shlizerman, Maximilian Puelma Touzel, 02:00 PM Ila's Talk Fiete
Jessica Thompson, Konrad Kording
02:30 PM Surya's Talk Ganguli
Recent years have witnessed an explosion of progress in AI. With it, a 03:15 PM Contributed Talk #5
proliferation of experts and practitioners are pushing the boundaries of 03:30 PM Coffee Break + Posters
the field without regard to the brain. This is in stark contrast with the
field's transdisciplinary origins, when interest in designing intelligent Sainath, Akrout, Delahunt,
algorithms was shared by neuroscientists, psychologists and computer Kutz, Yang, Marino, Abbott,
scientists alike. Similar progress has been made in neuroscience where Vecoven, Ernst, warrington, ,
novel experimental techniques now afford unprecedented access to Kagan, Cho, Harris, Grinberg,
brain activity and function. However, it is unclear how to maximize them Hopfield, Krotov, Muhammad,
to truly advance an end-to-end understanding of biological intelligence. Cobos, Walker, Reimer,
The traditional neuroscience research program, however, lacks Tolias, Ecker, Sheth, Zhang,
frameworks to truly advance an end-to-end understanding of biological Wo■czyk, Tabor, Maszke,
intelligence. For the first time, mechanistic discoveries emerging from Pogodin, Corneil, Gerstner,
deep learning, reinforcement learning and other AI fields may be able to Lin, Cecchi, Reinen, Rish,
steer fundamental neuroscience research in ways beyond standard uses Bellec, Salaj, Subramoney,
of machine learning for modelling and data analysis. For example, Maass, Wang, Pakman, Lee,
successful training algorithms in artificial networks, developed without Paninski, Tripp, Graber,
biological constraints, can motivate research questions and hypotheses Schwing, Prince, Ocker,
about the brain. Conversely, a deeper understanding of brain Buice, Lansdell, Kording,
computations at the level of large neural populations may help shape Lindsey, Sejnowski, Farrell,
future directions in AI. This workshop aims to address this novel situation Shea-Brown, Farrugia,
04:15 PM Poster Session
by building on existing AI-Neuro relationships but, crucially, outline new Nepveu, Im, Branson, Hu,
directions for artificial systems and next-generation neuroscience Iyer, Mihalas, Aenugu, Hazan,
experiments. We invite contributions concerned with the modern Dai, Nguyen, Tsao, Baraniuk,
intersection between neuroscience and AI and in particular, addressing Anandkumar, Tanaka, Nayebi,
questions that can only now be tackled due to recent progress in AI on Baccus, Ganguli, Pospisil,
the role of recurrent dynamics, inductive biases to guide learning, global Muller, Cheng, Varoquaux,
versus local learning rules, and interpretability of network activity. This Dadi, Gklezakos, Rao, Louis,
workshop will promote discussion and showcase diverse perspectives on Papadimitriou, Vempala,
these open questions. Yadati, Zdeblick, Witten,
Roberts, Prabhu, Bellec,
Schedule Ramesh, Macke, Cadena,
Bellec, Scherr, Marschall, Kim,
Rapp, Fonseca, Armitage, Im,
Lajoie, Thompson, Puelma
08:15 AM Opening Remarks Hardcastle, Sharma, Bair,
Touzel, Shlizerman, Kording
Valente, Shang, Stern, Patil,
Learning to be surprised - Wang, Gorantla, Stratton,
evidence for emergent Edwards, Lu, Ester, Vlasov
08:30 AM Richards
surprise responses visual
05:00 PM Doina's Talk Precup
cortex
Bengio, Richards, Lillicrap,
09:00 AM Tim's Talk Lillicrap Panel Session: A new hope
05:30 PM Fiete, Sussillo, Precup,
for neuroscience
09:30 AM Contributed Talk #1 Kording, Ganguli
Page 38 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Fair ML in Healthcare David Rolnick, Priya Donti, Lynn Kaack, Alexandre Lacoste, Tegan
Maharaj, Andrew Ng, John Platt, Jennifer Chayes, Yoshua Bengio
Shalmali Joshi, Irene Y Chen, Ziad Obermeyer, Sendhil Mullainathan
East Ballroom C, Sat Dec 14, 08:00 AM
East Ballroom B, Sat Dec 14, 08:00 AM
Climate change is one of the greatest problems society has ever faced,
Clinical healthcare has been a natural application domain for ML with a with increasingly severe consequences for humanity as natural disasters
few modest success stories of practical deployment. Inequity and multiply, sea levels rise, and ecosystems falter. Since climate change is
healthcare disparity has long been a concern in clinical and public health a complex issue, action takes many forms, from designing smart electric
for decades. However, the challenges of fair and equitable care using ML grids to tracking greenhouse gas emissions through satellite imagery.
in health has largely remained unexplored. While a few works have While no silver bullet, machine learning can be an invaluable tool in
attempted to highlight potential concerns and pitfalls in recent years, fighting climate change via a wide array of applications and techniques.
there are massive gaps in academic ML literature in this context. The These applications require algorithmic innovations in machine learning
goal of this workshop is to investigate issues around fairness that are and close collaboration with diverse fields and practitioners. This
specific to ML based healthcare. We hope to investigate a myriad of workshop is intended as a forum for those in the machine learning
questions via the workshop. community who wish to help tackle climate change.
Schedule Schedule
Page 39 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Sun, Veeramachaneni,
The Workshop on AI for Social Good addresses these challenges by Ramirez Diaz, Cuesta-Infante,
bringing together machine learning researchers, social impact leaders, 04:00 PM Track 3: Public Policy
Elzayn, Gamper, Schim van
ethicists, and public policy leaders to present their ideas and applications der Loeff, Green
for maximizing the social good. This workshop is a collaboration of three
formerly separate lines of research (i.e., this is a "joint" workshop),
including researchers in applications-driven AI research, applied ethics,
and AI policy. Each of these research areas are unified into a 3-track
Machine Learning for Autonomous Driving
framework promoting the exchange of ideas between the practitioners of
each track.
Rowan McAllister, Nick Rhinehart, Fisher Yu, Li Erran Li, Anca
Dragan
We hope that this gathering of research talent will inspire the creation of
new approaches and tools, provide for the development of intelligent East Meeting Rooms 1 - 3, Sat Dec 14, 08:00 AM
systems benefiting all stakeholders, and converge on public policy
mechanisms for encouraging these goals. Autonomous vehicles (AVs) provide a rich source of high-impact
research problems for the machine learning (ML) community at NeurIPS
Schedule in diverse fields including computer vision, probabilistic modeling,
gesture recognition, pedestrian and vehicle forecasting, human-machine
interaction, and multi-agent planning. The common goal of autonomous
08:00 AM Opening remarks Bengio
driving can catalyze discussion between these subfields, generating a
Dietterich, Gomes, cross-pollination of research ideas. Beyond the benefits to the research
Track 1: Producing Good
08:05 AM Luengo-Oroz, Dilkina, community, AV research can improve society by reducing road
Outcomes
Cornebise accidents; giving independence to those unable to drive; and inspiring
younger generations towards ML with tangible examples of ML-based
10:30 AM Break
technology clearly visible on local streets.
As many NeurIPS attendees are key drivers behind AV-applied ML, the
proposed NeurIPS 2019 Workshop on Autonomous Driving intends to
bring researchers together from both academia and industries to discuss
machine learning applications in autonomous driving. Our proposal
includes regular paper presentations, invited speakers, and technical
Page 40 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
benchmark challenges to present the current state of the art, as well as Borja Balle, Kamalika Chaudhuri, Antti Honkela, Antti Koskela,
the limitations and future directions for autonomous driving. Casey Meehan, Mi Jung Park, Mary Anne Smart, Adrian Weller
Page 41 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abstract 3: Poster Session in Privacy in Machine Learning (PriML), stage for this workshop.
Canonne, Jun, Neel, Wang, vietri, Song, Lebensold, Zhang, Gondara, Li,
Mireshghallah, Dong, Sarwate, Koskela, Jälkö, Kusner, Chen, Park, Schedule
Machanavajjhala, Kalpathy-Cramer, , Feldman, Tomkins, Phan,
Esfandiari, Jaiswal, Sharma, Druce, Meehan, Zhao, Hsu, Railsback,
Baydin, Carrasquilla, Ho,
Flaxman, , Adebayo, Korolova, Xu, Holohan, Basu, Joseph, Thai, Yang,
Kashinath, Paganini, Thais,
Vitercik, Hutchinson, Wang, Yauney, Tao, Jin, Lee, McMillan, Izmailov, 08:10 AM Opening Remarks
Anandkumar, Cranmer, Melko,
Guo, Swaroop, Orekondy, Esmaeilzadeh, Procopio, Polyzotis,
Prabhat, Wood
Mohammadi, Agrawal 11:30 AM
08:20 AM Bernhard Schölkopf Schölkopf
Schedule is not final. Details TBA.
Towards physics-informed
Abstract 4: TBA: Lalitha Sankar in Privacy in Machine Learning 09:00 AM deep learning for turbulent Yu
(PriML), Sankar 02:00 PM flow prediction
Page 42 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Topics:
Komkov, Fort, Wang, Yu,
—Abstractions and syntax (beyond meta-programming and operator
Park, Schoenholz, Cheng,
overloading) to naturally express a program (expression, or procedure)
Griffiths, Shimmin, Mukkavili,
as an object to be manipulated.
Schwaller, Knoll, Sun,
—Techniques from AD and PPL the ML community could adopt to
Kisamori, Graham, Portwood,
enable research on new models
Huang, Novello, Munchmeyer,
—How to overcome challenges due to the ML’s specific hardware
Jungbluth, Levine, Ayed,
(GPUs, specialized chips) and software (Python) stacks, and the
Atkinson, Hermann,
Afternoon Coffee Break & particular demands of practitioners for their tools
03:20 PM Grönquist, , Saha, Glaser, Li,
Poster Session —Greater collaboration between ML and programming languages
Iiyama, Anirudh, Koch-Janusz,
communities
Sundar, Lanusse, , Köhler,
Yip, guo, Ju, Hanuka, Albert,
Schedule
Salvatelli, Verzetti, Duarte,
Moreno, de Bézenac,
Vlontzos, Singh, Klijnsma, 08:30 AM Opening statements
Neuberg, Wright, Mustafa,
Jan-Willem van de Meent -
Schmidt, Farrell 08:40 AM van de Meent
TBA
04:20 PM Yasaman Bahri Bahri
Applications of a
Equivariant Hamiltonian 09:30 AM disintegration Narayanan
05:00 PM Jimenez Rezende
Flows transformation
Page 43 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abstracts (1):
MineRL Competition on Sample Efficient Reinforcement Learning.
Treguer, Kim, Guo, Luo, Zhao, -- Speakers: Sangyun Shin, David Hyunchul Shim, Ratnesh Madaan,
12:00 PM The AutoDL Challenge Keiko Nagami
Li, Guo, Zhang, Ota
Page 44 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Abhinav Gupta, Michael Noukhovitch, Cinjon Resnick, Natasha 11:30 AM Spotlight presentations x5
Jaques, Angelos Filos, Marie Ossenkopf, Angeliki Lazaridou, Jakob
02:00 PM Invited Talk - 3 Eisner
Foerster, Ryan Lowe, Douwe Kiela, Kyunghyun Cho
02:45 PM Contributed Talk - 3 Brown
West 118 - 120, Sat Dec 14, 08:00 AM
03:00 PM Invited Talk - 4 Andreas
Communication is one of the most impressive human abilities but
Coffee Break / Poster
historically it has been studied in machine learning on confined datasets 03:45 PM
Session
of natural language, and by various other fields in simple
low-dimensional spaces. Recently, with the rise of deep RL methods, the 04:15 PM Invited Talk - 5 Lee
questions around the emergence of communication can now be studied
05:00 PM Panel Discussion
in new, complex multi-agent scenarios. Two previous successful
workshops (2017, 2018) have gathered the community to discuss how, 05:55 PM Closing Remarks
when, and to what end communication emerges, producing research that
was later published at top ML venues such as ICLR, ICML, AAAI. Now,
we wish to extend these ideas and explore a new direction: how
emergent communication can become more like natural language, and Science meets Engineering of Deep Learning
what natural language understanding can learn from emergent
communication. Levent Sagun, CAGLAR Gulcehre, Adriana Romero, Negar
Rostamzadeh, Nando de Freitas
The push towards emergent natural language is a necessary and
important step in all facets of the field. For studying the evolution of West 121 + 122, Sat Dec 14, 08:00 AM
human language, emerging a natural language can uncover the
requirements that spurred crucial aspects of language (e.g. Deep learning can still be a complex mix of art and engineering despite
compositionality). When emerging communication for multi-agent its tremendous success in recent years, and there is still progress to be
scenarios, protocols may be sufficient for machine-machine interactions, made before it has fully evolved into a mature scientific discipline. The
but emerging a natural language is necessary for human-machine interdependence of architecture, data, and optimization gives rise to an
interactions. Finally, it may be possible to have truly general natural enormous landscape of design and performance intricacies that are not
language understanding if agents learn the language through interaction well-understood. The evolution from engineering towards science in
as humans do. To make this progress, it is necessary to close the gap deep learning can be achieved by pushing the disciplinary boundaries.
between artificial and natural language learning. Unlike in the natural and physical sciences -- where experimental
capabilities can hamper progress, i.e. limitations in what quantities can
To tackle this problem, we want to take an interdisciplinary approach by be probed and measured in physical systems, how much and how often
inviting researchers from various fields (machine learning, game theory, -- *in deep learning the vast majority of relevant quantities that we wish
evolutionary biology, linguistics, cognitive science, and programming to measure can be tracked in some way*. As such, a greater limiting
languages) to participate and engaging them to unify the differing factor towards scientific understanding and principled design in deep
perspectives. We believe that the third iteration of this workshop with a learning is how to *insightfully harness the tremendous collective
novel, unexplored goal and strong commitment to diversity will allow this experimental capability of the field*. As a community, some primary aims
burgeoning field to flourish. would be to (i) identify obstacles to better models and algorithms, (ii)
identify the general trends that are potentially important which we wish to
Schedule understand scientifically and potentially theoretically and; (iii) careful
design of scientific experiments whose purpose is to clearly resolve and
pinpoint the origin of mysteries (so-called 'smoking-gun' experiments).
LaCroix, Ossenkopf, Lee,
Fitzgerald, Mihai, Hare, Zaidi, Schedule
Cowen-Rivers, Brown,
08:00 AM Posters Marzoev, Kharitonov, Yuan,
Korbak, Liang, Ren, Dessì, Welcoming remarks and Sagun, Gulcehre, Romero,
08:00 AM
Potash, Guo, Hashimoto, introduction Rostamzadeh, de Freitas
Liang, Zubek, Fu, Zhu
Krzakala, Bahri, Ganguli,
08:15 AM Session 1 - Theory
08:55 AM Intro Remarks Zdeborová, Dieng, Bruna
Page 45 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Lakshmiratan, Yakubova,
Panel - The Role of
04:15 PM Doshi-Velez, Ganguli, Lipton, 09:00 AM Opening
Communication at Large
Paganini, Anandkumar
09:10 AM Invited Speaker 1 Bakshy
Frankle, Schwab, Morcos, Ma,
09:45 AM Break
Contributed Session - Tsai, Salakhutdinov, Jiang,
05:10 PM
Spotlight Talks Krishnan, Mobahi, Bengio, Mao, Nathan, Baldini,
Yaida, Yang Sivakumar, Wang, Magalle
10:30 AM Poster Session 1
Hewa, Shi, Kaufman, Fang,
Zhou, Ding, He, Lubin
Abstracts (1):
Contributed Talk 1: A Weak
Abstract 5: Lunch Break and posters in Science meets Engineering Supervision Approach to
of Deep Learning, Song, Hoffer, Chang, Cohen, Islam, Blumenfeld, 11:00 AM Detecting Visual Anomalies Szeskin
Madsen, Frankle, Goldt, Chatterjee, Panigrahi, Renda, Bartoldson, for Automated Testing of
Birhane, Baratin, Chatterji, Novak, Forde, Jiang, Du, Adilova, Kamp, Graphics Units
Weinstein, Hubara, Ben-Nun, Hoefler, Soudry, Yu, Zhong, Yang, Dhillon,
Contributed Talk 2: Learned
Carbonell, Zhang, Gilboa, Brandstetter, Johansen, Dziugaite, Somani,
11:15 AM TPU Cost Model for XLA Kaufman
Morcos, Kalaitzis, Sedghi, Xiao, Zech, Yang, Kaur, Ma, Tsai,
Tensor Programs
Salakhutdinov, Yaida, Lipton, Roy, Carbin, Krzakala, Zdeborová, Gur-Ari,
Dyer, Krishnan, Mobahi, Bengio, Neyshabur, Netrapalli, Sankaran, Contributed Talk 3: Learned
11:30 AM Nathan
Cornebise, Bengio, Michalski, Ebrahimi Kahou, Arefin, Hron, Lee, Multi-dimensional Indexing
Sohl-Dickstein, Schoenholz, Schwab, Li, Choe, Petzka, Verma, Lin,
Contributed Talk 4: Neural
Sminchisescu 12:00 PM
11:45 AM Hardware Architecture Lin
Search
Since we are a small workshop, we will hold the poster sessions during
the day, including all the breaks as the authors wish. 12:00 PM Lunch
Page 46 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
The third Conversational AI workshop – today's practice and 11:25 AM Contributed talk 4
tomorrow's potential 11:40 AM Invited talk - Alan Ritter
Alborz Geramifard, Jason Williams, Bill Byrne, Asli Celikyilmaz, 12:10 PM Lunch
Milica Gasic, Dilek Hakkani-Tur, Matt Henderson, Luis Lastras, Mari
01:45 PM Invited talk - David Traum Traum
Ostendorf
02:15 PM Invited talk - Y-Lan Boureau Boureau
West 205 - 207, Sat Dec 14, 08:00 AM
02:45 PM Contributed talk 5
In the span of only a few years, conversational systems have become 03:00 PM Contributed talk 6
commonplace. Every day, millions of people use natural-language
interfaces such as Siri, Google Now, Cortana, Alexa and others via 03:15 PM Contributed talk 7
in-home devices, phones, or messaging channels such as Messenger,
03:30 PM Posters + coffee break
Slack, Skype, among others. At the same time, interest among the
research community in conversational systems has blossomed: for Invited Talk - Ryuichiro
04:15 PM Higashinaka
supervised and reinforcement learning, conversational systems often Higashinaka
serve as both a benchmark task and an inspiration for new ML methods
04:45 PM Contributed talk 8
at conferences which don't focus on speech and language per se, such
as NIPS, ICML, IJCAI, and others. Such movement has not been 05:00 PM Panel discussion
unnoticed by major publications. This year in collaboration with AAAI
05:50 PM Closing Geramifard, Williams
community, the AI magazine will have a special issue on conversational
AI (https://tinyurl.com/y6shq2ld). Moreover, research community
challenge tasks are proliferating, including the seventh Dialog Systems
Technology Challenge (DSTC7), the Amazon Alexa prize, and the
Conversational Intelligence Challenge live competitions at NIPS (2017, Document Intelligence
2018).
Nigel Duffy, Rama Akkiraju, Tania Bedrax Weiss, Paul Bennett,
Hamid Reza Motahari-Nezhad
Following the overwhelming participation in our last two NeurIPS
workshops:
West 208 + 209, Sat Dec 14, 08:00 AM
2017: 9 invited talks, 26 submissions, 3 oral papers, 13 accepted papers,
37 reviewers
Business documents are central to the operation of business. Such
2018: 4 invited talks, 42 submission, 6 oral papers, 23 accepted papers,
documents include sales agreements, vendor contracts, mortgage terms,
58 reviewers, we are excited to continue promoting cross-pollination of
loan applications, purchase orders, invoices, financial statements,
ideas between academic research centers and industry. The goal of this
employment agreements and a wide many more. The information in such
workshop is to bring together researchers and practitioners in this area,
business documents is presented in natural language, and can be
to clarify impactful research problems, understand well-founded
organized in a variety of ways from straight text, multi-column formats,
methods, share findings from large-scale real-world deployments, and
and a wide variety of tables. Understanding these documents is made
generate new ideas for future lines of research.
challenging due to inconsistent formats, poor quality scans and OCR,
internal cross references, and complex document structure. Furthermore,
This one day workshop will include invited talks and a panel from
these documents often reflect complex legal agreements and reference,
Page 47 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
explicitly or implicitly, regulations, legislation, case law and standard tasks ranging from playing games such as Go and StarCraft to learning
business practices. dexterity. However, one attribute of intelligence that still eludes modern
The ability to read, understand and interpret business documents, learning systems is generalizability. Until very recently, the majority of
collectively referred to here as “Document Intelligence”, is a critical and reinforcement learning research involved training and testing algorithms
challenging application of artificial intelligence (AI) in business. While a on the same, sometimes deterministic, environment. This has resulted in
variety of research has advanced the fundamentals of document algorithms that learn policies that typically perform poorly when deployed
understanding, the majority have focused on documents found on the in environments that differ, even slightly, from those they were trained
web which fail to capture the complexity of analysis and types of on. Even more importantly, the paradigm of task-specific training results
understanding needed across business documents. Realizing the vision in learning systems that scale poorly to a large number of (even
of document intelligence remains a research challenge that requires a interrelated) tasks.
multi-disciplinary perspective spanning not only natural language
processing and understanding, but also computer vision, knowledge Recently there has been an enduring interest in developing learning
representation and reasoning, information retrieval, and more -- all of systems that can learn transferable skills. This could mean robustness to
which have been profoundly impacted and advanced by neural changing environment dynamics, the ability to quickly adapt to
network-based approaches and deep learning in the last few years. environment and task variations or the ability to learn to perform multiple
We propose to organize a workshop for AI researchers, academics and tasks at once (or any combination thereof). This interest has also
industry practitioners to discuss the opportunities and challenges for resulted in a number of new data sets and challenges (e.g. Obstacle
document intelligence. Tower Environment, Animal-AI, CoinRun) and an urgency to standardize
the metrics and evaluation protocols to better assess the generalization
Schedule abilities of novel algorithms. We expect this area to continue to increase
in popularity and importance, but this can only happen if we manage to
build consensus on which approaches are promising, and, equally
08:30 AM Opening Remarks
important, how to test them.
08:30 AM David Lewis Lewis
The workshop will include a mix of invited speakers, accepted papers
09:30 AM Ndapa Nakashole Nakashole
(oral and poster sessions) and a panel discussion. The workshop
10:30 AM Coffee Break welcomes both theoretical and applied research, in addition to novel data
sets and evaluation protocols.
Discussion Session /
11:00 AM
Posters Schedule
02:00 PM Rajasekar Krishnamurthy Krishnamurthy
Page 48 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 49 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
05:40 PM Closing Remarks Conditional Invertible Flow for Point Cloud Generation. Stypulkowski
Michal, Zamorski Maciej, Zieba Maciej, Chorowski Jan
Abstracts (3):
Getting Topology and Point Cloud Generation to Mesh. Austin Dill,
Chun-Liang Li, Songwei Ge, Eunsu Kang
Abstract 3: Coffee Break & Poster Session 1 in Sets and Partitions,
Zhang, Hare, Prugel-Bennett, Leung, Flaherty, Wiratchotisatian, Epasto,
Distributed Balanced Partitioning and Applications in Large-scale Load
Lattanzi, Vassilvitskii, Zadimoghaddam, Tulabandhula, Fuchs, Kosiorek,
Balancing. Aaron Archer, Kevin Aydin, MohammadHossein Bateni,
Posner, Hang, Goldie, Ravi, Mirhoseini, Xiong, Ren, Liao, Urtasun,
Vahab Mirrokni, Aaron Schild, Ray Yang, Richard Zhuang
Zhang, Borassi, Luo, Trapp, Dubourg-Felonneau, Kussad, Bender,
Zaheer, Oliva, Stypu■kowski, Zieba, Dill, Li, Ge, Kang, Parker Jones,
Abstract 8: Contributed Talk - Limitations of Deep Learning on Point
Wong, Payne, Li, Nazi, Erdem, Erdem, O'Connor, Garcia, Zamorski,
Clouds in Sets and Partitions, Bueno 02:00 PM
Chorowski, Sinha, Clifford, Cassidy 09:45 AM
Limitations of Deep Learning on Point Clouds
Poster Session 1 Paper Titles & Authors:
Christian Bueno, Alan G. Hylton
Deep Set Prediction Networks. Yan Zhang, Jonathon Hare, Adam Abstract 11: Coffee Break & Poster Session 2 in Sets and Partitions,
Prügel-Bennett Lee, Lee, Teh, Yeh, Hu, Schwing, Ahmadian, Epasto, Knittel, Kumar,
Mahdian, Bueno, Sanghi, Jayaraman, Arroyo-Fernández, Hryniowski,
Deep Hyperedges: a Framework for Transductive and Inductive Learning Mathur, Singh, Haddadan, Portilheiro, Zhang, Yuksekgonul, Arias
on Hypergraphs. Joshua Payne Figueroa, Maurya, Ravindran, NIELSEN, Pham, Payan, McCallum,
Mehta, Sun 03:15 PM
FSPool: Learning Set Representations with Featurewise Sort Pooling.
Yan Zhang, Jonathon Hare, Adam Prügel-Bennett Poster Session 2 Paper Titles & Authors:
Deep Learning Features Through Dictionary Learning with Improved Towards deep amortized clustering. Juho Lee, Yoonho Lee, Yee Whye
Clustering for Image Classification. Shengda Luo, Alex Po Leung, Haici Teh
Zhang
Chirality Nets: Exploiting Structure in Human Pose Regression.
Globally Optimal Model-based Clustering via Mixed Integer Nonlinear Raymond Yeh, Yuan-Ting Hu, Alexander Schwing
Programming. Patrick Flaherty, Pitchaya Wiratchotisatian, Andrew C.
Trapp Fair Hierarchical Clustering. Sara Ahmadian, Alessandro Epasto, Marina
Page 50 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Knittel, Ravi Kumar, Mohammad Mahdian, Philip Pham instead embraced RNNs, Temporal CNNs and Transformers, which
incorporate contextual information at varying timescales. While these
Limitations of Deep Learning on Point Clouds. Christian Bueno, Alan G. architectures have lead to state-of-the-art performance on many difficult
Hylton language understanding tasks, it is unclear what representations these
networks learn and how exactly they incorporate context. Interpreting
How Powerful Are Randomly Initialized Pointcloud Set Functions? Aditya these networks, systematically analyzing the advantages and
Sanghi, Pradeep Kumar Jayaraman disadvantages of different elements, such as gating or attention, and
reflecting on the capacity of the networks across various timescales are
On the Possibility of Rewarding Structure Learning Agents: Mutual open and important questions.
Information on Linguistic Random Sets. Ignacio Arroyo-Fernández,
Mauricio Carrasco-Ruiz, José Anibal Arias-Aguilar On the biological side, recent work in neuroscience suggests that areas
in the brain are organized into a temporal hierarchy in which different
Modelling Convolution as a Finite Set of Operations Through areas are not only sensitive to specific semantic information but also to
Transformation Semigroup Theory. Andrew Hryniowski, Alexander Wong the composition of information at different timescales. Computational
neuroscience has moved in the direction of leveraging deep learning to
HCA-DBSCAN: HyperCube Accelerated Density Based Spatial gain insights about the brain. By answering questions on the underlying
Clustering for Applications with Noise. Vinayak Mathur, Jinesh Mehta, mechanisms and representational interpretability of these artificial
Sanjay Singh networks, we can also expand our understanding of temporal
hierarchies, memory, and capacity effects in the brain.
Finding densest subgraph in probabilistically evolving graphs. Sara
Ahmadian, Shahrzad Haddadan In this workshop we aim to bring together researchers from machine
learning, NLP, and neuroscience to explore and discuss how
Representation Learning with Multisets. Vasco Portilheiro computational models should effectively capture the multi-timescale,
context-dependent effects that seem essential for processes such as
PairNets: Novel Fast Shallow Artificial Neural Networks on Partitioned language understanding.
Subspaces. Luna Zhang
We invite you to submit papers related to the following (non-exahustive)
Fair Correlation Clustering. Sara Ahmadian, Alessandro Epasto, Ravi topics:
Kumar, Mohammad Mahdian * Contextual sequence processing in the human brain
* Compositional representations in the human brain
Learning Maximally Predictive Prototypes in Multiple Instance Learning. * Systematic generalization in deep learning
Mert Yuksekgonul, Ozgur Emre Sivrikaya, Mustafa Gokce Baydogan * Compositionality in human intelligence
* Compositionality in natural language
Deep Clustering using MMD Variational Autoencoder and Traditional * Understanding composition and temporal processing in neural network
Clustering Algorithms. Jhosimar Arias models
* New approaches to compositionality and temporal processing in
Hypergraph Partitioning using Tensor Eigenvalue Decomposition. language
Deepak Maurya, Balaraman Ravindran, Shankar Narasimhan * Hierarchical representations of temporal information
* Datasets for contextual sequence processing
Information Geometric Set Embeddings: From Sets to Distributions. Ke * Applications of compositional neural networks to real-world problems
Sun, Frank Nielsen
Submissions should be up to 4 pages excluding references, and should
Document Representations using Fine-Grained Topics. Justin Payan, be NIPS format and anonymous. The review process is double-blind.
Andrew McCallum
We also welcome published papers that are within the scope of the
workshop (without re-formatting). This specific papers do not have to be
Context and Compositionality in Biological and Artificial anonymous. They will only have a very light review process.
Neural Systems
Schedule
Javier Turek, Shailee Jain, Alexander Huth, Leila Wehbe, Emma
Strubell, Alan Yuille, Tal Linzen, Christopher Honey, Kyunghyun 08:00 AM Opening Remarks Huth
Cho
08:15 AM Patricia Churchland Churchland
West 217 - 219, Sat Dec 14, 08:00 AM
09:00 AM Gina Kuperberg Kuperberg
The ability to integrate semantic information across narratives is 09:45 AM Poster Session + Break
fundamental to language understanding in both biological and artificial
cognitive systems. In recent years, enormous strides have been made in 10:30 AM Spotlights - TBA
NLP and Machine Learning to develop architectures and techniques that 11:00 AM Tom Mitchell Mitchell
effectively capture these effects. The field has moved away from
traditional bag-of-words approaches that ignore temporal ordering, and
Page 51 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Note: schedule not final and may change 09:45 AM Coffee Break
Note: schedule not final and may change Contributed Talk (Best
04:00 PM Paper) - Michelle Lee & Florensa, Lee
Abstract 11: Panel in Context and Compositionality in Biological and Carlos Florensa
Artificial Neural Systems, Willke, Fedorenko, Lee, Smolensky, Marcus
04:15 PM Invited Talk - Angela
04:15 PM Schoellig
Schoellig
Note: schedule not final and may change
Page 52 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
04:45 PM Invited Talk - Edward Johns Johns 10:30 AM Yann LeCun LeCun
Abstract 3: Posters 1 in Robot Learning: Control and Interaction in Transform the Set: Memory
the Real World, Genc, Clavera Gilaberte, Zimmer, Smith, Xiao, Fu, Attentive Generation of
11:10 AM Jetchev, Vollgraf
Ding, Stepputtis, Mallya, Bodapati, Lin 10:30 AM Guided and Unguided Image
Collages
All poster presenters are welcome to present at both poster sessions.
Paper Dreams: An
Interactive Interface for
Abstract 4: Contributed Talk - Laura Smith in Robot Learning: 11:20 AM Bernal, Zhou
Generative Visual
Control and Interaction in the Real World, Smith 11:15 AM
Expression
AVID: Translating Human Demonstrations for Automated Learning Deep reinforcement learning
11:30 AM Rojas
for 2D soft body locomotion
Abstract 11: Contributed Talk (Best Paper) - Michelle Lee & Carlos
Florensa in Robot Learning: Control and Interaction in the Real Towards Sustainable
World, Florensa, Lee 04:00 PM Architecture: 3D
Convolutional Neural
Combining Model-Free and Model-Based Strategies for Sample-Efficient 11:40 AM Networks for Computational Musil
Reinforcement Learning Fluid Dynamics Simulation
and Reverse Design
Workflow
NeurIPS Workshop on Machine Learning for Creativity and
Human and GAN
Design 3.0
11:50 AM collaboration to create Seita, Koga
haute couture dress
Luba Elliott, Sander Dieleman, Adam Roberts, Jesse Engel, Tom
White, Rebecca Fiebrink, Parag Mital, Christine Payne, Nao Tokui Lee, Saeed, Broad, Gillick,
Hertzmann, Aggarwal, Sung,
West 223 + 224, Sat Dec 14, 08:00 AM 01:30 PM Poster Session 1 Champandard, Park, Mellor,
Herrmann, Wu, Lee, Jieun,
Generative machine learning and machine creativity have continued to
Han, jung, Kim
grow and attract a wider audience to machine learning. Generative
models enable new types of media creation across images, music, and 02:30 PM Sougwen Chung Chung
text - including recent advances such as StyleGAN, MuseNet and
03:45 PM Claire Evans Evans, Bechtolt, Kieswetter
GPT-2. This one-day workshop broadly explores issues in the
applications of machine learning to creativity and design. We will look at MidiMe: Personalizing a
algorithms for generation and creation of new media, engaging 04:15 PM MusicVAE model with user Dinculescu
researchers building the next generation of generative models (GANs, data
RL, etc). We investigate the social and cultural impact of these new
First Steps Towards
models, engaging researchers from HCI/UX communities and those
04:25 PM Collaborative Poetry Uthus, Voitovich
using machine learning to develop new creative tools. In addition to
Generation
covering the technical advances, we also address the ethical concerns
ranging from the use of biased datasets to the use of synthetic media 04:35 PM Panel Discussion
such as “DeepFakes”. Finally, we’ll hear from some of the artists and
Saxena, Frosst, Cabannes,
musicians who are adopting machine learning including deep learning
Kogan, Dill, Sarkar, Moniz,
and reinforcement learning as part of their own artistic process. We aim
Thio, Sievert, Coleman, De
to balance the technical issues and challenges of applying the latest 05:00 PM Poster Session 2
Bleser, Quanz, Kereliuk,
generative models to creativity and design with philosophical and cultural
Achlioptas, Elhoseiny, Ge,
issues that surround this area of research.
Gomez, Brew
Schedule Sarin, Bourached, Carr,
Zukowski, Zhou, Malakhova,
05:05 PM Artwork
08:15 AM Welcome and Introduction Petric, Laurenzo, O'Brien,
Wegner, Kishi, Burnam
08:30 AM Alec Radford Radford
Page 53 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
West 301 - 305, Sat Dec 14, 08:00 AM Liu, Sharan, Abolmaesumi,
Parmar, Lei, Gavves, Nabi,
Medical imaging and radiology are facing a major crisis with an Namdar, Chen, Modi, Fels,
ever-increasing complexity and volume of data along an immense Rauscher, Li, Chung, Oktay,
economic pressure. The current advances and widespread use of Coffee Break + Poster
03:00 PM Gopinath, Selvan, Adiga
imaging technologies now overload the human capacity of interpreting Session
Vasudeva, Poblenz, Baltatzis,
medical images, dangerously posing a risk of missing critical patterns of Wei, Velayutham, Garyfallidis,
diseases. Machine learning has emerged as a key technology for Ellis, Bhatia, Galitz, Muckley,
developing novel tools in computer aided diagnosis, therapy and Cai, Prasanna
intervention. Still, progress is slow compared to other fields of visual
recognition, which is mainly due to the domain complexity and Session 4 (Invited Talk +
03:30 PM
constraints in clinical applications, i.e., robustness, high accuracy and presentations)
reliability.
Yakubova, Pezzotti, Wang,
05:00 PM fastMRI Challenge Talks Zitnick, Karkalousos, Sun,
“Medical Imaging meets NeurIPS” aims to bring researchers together
Caan, Murrell
from the medical imaging and machine learning communities to discuss
the major challenges in the field and opportunities for research and novel 06:00 PM Closing Remarks
applications. The proposed event will be the continuation of a successful
workshop organized in NeurIPS 2017 and 2018
(https://sites.google.com/view/med-nips-2018). It will feature a series of Abstracts (1):
invited speakers from academia, medical sciences and industry to give
Abstract 9: fastMRI Challenge Talks in Medical Imaging meets
an overview of recent technological advances and remaining major
NeurIPS, Yakubova, Pezzotti, Wang, Zitnick, Karkalousos, Sun, Caan,
challenges.
Murrell 05:00 PM
Schedule
tentative
Lombaert, Glocker,
08:15 AM Opening Remarks Konukoglu, de Bruijne, Learning with Temporal Point Processes
Feragen, Oguz, Teuwen
Manuel Rodriguez, Le Song, Isabel Valera, Yan Liu, Abir De,
Session 1 (Invited Talk + Schnabel, Vidal, Sodickson,
08:30 AM Hongyuan Zha
presentations) Grady, Vidal
Page 54 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Schedule
Cakmak, Zhang,
Prabhakarannair Kusumam,
Welcome Address and Ahmed, Wu, Choudhari,
08:30 AM
Introduction 05:00 PM Poster Session Inouye, Taylor, Besserve,
Turkmen, Islam, Artés, Setlur,
Invited Talk by Negar
08:35 AM Kiyavash Fu, Han, De, Du, Sanchez
Kiyavash
Martin
Fused Gromov-Wasserstein
09:15 AM Alignment for Hawkes
Processes
The Optimization Foundations of Reinforcement Learning
Insider Threat Detection via
09:30 AM Hierarchical Neural Wu
Bo Dai, Niao He, Nicolas Le Roux, Lihong Li, Dale Schuurmans,
Temporal Point Processes
Martha White
09:45 AM Coffee Break
West Ballroom A, Sat Dec 14, 08:00 AM
Invited Talk By Niloy
10:30 AM ganguly
Ganguly Interest in reinforcement learning (RL) has boomed with recent
improvements in benchmark tasks that suggest the potential for a
Intermittent Demand
revolutionary advance in practical applications. Unfortunately, research
11:10 AM Forecasting with Deep Turkmen
in RL remains hampered by limited theoretical understanding, making
RenewalProcesses
the field overly reliant on empirical exploration with insufficient principles
Temporal Logic Point to guide future development. It is imperative to develop a stronger
11:25 AM
Processes fundamental understanding of the success of recent RL methods, both to
expand the useability of the methods and accelerate future deployment.
The Graph Hawkes Network
Recently, fundamental concepts from optimization and control theory
11:40 AM for Reasoning on Temporal
have provided a fresh perspective that has led to the development of
Knowledge Graphs
sound RL algorithms with provable efficiency. The goal of this workshop
Multivariate coupling is to catalyze the growing synergy between RL and optimization
estimation between research, promoting a rational reconsideration of the foundational
11:55 AM Besserve
continuous signals and principles for reinforcement learning, and bridging the gap between
point processes theory and practice.
Page 55 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 56 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Then, we consider sample-based TRPO and establish $\tilde is the fixed and known lead time, and D is an unknown parameter of the
O(1/\sqrt{N})$ convergence rate to the global optimum. Importantly, the demand distribution described roughly as the number of time steps
adaptive scaling mechanism allows us to analyze TRPO in regularized needed to generate enough demand for depleting one unit of inventory.
MDPs for which we prove fast rates of $\tilde O(1/N)$, much like results Our results significantly improve the existing regret bounds for this
in convex optimization. This is the first result in RL of better rates when problem. Notably, even though the state space of the underlying Markov
regularizing the instantaneous cost or reward. Decision Process (MDP) in this problem is continuous and
L-dimensional, our regret bounds depend linearly on L. Our techniques
Abstract 6: The Provable Effectiveness of Policy Gradient Methods utilize convexity of the long run average cost and a newly derived bound
in Reinforcement Learning in The Optimization Foundations of on the `bias' of base-stock policies, to establish an almost blackbox
Reinforcement Learning, Kakade 10:30 AM connection between the problem of learning and optimization in such
MDPs and stochastic convex bandit optimization. The techniques
Reinforcement learning is now the dominant paradigm for how an agent presented here may be of independent interest for other settings that
learns to interact with the world in order to achieve some long term involve learning large structured MDPs but with convex cost functions.
objectives. Here, policy gradient methods are among the most effective
methods in challenging reinforcement learning problems, due to that Abstract 12: Plenary Talk 5 in The Optimization Foundations of
they: are applicable to any differentiable policy parameterization; admit Reinforcement Learning, Yu 04:20 PM
easy extensions to function approximation; easily incorporate structured
state and action spaces; are easy to implement in a simulation based, TBA
model-free manner.
Abstract 13: Continuous Online Learning and New Insights to Online
However, little is known about even their most basic theoretical Imitation Learning in The Optimization Foundations of
convergence properties, including: Reinforcement Learning, Lee, Cheng, Goldberg, Boots 05:00 PM
- do they converge to a globally optimal solution, say with a sufficiently
rich policy class? Online learning is a powerful tool for analyzing iterative algorithms.
- how well do they cope with approximation error, say due to using a However, the classic adversarial setup sometimes fails to capture certain
class of neural policies? regularity in online problems in practice. Motivated by this, we establish a
- what is their finite sample complexity? new setup, called Continuous Online Learning (COL), where the gradient
This talk will survey a number of results on these basic questions. We of online loss function changes continuously across rounds with respect
will highlight the interplay of theory, algorithm design, and practice. to the learner’s decisions. We show that COL covers and more
appropriately describes many interesting applications, from general
Joint work with: Alekh Agarwal, Jason Lee, Gaurav Mahajan equilibrium problems (EPs) to optimization in episodic MDPs. Using this
new setup, we revisit the difficulty of achieving sublinear dynamic regret.
Abstract 7: Logarithmic Regret for Online Control in The We prove that there is a fundamental equivalence between achieving
Optimization Foundations of Reinforcement Learning, Agarwal, sublinear dynamic regret in COL and solving certain EPs, and we
Hazan, Singh 11:10 AM present a reduction from dynamic regret to both static regret and
convergence rate of the associated EP. At the end, we specialize these
We study optimal regret bounds for control in linear dynamical systems new insights into online imitation learning and show improved
under adversarially changing strongly convex cost functions, given the understanding of its learning stability.
knowledge of transition dynamics. This includes several well studied and
fundamental frameworks such as the Kalman filter and the linear Abstract 14: Panel Discussion in The Optimization Foundations of
quadratic regulator. State of the art methods achieve regret which scales Reinforcement Learning, Sutton, Precup 05:20 PM
as T^0.5, where T is the time horizon.
TBA
We show that the optimal regret in this setting can be significantly
Abstract 15: Closing Remarks in The Optimization Foundations of
smaller, scaling as polylog T. This regret bound is achieved by two
Reinforcement Learning, Dai, He, Le Roux, Li, Schuurmans, White
different efficient iterative methods, online gradient descent and online
05:45 PM
natural gradient.
Awards Announcement
Abstract 9: Plenary Talk 3 in The Optimization Foundations of
Reinforcement Learning, Van Roy 02:00 PM
Abstract 10: Learning in structured MDPs with convex cost function: Ben London, Gintare Karolina Dziugaite, Dan Roy, Thorsten
improved regret bounds for inventory management in The Joachims, Aleksander Madry, John Shawe-Taylor
Optimization Foundations of Reinforcement Learning, Agrawal 02:40
PM West Ballroom B, Sat Dec 14, 08:00 AM
We present a learning algorithm for the stochastic inventory control As adoption of machine learning grows in high-stakes application areas
problem under lost sales penalty and positive lead times, when the (e.g., industry, government and health care), so does the need for
demand distribution is a priori unknown. Our main result is a regret guarantees: how accurate a learned model will be; whether its
bound of O(L\sqrt{T}+D) for the algorithm, where T is the time horizon, L predictions will be fair; whether it will divulge information about
Page 57 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
08:45 AM Welcome Address London Abstract 10: Coffee Break / Poster Session 2 in Machine Learning
09:00 AM TBD Roth with Guarantees, 03:15 PM
11:00 AM Mehryar Mohri Mohri The purpose of this workshop is to bring together experts from different
Page 58 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Schedule
Abstracts (5):
Joachims, Kallus,
Abstract 4: Poster Spotlights in “Do the right thing”: machine
08:45 AM Opening Remarks Santacatterina, Swaminathan,
learning and causal inference for improved decision making,
Sontag, Zhou
Namkoong, Charpignon, Rudolph, Coston, Saito, Dhillon, Markham
09:00 AM Susan Athey Athey 10:00 AM
09:30 AM Andrea Rotnitzky Rotnitzky Poster spotlights ID: 10, 11, 16, 17, 20, 24, 31
Namkoong, Charpignon,
Abstract 8: Tentative topic: Reasoning about untestable
10:00 AM Poster Spotlights Rudolph, Coston, Saito,
assumptions in the face of unknowable counterfactuals in “Do the
Dhillon, Markham
right thing”: machine learning and causal inference for improved
Lu, Chen, Namkoong, decision making, 12:00 PM
Charpignon, Rudolph, Coston,
von Kügelgen, Prasad, Tentative topic: How machine learning, and causal inference work
Dhillon, Xu, Wang, Markham, together: cross-pollination and new challenges.
Rohde, Singh, Zhang,
Abstract 11: Contributed talk 1 in “Do the right thing”: machine
Coffee break, posters, and Hassanpour, Sharma, Lee,
10:15 AM learning and causal inference for improved decision making, Chen,
1-on-1 discussions Pouget-Abadie, Krijthe,
Boehnke, Wang, Bonaldi 03:00 PM
Mahajan, Ke, Wirnsberger,
Semenova, Mykhaylov, Shen,
Oral Spotlights ID: 8,9, 27
Takatsu, Sun, Yang, Franks,
Wong, Zaman, Mitchell, kang, Abstract 12: Contributed talk 2 in “Do the right thing”: machine
Yang learning and causal inference for improved decision making,
11:00 AM Susan Murphy Murphy Mahajan, Khosravi, D'Amour 03:15 PM
11:30 AM Ying-Qi Zhao Zhao Oral Spotlights ID: 57, 93, 113
03:15 PM Contributed talk 2 Mahajan, Khosravi, D'Amour Ioannis Mitliagkas, Gauthier Gidel, Niao He, Reyhane Askari
Hemmat, Nika Haghtalab, Simon Lacoste-Julien
Griveau-Billion, Singh, Zhang,
Lee, Krijthe, Charles,
03:30 PM Poster Spotlights West Exhibition Hall A, Sat Dec 14, 08:00 AM
Semenova, Ladhania,
Oprescu Advances in generative modeling and adversarial learning gave rise to a
recent surge of interest in differentiable two-players games, with much of
Page 59 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 60 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
games. Using this insight, we develop two-timescale learning dynamics field of deep reinforcement learning has led to remarkable empirical
that converge to Stackelberg equilibria in zero-sum games and the set of results in rich and varied domains like robotics, strategy games, and
stable attractors in general-sum games. multiagent interaction. This workshop will bring together researchers
working at the intersection of deep learning and reinforcement learning,
Abstract 10: Contributed talk: On Solving Local Minimax and it will help interested researchers outside of the field gain a
Optimization: A Follow-the-Ridge Approach in Bridging Game high-level view about the current state of the art and potential directions
Theory and Deep Learning, Wang 02:30 PM for future contributions.
In recent years, the use of deep neural networks as function NeurIPS RL Competitions
04:00 PM
approximators has enabled researchers to extend reinforcement learning Results Presentations
techniques to solve increasingly complex control tasks. The emerging
Page 61 of 62
NeurIPS 2019 Workshop book Generated Thu Nov 28, 2019
Page 62 of 62