Sei sulla pagina 1di 32

Citation:

Azencott, C.-A. (2018). Machine learning and genomics: precision medicine versus patient
privacy. ​Philosophical Transactions of the Royal Society A,​ ​376​(2128).
https://doi.org/10.1098/rsta.2017.0350

Summary:
The growth in the availability of big data has allowed data processing techniques, including
machine learning, to prevail in the field of medicine. Precision medicine is a specific application
of this. Using specific clinical, environmental, and genetic information concerning the patient,
risk prediction, prognosis, and treatment plans can be tailored to the individual. The first
complication in doing this arises at the data collection level: there are often more variables
collected for each patient than there are patients. Thus, data sharing throughout this field is
crucial to ensure that no statistical difficulties arise from this issue. Data sharing, however, is not
without its own concerns. A patient’s genomic sequence contains some of the most private
information about them. As more research is done in this field, the wealth of information that can
be extracted from a genome is continuously expanded. Genetic discrimination has become a
concern; acts in the US make it illegal for employers or health insurers to discriminate against
individuals with certain genetic mutations. While patient identifiers are removed from data
before it is shared, anonymization does not guarantee unidentifiability if only one patient at a
given hospital matches the disease data. Public genetic genealogy databases have made it
possible for patient’s surnames to be discovered from limited amounts of their genome. In order
to ensure the privacy of databases necessary for machine learning, these issues of data security
and cryptography must be studied.

Annotation:
Author Chloé-Agathe Azencott is a researcher at the Centre for Computational Biology of Mines
ParisTech, and is additionally affiliated with the Institut Curie of the PSL Research University
and Inserm. Written in 2018, this article is still current on specific data privacy issues related to
sharing of genomic data. The information in this article can be corroborated. The authors of the
article, “Unintended consequences of machine learning in medicine,” agree that, “…the
introduction of new technologies in health care has not always been straightforward or without
unintended and adverse effects,” and that research should be done to combat this deficiency
(Cabitza et al., 2017). The coverage of this article is both broad and deep. The authors begin with
a broad overview of what types of data are used in medical fields, and why genomic data, even
when reported anonymously, poses a risk for privacy. They then explore each fallacy deeper,
describing their cryptographic or legislative solutions. The authors explicitly state their purpose
in a call to action, writing, “regulators and scientists must work together to develop appropriate
sharing frameworks, based on ethical concerns and technical solutions.” The purpose is to
encourage work by data scientists, policymakers, and other people in a position to solve ethical
concerns, to further research solutions for data sharing privacy issues.

Citation:
Cabitza, F., Rasoini, R., & Gensini, G. F. (2017). Unintended consequences of machine learning
in medicine. ​Journal of the American Medical Association​, ​318​(6), 517-518.
https://doi.org/10.1001/jama.2017.7797

Summary:
Development of machine learning techniques have driven changes in how health care problems
are approached. While these have the possibility to improve the accuracy of prognostic models
and other pattern recognition applications, as with any novel technology it is necessary to
consider potential adverse consequences. First, application of machine learning could lead to an
overreliance on technology, which may lead to deskilling of physicians and thus there would be
no backup for when the automated program fails. Also, by their nature, machine learning
algorithms focus on the data presented, which cannot always be trusted to provide a complete
representation of the problem. If certain aspects are not included in the data set, their existence
may easily be forgotten, and results may be misinterpreted. Uncertainty is also intrinsically
weaved into medical decisions. Clinicians often disagree about diagnoses, and observer
variability may limit the accuracy of a machine learning algorithm which uses this data. Finally,
factors other than accuracy are important to consider when determining what a “good” algorithm
is. Relying solely on accuracy makes the algorithm “opaque”, where the developer does not
consider the logic which is actually occurring in the machine. This makes the developer
susceptible to missing any logical shortcomings with the algorithm.

Annotation:
Author Federico Cabitza is associated with the Department of Informatics at the University of
Milano-Bicocca and the IRCCS Istituto Ortopedico (orthopedic institute), while Raffaele Rasoini
and Gian franco gensini are associated with the Centro Studi Medicina Avanzata (advanced
medical study center). Written in 2017, this article is still current on the potential consequences
of using machine learning in medical fields. The information in this article can be corroborated.
In “Machine learning and genomics: precision medicine versus patient privacy,” the authors
agree that while data “…has opened the door to new approaches in biomedical research,
allowing us to use statistical and machine learning approaches to objectively analyze and
generate new hypotheses,” it is important to consider other non-technical issues before using
them. The coverage of this article is both broad and deep (Azencott, 2018). The authors present
many different possible consequences of machine learning in medicine, ranging from
over-automation of the medical field to intrinsic opacity of algorithms. Each of these subjects are
illustrated with an example: when describing a failure to consider cofactors of pneumonia
mortality they describe an algorithm which seemed to suggest (due to a wrong assumption that
all pneumonia patients are treated equally) that patients with both pneumonia and asthma had
lower mortality rates than patients with only pneumonia. The purpose of this article is twofold:
first, it aims to provoke more detailed research on unintended consequences of machine learning.
Second, it encourages a careful approach when designing algorithms. The audience suits both of
these purposes. Machine learning developers are situated where knowledge of the consequences
of their work would be useful, and they have enough experience to be able to research them
further.

Citation:
Camporeale, E. (2018). The challenge of machine learning in space weather: Nowcasting and
forecasting. ​Space Weather,​ ​18,​ 1166-1207. ​https://doi.org/10.1029/2018SW002061​.

Summary:
The author begins her paper with a discussion of the current state of machine learning: broad
classifications of algorithms and fields where such algorithms have been successfully applied. he
cites the growth of big data, technological advancements such as GPUs, and investment in
computer science as the reasons why machine learning is seeing a spike in popularity in recent
years. For the space weather question, the author provides a literature review proving machine
learning applications in forecasting aspects of geomagnetic activity, among other related areas.
Camporeale then discusses some of the problems with machine learning and space weather.
These include limited information with which to make a prediction, the gray-box problem, the
surrogate problem (replacing aspects with a black box), and associated uncertainties. Similar
problems exist with all applications of machine learning.

Annotation:
Camporeale is affiliated with CIRES, a partnership between NOAA and the University of
Colorado Boulder focused on environmental science, and is also with the Centrum Wiskunde &
Informatica, a national research institute in Amsterdam. Written in 2018, this article is still
current on the effects of space weather on machine learning’s applications in space. The
information in this article can be corroborated. The article, “Automated solar activity prediction:
A hybrid computer platform using machine learning and solar imaging for automated prediction
of solar flares,” agrees that machine learning can be applied to predict specific events, including
space weather ones like solar flares (Colak & Qahwaji 2009). The coverage of this article is both
broad and deep. It discusses a broad range of topics related to machine learning for space
weather, including the type of problems it can be applied to and the problems that exist. For each
application, the author dives deep into an analysis of past successes and failures of both machine
learning algorithms and non-machine learning imaging techniques. This article is intended for
those who work in predicting space weather who may be interested in applying machine learning
to their field. It describes the potential for its application without delving into complicated
mathematics. It also explains the current challenges which future work should address. The
purpose is to provide a gentle introduction to machine learning for those unfamiliar with the
subject and encourage them to further explore its applications to space weather.

Citation:
DeTone, D., Malisiewicz, T., & Rabinovich, A. (2016). Deep image homography estimation.
Computing Research Repository,​ ​https://arxiv.org/abs/1606.03798

Summary:
Currently, image homography is estimated with corner detection -- mapping the corners found on
the warped image to their respective locations on the original. Homography is needed for
panoramas and camera calibration, among other applications. The goal of using machine learning
in image homography is so that instead of manually selecting corner or line features, the
computer can find points of interest itself. To generate training data, the authors use cropped
MS-COCO images, converted to grayscale. They randomly crop a square patch from the larger
image, Patch A, and then randomly perturb each of Patch A's corners to create a Patch B. The
two patches are stacked channel-wise to create a 2-channel image (128x128x2), the input to a
convolutional neural network. Blurs or occlusions could be randomly inserted to make the model
more robust. The authors found that homography estimation could be solved as a machine
learning problem. A benefit to using machine learning is that if it is intended to be used for a
specific task, such as a robot navigating a factory floor, it can be trained on problem-specific
images. If it is just trained on images of a factory floor it would be optimized for that particular
situation.

Annotation:
All authors are affiliated with Magic Leap, a company that works on computer vision and
augmented reality technologies. Written in 2016, this article is current on the uses of machine
learning to solve image homography problems. The information in this article can be
corroborated. The authors of the article “Homography Estimation from Image Pairs with
Hierarchical Convolutional Networks” agree that with regression, convolutional neural networks
are successful at estimating the homography between a pair of images (Nowruzi et al. 2017). The
coverage of this article is both broad and deep. It begins by describing broadly what image
homography is, how it can be applied to real-life problems, and the current, non-machine
learning methods for estimating it. It then explains in depth how one would go about creating a
homography dataset from a directory of images, because none exist publicly. It continues a deep
explanation of the parameterization they use to estimate the rotational and translational changes
between an original and warped image. The authors' purpose is to present machine learning as a
viable option for solving homography problems. They mention that this is a little-explored field,
as large, publicly available test datasets for homography needed for machine learning do not
exist. Thus, the authors' audience is people who model image homography, who may be looking
for an alternative to common methods.

Citation:
Diamantas, S., Oikonomidis, A., & Crowder, R. M. (2011). Depth computation using optical
flow and least squares. ​IEEE/SICE International Symposium on System Integration (SII).​
https://10.1109/SII.2010.5708293

Summary:
The computer vision this paper seeks to address is depth computation -- finding the distance
between a robot and a landmark -- in such a way that sharp corner features are not needed. One
of the methods the authors test is optical flow, a model inspired by how insects, particularly
honeybees, navigate by judging the rate of change of features in the image. The particular optical
flow algorithm employed in this paper is the Lucas-Kande (LK) algorithm. The authors note that
this algorithm requires the images to fulfill three key criteria: constant brightness, small
movements between frames, and all pixels moving in the same way. The model is tested on a
simulation of various distances with a robot moving at various velocities. Combining optical
flow and least squares offers a more accurate distance value, as the two models separately
produce independent estimations. The combination was found to perform better than either of the
two models did on their own. However, the authors still found a tradeoff between accuracy and
computational cost, and that accuracy improved the closer the robot was to the landmark.

Annotation:
All authors are affiliated with the University of Southampton, either through the Intelligence,
Agents, Multimedia Group at their School of Electronics and Computer Science or through the
Centre for Risk Research at their School of Management. Despite being written in 2011, this
article is still current because the method with which it is concerned, optical flow, is not a recent
technological advancement. The basic theory behind how optical flow works has not changed
since this article was published. The information in this article can be corroborated. Marcus Pan
agrees with the findings of this article in his work, “Depth Estimation of Light Field Imaged by
Optical Flow”, as he too argues that optical flow is a successful model from which to estimate
depth from an image (2017).​ ​The coverage of this article is both broad and deep. In its
exploration of a model to find depth estimation, it examines a broad range of algorithms: optical
flow, least squares, and the resulting combination of the two. For each model, the authors write
in depth of the mathematics behind the algorithm and of the rationale behind their development
as well. For the target audience of model developers, this article’s purpose is to present the
combination of least squares and optical flow as a valuable tool with which to estimate depth.
They note that such information would be useful for those working in the fields of autonomous
navigation and localization for applications such as search and rescue or planet exploration.

Citation:
Diao, J. A., Kohane, I. S., & Manrai, A. K. (2018). Biomedical informatics and machine learning
for clinical genomics. ​Human Molecular Genetics,​ ​27,​ R29-R34.
https://doi.org/10.1093/hmg/ddy088

Summary:
Genes play an integral role in the occurrence of many diseases, and genome sequencing is a way
in which this role can be extrapolated for clinical benefit. The cost of gene sequencing has
recently dropped significantly, and similarly, genetic tests exist for over 11,000
inheritance-linked conditions. Thus, over time, the ability to and request for genetic counseling
on disease risk has increased. Machine learning has an application in this field, in filtering and
linking diverse data about diseases and patient allele frequency. While large public databases
exist, such as NIH’s ClinVar, data storage, analysis, and security remain issues. Two key ideas
when discussing clinical genomics are pathogenicity, the likelihood that a certain genetic
combination will result in a disease, and penetrance, the probability of disease among patients
with those genes. Both of these measures may vary greatly between laboratories which calculate
them. Machine learning in clinical genomics aims to improve the quality of gene-linked disease
prediction. Neural networks have been notably successful in handling genomic data, in
particular, predicting tissue-specific sequence function. The nodular workflow of these networks
is reminiscent of the differential diagnosis process of a clinician. Linear models and support
vector machines have also been used. Current limitations of machine learning center around
unlabeled or mislabeled data, often by clinicians or genetic counselors who are uncertain of their
classification or disagree with others on it.

Annotation:
Authors James A. Diao, Isaac S. Kohane, and Arjun K. Manrai are all associated with the
Department of Biomedical Informatics at Harvard Medical School. Diao is also affiliated with
the Department of Statistics and Data Science at Yale University. Written in 2018, this article
provides information on the current state and prospective future of machine learning as it is used
in clinical genomics. The information in this article can be corroborated. In “Machine learning in
genetics and genomics,” the authors agree that with machine learning, “gene expression data can
be used to learn to distinguish between different disease phenotypes and, in the process, to
identify potentially valuable disease biomarkers,” (Libbrecht & Noble, 2015). This article is both
broad and deep. It begins by covering a broad range of general topics related to biomedical
informatics and clinical genomics, including a discussion of the state of the latter topic and how
the former is now being applied to it. Then, the authors delve deep into a discussion of machine
learning, and consider multiple types of machine learning algorithms and the specific progress of
each one. The audience for which this article is intended is data scientists with some background
in machine learning. The article assumes the reader has little knowledge of clinical genomics and
explains the current issue of genetic diseases and their prediction in detail. However, when
discussing how machine learning fits as a solution, the reader is assumed to have some
knowledge of the algorithm names used. As machine learning is a growing field, the purpose of
this article is to inform readers of its applications in clinical genomics and encourage them to
study it further. The authors, while acknowledging current limitations of the technology, take the
stance that machine learning can be highly beneficial as a disease predictor.

Citation:
Dong, H., Hu, Q., & Akella, M. R. (2017). Safety control for spacecraft autonomous rendezvous
and docking under motion constraints. ​Journal of Guidance, Control, and Dynamics,​
40​(7), 1680-1692. ​https://doi.org/10.2514/1.G002322

Summary:
Since its first successful test in 1966, spacecraft rendezvous and docking mechanisms have been
integral in refueling and on-orbit assembly. The main phases of spacecraft rendezvous are the
homing phase, closing phase, final translation, and proximity operations. Throughout this whole
process, flight safety is the first priority. Threatening this are noncooperative, exterior obstacles
such as space debris, which make real-time implementation of trajectory planning methods
difficult. This article presents an artificial potential function (APF) method, to encompass
constraints such as keep-out zones and obstacles in the docking execution during the final phase
in particular. Similar methods have been previously used in robotic and spacecraft motion
planning and collision avoidance. Its benefits are real-time response ability to avoid collisions,
even when confronted with a noncooperative obstacle or a nontrivial mechanism fault.

Annotation:
Authors Dong, Hu, and Akella, are affiliated with the Departments of Engineering at Harbin
Institute of Technology, Beihang University, and the University of Texas at Austin, respectively.
Written in 2017, this article is still current on applicable methods to prevent collision with
obstacles in the final phase of rendezvous and docking. The information in this article can be
corroborated. In “Rendezvous and docking for space exploration,” the authors agree that in
automating rendezvous and docking procedures, a system must be created such that may function
under a variety of unforeseen circumstances, including physical obstructions (Manchula &
Sandhoo, 2005). The coverage of this article is both broad and deep. It begins with a broad
outline of the problem: that automated rendezvous maneuvers may be thrown off their planned
trajectory if they encounter an object. The authors’ solution is an artificial potential function,
which they discuss broadly in its prior uses. They continue to dive deeper into this solution, and
support the technical aspect using diagrams of vectors, equations, and detailed proofs of the
theories they use. The purpose of this article is to outline a function that should be used for
real-time motion planning of the spacecraft during rendezvous. Its audience is the aerospace
engineers who would be implementing it, apparent by the fact that the majority of the article is
devoted to technical outlines of the function itself.

Citation:
Erhan, D., Szegedy, C., Toshev, A., & Anguelov, D. (2014). Scalable object detection using deep
neural networks. ​The IEEE Conference on Computer Vision and Pattern Recognition
(CVPR)​, 2147-2154. ​https://ieeexplore.ieee.org/document/6909673/citations

Summary:
Object detection, a key component in computer vision, is traditionally done in an exhaustive
search of all possible locations in an image. A separate detector is trained per object being
searched for. This paper presents a single deep neural network to detect objects in a “class
agnostic manner”. The model scores the likelihood of a box containing an object, as opposed to
the slow process of scoring features within predefined boxes. Current models scale linearly,
evaluating more locations or more classes relies more parts in the model. Past class agnostic
approaches have created classless object boundaries. For models like these, classless
segmentation is done in one layer, and then in another, the class of each segment is determined.
In the author’s form of classless segmentation, the model outputs a number of bounding boxes
with confidence scores of the likelihood of containing an object. A useful characteristic of a class
agnostic segmentation model is that it can also segment an image with a single class, depending
on the class or classes on which it was trained. Their approach, unlike “one-box-per-class”
approaches, is also able to generalize if there are multiple instances of the same object.

Annotation:
All authors work with machine learning for Google: Erhan works with video generation and
prediction for Google Brain, Szegedy and Toshev as research scientists of machine learning, and
Anguelov in Google’s Engineering Department. Written in 2014, this article is still current on the
ways to effectively scale object-detecting neural networks. The information in this article can be
corroborated. Authors of the article, “Application of Deep Learning for Object Detection,” agree
in their explanation of object detection, writing that it “[determines] the instance of the class to
which the object belongs…by outputting the bounding box…” (Pathak et al. 2018).​ T ​ he
coverage of this article is both broad and deep. The authors begin with a discussion of the
different approaches researchers have previously taken to use classless segmentation to make
neural networks more scalable, as well as the shortcomings of models which do not. They
explain deeper the formation of their network and the mathematical formulas behind it, including
the vectors and sums to create their layers, confidence arrays, and confidence scores.​ T ​ he
audience of this article is researchers who already have experience in writing neural networks,
though not class agnostic ones. The article focuses on the differences between their model and
standard neural networks, assuming the reader could bring background on the standard ones. For
example, their statement, “We encode the ​i-​ th object box and its associated confidence as node
values of the last layer,” requires the reader to have prior knowledge that nodes in the last layer
of the network determine the output of the model. Thus, the purpose of the article is to present
classless segmentation as a useful alteration to make in standard neural networks.

Citation:
Finlayson, G., Gong, H. & Fisher, R. B. (2019). Color homography: Theory and applications.
IEEE Transactions on Pattern Analysis and Machine Intelligence,​ ​41(​ 1), 20-33.
https://ieeexplore.ieee.org/abstract/document/8141909​.

Summary:
In computer vision there are two important parts of an image: how edges, points, and surfaces
combine to form the shapes of an image, and how light and shadows create different shades of
color. This paper focuses on how a homography can be estimated on color image pairs. It argues
that a model must map the colors of an image independent of shading, as brightness only affects
the magnitude of an RGB vector. Through testing, they determine that color change can also be
represented as a part of the homography matrix. Calculating a color homography and applying its
inverse has applications throughout the field of computer vision, from color correction problem
and camera calibration, to color-based object recognition.

Annotation:
Finlayson and Gong are associated with the School of Computing Sciences at the University of
East Anglia, where Finlayson is a professor of computing sciences, and Fisher is a professor at
the University of Edinburgh’s School of Informatics. Written in 2019, this article is still current
on how homographies are computed on problems with color images. The idea that color can be
encoded in a homography matrix is relatively unexplored, as this paper was one of the first to
suggest the idea. However, certain aspects of the work can still be corroborated. The article,
“Color-based object recognition,” agrees that changes in color do degrade recognition accuracy,
so it follows that the claims of Finlayson et al. that an estimate of color change can aid in
color-based object recognition is viable (Gevers & Smeulders 1999). The coverage of this article
is both broad and deep. The authors test their homography estimation model with several
different goals, from color correction to color object recognition and color transfer. On all of
these problems, the authors compare how homography solves the problem to how other models,
specifically one dealing with 3D similarity, perform. Their focus is deeper in terms of the math
behind computing the color homography, as they explain the ALS method to turn four pairs of
corresponding colors into a homography. The article is one of the first to discuss the relationship
between color and homography, and its purpose is to bring the subject to light and encourage
further research in it. Because little background exists on the subject, the paper is very thorough
and detailed in its description of the mathematics behind the algorithm. Thus, it is quite difficult
for anyone not already very familiar with the theory of homography to understand.

Citation:
Fomin, I. S., Bakhshiev, A. V., & Gromoshinskii, D. A. (2017). Study of using deep learning
nets for mark detection in space docking images. ​Procedia Computer Science,​ ​103,​
59-66. ​https://doi.org/10.1016/j.procs.2017.01.011​.

Summary:
The problem addressed in this paper is of determining relative position of spacecrafts as they
approach to dock using external cameras and visual marks on the target. The current method for
this problem uses detected features and descriptors to create a 3D model of a spacecraft’s relative
position to the target. New models should be robust, and able to cope with distortions, low
quality images, and poor signal. This paper uses a model called Fast-RCNN, which quickly
detects objects in a given region of interest. The speed is due in part to training the kernel on a
full image set, rather than a more limited number. Testing on a dataset with noise, blurs, and
exposure changes, the authors found that the neural network object detector performed
significantly better than the cascade detector, the method currently used. Both models decreased
in accuracy the closer the spacecraft got to the target.

Annotation:
All authors are affiliated with the Russian State Scientific Center for Robotics and Technical
Cybernetics. Written in 2017, this article is still current on the prospects for using machine
learning to control autonomous spacecraft in docking procedures. The information in this article
can be corroborated. The information in this article can be corroborated. The authors of
“​Convolutional neural network-based spacecraft attitude control for docking port alignment”
agree that deep learning improves the accuracy of object detection in a docking procedure when
compared to currently used methods (Kim & Choi 2017). The coverage of this article is less
broad than it is deep. It focuses on a single issue: docking spacecraft. For this, the article mostly
focuses on the authors’ proposed model, which is a particularly fast version of a convolutional
neural network. With this algorithm, they test on various nodes using a dataset that includes poor
quality images to test the model’s robustness. Later, they extend their testing to include a
comparison of the CNN and the current, non-machine-learning models. This article was intended
for aerospace engineers who are working on automated spacecraft docking. Though the authors
do not intend that others use their exact model, as there is not much description of the
architecture, they do want to support the use of deep learning in general.
Citation:
Jordan, M. I. & Mitchell, T. M. (2015). Machine learning: Trends, perspectives, and prospects.
Science,​ ​349(​ 6245), 255-260. ​https://doi.org/10.1126/science.aaa8415​.

Summary:
According to the author, the foundational questions of machine learning are, “How can one
construct computer systems that automatically improve through experience?” and “What are the
fundamental statistical-computation- information-theoretic laws that govern all learning
systems?” Machine learning is a type of artificial intelligence by which one trains a system to
recognize future examples by showing it past examples of a certain input-output pair. Different
machine learning algorithms exist which implement this goal in different ways. For any machine
learning algorithm, the questions one should ask concern how accurate and robust the model is.
Different types of machine learning can be classified as supervised or unsupervised, or
reinforcement, an intermediate of the other two categories. Most current models blend the three
categories. Broad questions for machine learning still exist however. Humans and other animals
are lifelong learners, and can learn a variety of related tasks based on things they have already
learned. The question of whether a computer may work exactly the same as a human brain is still
theoretical. Ethical questions surrounding personal data and privacy have also arisen.

Annotation:
Jordan is a part of the the Department of Electrical Engineering and Computer Sciences at the
University of California, Berkeley, and Mitchell is affiliated with the Machine Learning
Department of Carnegie Mellon University. Written in 2015, this article is still current on the
prospects for machine learning and the problems which may hinder its full application in some
fields. The information in this article can be corroborated. “Learning to act by predicting the
future” corroborates the categories into which machine learning algorithms can be broadly
divided: supervised learning, unsupervised learning, and reinforcement learning (Dosovitskiy &
Koltun 2017). The coverage of this article is primarily broad. It presents an overview of machine
learning, giving definitions of the term, as well as of artificial intelligence and various specific
types of learning. It also covers broadly the big questions facing machine learning today, though
approaches them hypothetically and, because of their complexity, does not delve into solutions.
The article is intended to provide a broad overview of a topic. The language indicates that it is
written for collegiate level students, though there is nothing so technical that one needs much
background in computer science to understand it.

Citation:
Kang, L., Wei, Y., Xie, Y., Jiang, J. & Guo, Y. (2019). Combining convolutional neural network
and photometric refinement for accurate homography estimation. ​Digital Object
Identifier,​ ​7,​ 109460-109473. ​https://doi.org/10.1109/ACCESS.2019.2933635​.
Summary:
The problem which this paper sets out to address is the estimation of a 3x3 homography matrix
representing the shift between points of one image and another. The proposed model combines a
convolutional neural network (deep learning) and energy minimization. The training and testing
computation time is less than other CNNs used for homography because the network is relatively
small. The neural network calculates an initial homography estimate, and then gradient-descent
is used to further refine the prediction. Gradient-descent minimizes the distance between
corresponding pixels of the warped and original images. The authors test this new model on a
synthetic dataset, made by warping images from the MS-COCO database. They found that their
model performs better than traditional methods based on feature detection, and performs better
than regular CNNs.

Annotation:
Kang, the primary author, is an Assistant Professor at the College of Systems Engineering,
whose research specialities include computer vision, graphics, and image analysis, as well as the
real-world applications of such areas. Written in 2019, this article is still current on the success
of a convolutional neural network-based model for homography estimation. The information in
this article can be corroborated. The authors of this paper assume that creating a synthetic
dataset, by warping images from the MS-COCO dataset and creating stacked image pairs will be
a good test of how the homography estimation model works. The authors of “Deep image
homography estimation” create a validation dataset in an almost identical fashion (DeTone et al.
2016). The coverage of this article is both broad and deep. It details the mathematics behind all
aspects of the model, from the original architecture of the neural network to the gradient-descent
algorithm used to refine the accuracy of the model with backpropagation. However, it also
covers a range of other important topics, including how the testing dataset was generated, how
the model was trained and tested, and what other studies on similar models have found. The
purpose of this article is to present a working model to other computer scientists who may be
trying to accomplish a similar task. The authors not only provide enough math that one can
follow what their algorithm is doing, but they also provide numbered pseudocode in case the
reader wishes to write a similar model.

Citation:
Kim, S.-H. & Choi, H.-L. (2017). Convolutional neural network-based spacecraft attitude control
for docking port alignment. ​14th International Conference on Ubiquitous Robots and
Ambient Intelligence​. ​https://doi.org/10.1109/URAI.2017.7992783

Summary:
For autonomous docking, a spacecraft must calculate its relative position and angle to the target
from various sensors such as cameras and radar. Convolutional neural networks are used to avoid
the information loss of feature extraction, using the entire image rather than specific markers.
With a CNN, the spacecraft obtains information about the attitude of the target, and makes
corresponding adjustments to align the center of the camera with the target docking port. This
paper presents docking control as a classification problem; an image from a camera could either
require a lean to the left, lean to the right, lean upward, lean downward, or indicate a center
alignment. Thus, only two dimensions are considered, controlling yaw and pitch. The model is
trained on a three-dimensional spacecraft simulator, which uses OpenGL for graphics. In testing,
the authors found that the spacecraft’s attitude could be controlled by a CNN model, however,
further research is required to expand this conclusion to the roll axis.

Annotation:
Both authors are affiliated with the Department of Aerospace Engineering at KAIST, a national
research university in South Korea. Written in 2017, this article is still current on the applications
of convolutional neural networks to spacecraft docking. The authors of “Deep learning based
pose estimation in space,” agree that deep learning can be used to evaluate three-dimensional
locations in real time, with applications to a variety of space endeavours including docking and
navigation (Hirano et al. 2018). The coverage of this article is both broad and deep. It begins
with a literature review of a variety of proposed systems, some using machine learning and some
not, where automated navigation is attempted. The authors compare each previous study
discussed to what their own attempts to explore. When discussing their model, they explain in
depth how their model controls attitude, and why certain decisions were made regarding the
axes, markers, or spacecraft design models to use. The purpose of this article is to inform readers
of what the authors have determined in their study: that when modeling on two axes, yaw and
pitch, CNN is successful in controlling spacecraft attitude. However, they make sure to mention
a major shortcoming in their study, in that it does not account for roll. Thus, their audience is
other aerospace engineers who may be able to fill in this gap with further research.

Citation:
Kremer, J., Stensbo-Smidt, K., Gieseke, F., Pedersen, K. S., & Igel, C. (2017). Big universe, big
data: Machine learning and image analysis for astronomy. ​IEEE Intelligent Systems​, ​32​,
16-22. ​https://doi.org/10.1109/MIS.2017.40

Summary:
Astronomical surveys, images of the sky collected nightly, can be used to detect the periodic
brightness changes of transients through spectroscopic or photometric data. Multiple surveys
exist, some producing up to 90 TB of information per night. Spectroscopy uses photon counts to
identify chemical components of the observed object. Photometry captures images using only a
handful of broadband filters, and so is less informative than spectroscopy. However, photometry
is more sensitive than spectroscopy and thus works better for more distant objects, and can
capture multiple objects at the same time. Much of astronomical image analysis is still done by
humans, however with the masses of data produced automation would significantly lessen
associated costs. As astronomical images rely are often long-exposure to allow far away objects
to become visible, data is often very noisy. Machine learning may be used to recognize patterns
between the images and the physical properties of objects, and mark objects for re-observation.
When designing a machine learning model in astronomy, importance weighting can be used to
deal with sample selection bias, where more weight is given to examples in the training set
which have features underrepresented due to sample cost or availability.

Annotation:
Jan Kremer is a data scientist working at adform and specializing in machine learning and
computer vision. The other authors, Stensbo-Smidt, Gieseke, Pedersen, and Igel, are associated
with the Department of Computer Science at the University of Copenhagen. Written in 2017, this
article is still current on the uses of machine learning in analyzing images from astronomical
surveys. The information in this article can be corroborated. In the article, “Machine learning in
space: Extending our reach,” the authors agree that with machine learning, image analysis and
time series analysis may be done to detect target features or anomalies in data, (McGovern &
Wagstaff, 2011). The coverage of this article is both broad and deep. The authors broadly cover
the scenarios where machine learning offers a great advantage to traditional image analysis by
humans. However, they also explore deeper the ways in which needs related to the field affect
methodologies for data collection and preprocessing. The purpose of this article is to provide a
foundation and supporting information for an aerospace engineer or computer scientist to begin
using machine learning for their own image analysis. It covers topics such as how to ensure the
data is unbiased, where to find large collections of data, and how to decide between a variety of
models.

Citation:
Lavecchia, A. (2015). Machine-learning approaches in drug discovery: Methods and
applications. ​Drug Discovery Today​, ​20​(3), 318-331.
https://doi.org/10.1016/j.drudis.2014.10.012

Summary:
Virtual screening is a technique where a computer searches through chemical compounds and
virtually decides which will bind to a given protein receptor or enzyme which is necessary for
the drug. As data sets grow, machine learning can be applied to better judge patterns within the
information. Machine learning can process molecular properties including absorption,
distribution, metabolism, and excretion, among others. One of the most computationally
inexpensive methods is ligand-based virtual screening, where databases are searched to identify
compounds with particular components. Another method used is structure-based virtual
screening. However these often fail at “activity cliffs” points where small structural changes
drastically impact the compound’s behaviour. Support vector machines (SVM), decision trees
(DT), k-nearest neighbors (k-NN), naїve Bayesian models, and artificial neural networks (ANN)
can be used to determine if a certain chemical compound will be active or inactive for a
particular purpose.

Annotation:
Antonio Lavecchia is affiliated with the Department of Pharmacy of the Drug Discovery
Laboratory at the University of Napoli. Written in 2015, this article is still current on how
machine learning has been and can be applied to drug discovery. The information in this article
can be corroborated. The authors of “Machine learning in chemoinformatics and drug discovery”
agree, “‘activity cliffs’ where minor modification of functional groups causes an abrupt change
in activity violate this principle and can cause failure of QSAR models,” (Lo et al., 2018). The
coverage of this article is both broad and deep. The author covers a broad range of machine
learning algorithms which studies have previously applied to drug discovery, including SVM,
DT, k-NN, naïve Bayesian, and ANN, as well as discussing non-machine learning approaches to
virtual screening. For each of these algorithms, the author explains how one would develop the
algorithm as well as the benefits and drawbacks to that method in particular.​ T ​ he audience of this
article is those looking to use machine learning to aid in drug discovery, and the purpose is to
inform them of the benefits and drawbacks to the many different options for machine learning
algorithms. The majority of the paper is spent in a comparison of the accuracy, logic, and
interpretability of algorithms.

Citation:
Leung, M. K. K., Delong, A., Alipanahi, B., & Frey, B. J. (2015). Machine learning in genomic
medicine: a review of computational problems and data sets. ​Proceedings of the IEEE,​
104​(1), 176-197.​ https://doi.org/10.1109/JPROC.2015.2494198​.

Summary:
This paper focuses on the applications of machine learning in linking DNA with risks of certain
diseases, using gene expression, splicing, and protein binding as causal stepping stones between
the two. Genomic medicine, which encapsulates precision medicine (treatments personalized
based on an individual case, often using the patient’s genome), is growing in popularity in the
medical community. Currently, the most widely understood form of genomic medicine is the
identification of specific mutations and the diagnosis of the mutation’s effects on protein
synthesis. However, more research is needed in improving the prediction of whether or not a
mutation will harm the final protein. Machine learning has been used for issues like these in
interpreting the human genome. Much data exists, such as the Cancer Genome Atlas project,
which can be used for analysis. With the development of genome editing technology for use in
gene therapies, it is greatly important to understand the effects of certain, potentially beneficial,
changes. In theory, machine learning would be used to predict the phenotypes (including disease
risk) with biomarkers learned from the genomic data. Cell variables (gene expression, splicing,
and protein binding) are useful steps between DNA and phenotypes as they are more closely
linked to DNA than phenotypes are.

Annotation:
All authors are associated with the Department of Electrical and Computer Engineering at the
University of Toronto, and Brendan Frey is also associated with the Program on Genetic
Networks and the Program on Neural Computation and Adaptive Perception at the Canadian
Institute for Advanced Research in Toronto. Written in 2015, this article is still current on its
descriptions of the problems which still must be worked on and the importance of solving these
problems. The information in this article can be corroborated. The authors of “Machine learning
in genetics and genomics,” Maxwell W. Libbrecht and William Stafford Noble, agree that
machine learning is a valuable tool for identifying the genetic biomarkers of certain diseases
which relate to both the DNA and the phenotypes that result (2015). The coverage of this article
is both broad and deep. The authors begin by describing the growth of genomic medicine and
how it is necessary to understand the function of specific genes to be able to apply new gene
editing technologies to disease treatments. They describe in theory how machine learning could
be a tool to determine the phenotypes related to various genetic changes. The authors then
describe in detail the process, beginning with data collection, that one would take to go about
finding these links. The audience of this article is molecular biologists. The language used for
discussing machine learning is relatively basic, but the authors discuss in great detail the genetic
processes which the algorithms take advantage of. Understanding this, requires knowledge of
how proteins are formed and the mutations which may occur while this happens. The purpose of
this article is to encourage readers to not only conduct further research on machine learning as a
tool to help with disease prediction, but to actually work on the problem directly. The authors
spend a significant portion of the article discussing the process by which a machine learning for
this purpose would be created.

Citation:
Libbrecht, M. W., & Noble, W. S. (2015). Machine learning in genetics and genomics. ​Nature
Reviews Genetics​, ​16​(6), 321-332.​ https://doi.org/10.1038/nrg3920

Summary:
As the theoretical and technical capabilities of machine learning increase, applications for its
benefits must be studied. Relating to genomics, algorithms can be used to identify various
elements of a sequence, such as splice sites, promoters, enhancers, and transcription start sites,
among others. Some algorithms can determine the functionality of a certain gene, and others
predict gene expression. For different applications it may be necessary to vary the type of
algorithm used. Supervised learning is done on labeled training set, for example, a dataset where
splice sites are already labeled is used to determine where the splice sites are on future
sequences. Unsupervised learning is also useful, as it can be used when labeled data is
unavailable, or when a novel element is identified. Generative models are full distributions of
features and compare how the distributions differ, whereas discriminative models just separate
the two classes. The former is useful for an interpretation of data, as the model will show not
only the prediction but gives an explanation as to why. The latter, however, is performs better
when tasked with an infinite training data, and thus is more accurate for questions of prediction.
Current limitations of machine learning for genomics include the difficulty in distinguishing
which features are important and which are obsolete; there is a “curse of dimensionality”
encourages researchers to minimize the number of input dimensions. Many applications also
have the issues of imbalanced class sizes, where there are far more negative cases than positive
ones and difficulties arise if the population is too small. Finally, missing data values for some
patients, caused by defective cells, invalid genomic positions, or a mistake on the part of a
physician may complicate analysis.

Annotation:
Authors Maxwell W. Libbrecht and William Stafford Noble are both associated with the
University of Washington, and their expertise covers affiliations with the Department of
Computer Science and Engineering and the Department of Genome Sciences. Written in 2015,
this article is current on the possible applications of machine learning in the field of genomics.
The information in this article can be corroborated. In “Biomedical informatics and machine
learning for clinical genomics,” the authors agree that in determining which algorithm to use for
a given problem, “Semi-supervised methods use the structure of unlabeled data to improve label
predictions; this approach is especially useful for prediction tasks when data are plentiful but
labels are not,” (Diao et al., 2018). This article is very broad, yet is still deep in certain topics. It
discusses machine learning as a whole and its major classes, the tradeoffs and strategies which
are integral in the development of any model, and the future of machine learning in the field of
genomics. The authors discuss in depth the current limitations of supervised learning which
many inhibit success in applying these algorithms to clinical scenarios. This article is intended
for collegiate-level researchers with minimal background knowledge of machine learning. It
begins with an explanation of how machine learning algorithms work in theory, before beginning
to explain applications for these. The authors’ purpose is to encourage further research in
machine learning for gene-based questions. By describing both the prospects of this technology
and the current pitfalls, the reader is made aware of how important and beneficial overcoming
current limitations will be.
Citation:
Lo, Y.-C., Rensi, S. E., Torng, W., & Altman, R. B. (2018). Machine learning in
chemoinformatics and drug discovery. ​Drug Discovery Today,​ ​23(​ 8), 1538-1546.
https://doi.org/10.1016/j.drudis.2018.05.010

Summary:
Chemoinformatics is a broad field which combines large amounts of chemical data with novel
technologies for processing information. Machine learning can be applied here as a method of
predicting the properties of chemical compounds in new drugs, offering an alternative to the high
monetary and time costs of drug synthesis. Chemical graph theory presents an idea for how the
structures of chemicals influence their biological activity; it encodes information such as bond
length and other atomic properties. Chemical descriptors are features that can be extracted from
these, for example, atom counts, bond counts, molecular weight, and other variables. These
combine to form the chemical’s “fingerprint”, which machine learning algorithms use to for
digital representation. Most algorithms for this purpose rely on the assumption that small
changes in functional groups result in small changes in chemical behaviour. This approach fails
only at “activity cliffs” where small changes create radically different drugs. A multitude of
machine learning algorithms have been applied in various studies to this issue, including Naive
Bayes classifiers, regression analysis, k-Nearest neighbors, support vector machines, random
forest models, and artificial neural networks.

Annotation:
All authors are associated with the Department of Bioengineering at Stanford University. Written
in 2018, this article is current on the applications of machine learning in designing drugs based
on large amounts of chemical data. The information in this article can be corroborated. The
author of “Machine-learning approaches in drug discovery: Methods and applications,” agrees,
“‘activity cliffs’ in the activity landscape … [fall] outside the applicability domain of global
similarity approaches,” (Lavecchia, 2015). The coverage of this article is both broad and deep. It
includes a broad explanation of the types of machine learning algorithms which have been
applied to drug developments, providing mostly a general description of how the algorithm’s
logic works. It also includes a more in-depth explanation of how machine learning applies to this
problem in particular, with descriptions of chemical graph theory and how this provides the basis
for a digital feature-based prediction model.​ T​ he audience of this article is chemists of
researchers in chemoinformatics. Most basic explanations are dedicated to machine learning
topics, as opposed to explanations of the data itself which should be used. As the authors
assumes prior knowledge on the latter point, the article is aimed at this group. The purpose, then,
is to convince the audience to dedicate time, effort, or money to further research into machine
learning applications. The importance of this problem and the potential of the technology is
stressed in the article’s conclusions.

Citation:
Lorenz, R. D., Turtle, E. P., Barnes, J. W., Trainer, M. G., Adams, D. S., Hibbard, K. E., …
Bedini, P. D. (2018). Dragonfly: A rotorcraft lander concept for scientific exploration at
Titan. ​Johns Hopkins APL Technical Digest,​ ​34(​ 3), 374-387. Retrived from
https://dragonfly.jhuapl.edu/News-and-Resources/docs/34_03-Lorenz.pdf

Summary:
Titan, a moon of Saturn, is one of the most Earth-like bodies in the solar system, having a thick
nitrogen atmosphere. Methane works as a greenhouse gas, and forms clouds, rain, and rivers.
Life relies on compounds made up of carbon, nitrogen, oxygen, and hydrogen, along with trace
amounts of sulfur and phosphorus, and Titan is rich in the former two elements. These properties
make Titan a target for NASA’s fourth New Frontiers mission. The mission seeks to land a
mobile rotorcraft on the moon to then explore various points throughout the dry land of its
surface. One of the most recent lander flight design proposals has been a electrically-run
“over-under” quad octocopter. The lander will have two sample acquisition drills, one on each
skid, to allow for sample choice and redundancy. These are conveyed pneumatically through a
cyclone separator to the mass spectrometer. DraMS is used to analyze samples, however
DraGNS allows elemental composition of the ground under the lander to be determined without
any extractions. Other instruments include draGMet to sense atmospheric pressure, temperature,
and wind speed and direction, and the DragonCam for imaging. Specific sites to be investigated
include anywhere with evidence of pyrimidines or amino acids. One of the main design
challenges so far has been energy, and the design currently utilises output from a radioisotope
power source to recharge the battery.

Annotation:
The majority of authors, including Ralph D. Lorenz and Elizabeth P. Turtle work in the space
exploration sector of Johns Hopkins University Applied Physics Laboratory. The others are also
affiliated with the Dragonfly mission, through their work at the University of Idaho, NASA
Goddard Space Flight Center, Honeybee Robotics, Argo AI, Malin Space Science Systems, or
Penn State University. Two authors, Ralph Lorenz and Jason Barnes were the originators of the
idea for the Dragonfly mission. Written in 2018, this article is still current on the foundational
design of Dragonfly and the goals of its mission. The information in this article can be
corroborated. In “Dragonfly: Investigating the surface composition of Titan,” the authors agree
that the general goal of the Dragonfly mission is to explore various locations on Titan’s surface
to assess the presence of prebiotic compounds (Trainer et al., 2018).​ T ​ he coverage of this article
is both broad and deep. It begins with a broad overview of the Dragonfly lander’s tasks and the
reasons why it would be beneficial to collect information on Titan’s surface. It then provides a
deeper analysis of the optimal designs for the rotorcraft, including how it should fly, move,
sample, and recycle energy. The purpose of this article is to propose Dragonfly as both a
necessary and an achievable program, to support its candidacy as the next mission for NASA’s
New Frontiers program. The audience is scientists at NASA, who are well-informed on space
exploration and who will be choosing in which mission to invest hundreds of millions of dollars.

Citation:
Manchula, M. F., & Sandhoo, G. S. (2005). Rendezvous and docking for space exploration. ​1st
Space Exploration Conference: Continuing the Voyage of Discovery.​
https://doi.org/10.2514/6.2005-2716

Summary:
Previous programs, like the Space Shuttles, performed rendezvous using sensors added on as an
afterthought. However, it is becoming necessary to incorporate these into an entire system of
sensors, mechanisms, and thrusters in order to allow safe and successful autonomous rendezvous
and docking outside of low earth or lunar orbit. In developing these new systems, a balance must
be found between automation and human oversight. Relying on autonomous systems for tasks
which have long been successfully accomplished by humans may add potential risks to future
mission successes. However, the further a spacecraft is from Earth, the more detrimental
communications delay becomes. Critical flight management decisions must be made onboard.
Automated rendezvous fails when a problem occurs, and the current solution is to have a crew,
either onboard or at mission control, ready to intervene. Sensor data includes inertial
measurements of position, velocity, acceleration, range, range-rate, and bearing to the target. The
more distant the docking objects are, the more important range and direction become. The closer
the spacecraft is to the target the more accurate sensor measurements must be. Along with these
motion sensors, other mechanisms are needed to be able to function no matter the lighting,
darkness, target composition, or potential obstructions. Currently, no single sensor perfectly fits
these criteria. The docking or “mating” mechanism currently available either lack the ability to
become autonomous or require certain velocities and impact forces which are not compatible
with every spacecraft.

Annotation:
Both authors are aerospace engineers for NASA at the Johnson Space Center in Houston, Texas;
one specializes in rendezvous procedures and the other in automated vehicles. Though this article
was originally written in 2005, it still provides an explanation of the process of docking. It
presents the perspective held when automation of docking first began, which is useful for
determining its most foundational problems. The information in this article can be corroborated.
In “Safety control for spacecraft autonomous rendezvous and docking under motion constraints,”
the authors explain that a method for dealing with uncooperative obstacles or malfunctions must
be included in any automated rendezvous system. They express a similar point that the program
must work under any conditions (Dong et al., 2017). The coverage of this article is both broad
and deep. It includes a background of the phases of docking, from the rendezvous in orbit to the
“mating” mechanisms. For each phase, the authors describe the sensors which are needed and the
effect that varying distances from the target and from mission control on earth would have on
their utility. This article stresses the critical nature of docking for mission success to fulfill their
purpose of encouraging investment in the technologies to improve rendezvous sensors. Their
audience is systems engineers, highlighted by their call to action to create a more appropriate
design approach for the new semi-autonomous system. To this group, they explain different
phases of rendezvous and what sensors are needed for these, including why in each phase it is
necessary to balance the programmed machines with backup human controls.

Citation:
McGovern, A., & Wagstaff, K. L. (2011). Machine learning in space: Extending our reach.
Machine Learning​, ​84​(3), 335-340. ​https://doi.org/10.1007/s10994-011-5249-4

Summary:
Due to the increased risk, cost, and danger of sending humans increased distances from Earth,
most space missions in the future will be done robotically, through processes where autonomy
would be useful. Challenging factors associated with autonomy in space however, include
computational constraints of relatively old processors and memory capabilities that result from
necessary radiation-hardening, the extremely high cost of failure and scarce opportunity to repair
after launch, and the lag in communication between the mission’s parts in space and on Earth.
Though machine learning is only just beginning to be applied to space missions, it has useful
possibilities in obstacle avoidance, terrain navigation, detection of real-time events from orbit,
and identification of interesting formations to target for additional observation. These algorithms
are limited by how they are stored; some units, while they can store less information, are better
able to protect the information from radiation corruption. Improvements in autonomy and remote
decision making saves time and improves efficiency, which are valuable in spacecraft with
lifetimes limited by finite resources, degradation of components, environmental factors, and cost.

Annotation:
Author Amy McGovern is a student at the University of Oklahoma’s School of Computer
Science, and Kiri L. Wagstaff is associated with California Institute of Technology’s Jet
Propulsion Laboratory. Though written in 2011, this article is still current on the potential uses of
machine learning in space-related missions. The only topics which may have changed are the
exact scenarios algorithms have been developed for thus far. The information in this article can
be corroborated. In the article, “Risk-aware planetary rover operation: Autonomous terrain
classification and path planning.,” the authors agree that machine learning’s uses may extend the
lifetime of space missions, and they illustrate this point with Spirit, whose mission was
terminated after the Mars rover became stuck in soft sand, (Ono et al., 2015). The coverage of
this article is both broad and deep. It covers a list of many areas where machine learning may be
applied, and what forms of machine learning algorithms may be used. It then illustrates this with
an example of one specific application. It also dives deep into an explanation of the reasons why
machine learning can help extend the lifetime efficiency of missions and the challenges which
arise from using it in space, such as radiation corruption. The purpose of this article is to
highlight the potential uses for machine learning methods in space missions, while pointing out
challenges which exist inherently. Through their explanation of the benefits which may come
from their development, however, they are pushing for more implementation of autonomous
algorithms. Their audience is aerospace engineers in all fields of space exploration, who would
be researching and/or implementing these programs.

Citation:
Mohanty, V., Agrawal, S., Datta, S., Ghosh, A., Sharma, V., & Chakravarty, D. (2016).
DeepVO: A deep learning approach for monocular visual odometry. ​Computing Research
Repository​. ​https://arxiv.org/abs/1611.06069​.
Summary:
Convolutional neural networks have been used in object detection, segmentation, and
classification, outperforming traditional methods. However, the full potential of CNNs in visual
odometry has not been explored. Visual odometry is the process of controlling a robot based on
estimating its change in position. It is traditionally done by GPS, but recently it can be solved
with just images from the camera. It requires image acquisition, correction, feature detection,
feature tracking, and estimating motion. In the feature detection aspect, CNNs perform better
than traditional feature matchers such as ORB and SURF. Visual odometry can be done either
through feature-based methods which extract lines or corners, or appearance-based methods
which rely on pixel intensity, or a combination of both. When testing a CNN for this problem,
the authors determined that the more a computer knows about its environment, the better its
predictions are.

Annotation:
All authors are affiliated with the Indian Institute of Technology in Kharagpur, India. Written in
2016, this article is still current on the use of convolutional neural networks for visual odometry,
using data from sensors to change an automated object’s position. The information in this article
can be corroborated. The authors of the article “​Homography estimation from image pairs with
hierarchical convolutional networks” agree that convolutional neural networks are more
successful at determining perspective differences between two images than traditional, non
machine learning methods such as ORB (Nowruzi et al. 2017). The coverage of this article is
both broad and deep. The authors discuss broadly the different categories of visual odometry
methods, and list common examples of them. They discuss the training and testing of their model
in depth, including graphs of the predicted output vs ground truths, training and testing losses in
various trials, and deviation from ground truth over time.​ This article was intended for those
studying computer vision and deep learning, as it assumes a basic understanding of terms such as
“visual odometry” and “convolutional neural networks. It aims to educate readers on how a
specific application of convolutional neural networks in hopes it will be explored further.

Citation:
Nguyen, T., Chen, S. W., Shivakumar, S. S., Taylor, C. J., & Kumar, V. (2018). Unsupervised
deep homography: A fast and robust homography estimation model. ​IEEE Robotics and
Automation Letters,​ ​3(​ 3), 2346-2353.
https://ieeexplore.ieee.org/abstract/document/8302515

Summary:
A homography is a map between two images seemingly taken from different perspectives. They
can be applied in a number of fields, though to be useful in autonomous control the models must
be both fast and reliable. Homography can either be direct or feature-based, which use either
pixel to pixel matching or extracted key points respectively. The problems with feature-based
homography is that it requires there to be features that can be easily located on both images.
SIFT and ORB are two types of feature-based methods. The authors propose an unsupervised
deep learning algorithm to learn to find the best features for a given, specific dataset. The
authors' model combines the strengths of deep learning and the strengths of traditional
feature-based or direct methods. When tested, their unsupervised model performs better than a
comparable supervised model, and works faster and with better accuracy than traditional
methods. Their model could also handle large warpings and variations in brightness, important
for being robust in real-world applications.

Annotation:
All authors are affiliated with the General Robotics, Automation, Sensing, and Perception
Laboratory of the University of Pennsylvania. Written in 2018, this article is still current on the
effectiveness of unsupervised deep learning to evaluate image homography. The information in
this article can be corroborated. Authors of “​Deep image homography estimation” agree that a
convolutional neural network, and deep learning in general, is more successful at determining the
homography between two images than traditional methods (DeTone et al. 2016). The coverage
of this article is both broad and deep. It reviews five different approaches to determine which
performs better for image homography, and seven different combinations of different models and
CPU/GPUs to assess time vs. performance tradeoff. The authors dive deep into the specific
mathematics behind their models as well. This article is intended for computer scientists
well-versed in the field of machine learning or computer vision with the purpose of promoting
further research into unsupervised deep learning applied to image homography. Though its
conclusions are understandable without advanced technical knowledge, it discusses much of the
mathematics behind solving for a homography matrix, and training a neural network -- topics
which require the reader to understand linear algebra notation and the calculus behind
backpropagation.

Citation:
Nowruzi, F. E., Laganiere, R., & Japkowicz, N. (2017). Homography estimation from image
pairs with hierarchical convolutional networks. ​The IEEE International Conference on
Computer Vision (ICCV),​ 913-920.
http://openaccess.thecvf.com/content_ICCV_2017_workshops/w17/html/Nowruzi_Homo
graphy_Estimation_From_ICCV_2017_paper.html

Summary:
In computer vision, deep learning methods, when applied, have had great success compared to
traditional methods, particularly because machine learning does not need humans to hand-select
features. Convolutional neural networks in particular can extract image features, so present a
viable option to improve the accuracy of planar homography estimation. Traditional methods use
Direct Linear Transform to find corresponding points between the original and warped image.
These points can be found with methods like SIFT and ORB. The benefit of deep learning is that
it enables much stronger feature extraction. The model the authors create a novel framework for
image homography, which uses a smaller network, but warps the image between each step to
improve its accuracy and visualize how the model evolves. This is the characteristic which
makes the model hierarchical, as like iterative optimization, modules successively improve the
accuracy of the homography estimation. The final result is the sum of all previous modules’
estimates. The authors generate their dataset from MSCOCO images by randomly selecting the
corners of a 128x128 patch and perturbing each point. HomographyNet, a neural network
approach to image homography, represented a 21.37% error reduction from traditional methods
ORB and RANSAC. The proposed hierarchical network, by comparison was a 66.58% error
reduction.

Annotation:
Authors Nowruzi and Laganiere are affiliated with the Departments of Electrical Engineering
and Computer Science at the University of Ottawa, and Japkowicz with the Department of
Computer Science at American University. Written in 2017, this article is still current on the
performance of a hierarchical neural network for image homography in comparison with
HomographyNet, ORB, and RANSAC. The information in this article can be corroborated. The
authors of the article, “Deep Image Homography Estimation,” agree that the HomographyNet
uses convolutional neural networks to estimate image homography better than traditional
methods (DeTone et al. 2016). The coverage of this article is both broad and deep. The authors
discuss in depth the framework they created, explaining how they generated their dataset, the
loss functions used, and how the hierarchy in the model works. They extend their study to be
broad when they not only test that their model performs well, but compare it to a combination of
traditional, non-machine learning methods, as well as a previously created deep learning model,
HomographyNet. The authors’ purpose in writing this article is to present the model they created
to researchers looking to estimate image homography. They aim to prove that their model is one
of the best options by comparing its architecture, training loss, and accuracy with other
state-of-the-art methods.

Citation:
Ono, M., Fuchs, T. J., Steffy, A., Malmone, M., & Yen, J. (2015). Risk-aware planetary rover
operation: Autonomous terrain classification and path planning. ​IEEE Aerospace
Conference​. ​https://doi.org/10.1109/AERO.2015.7119022

Summary:
For planetary rovers, automated obstacle identification and avoidance is very helpful, even when
controlled primarily from the ground, because of the reduced risk and reduced cognitive load for
operators. This type of tool is built in two parts. First, terrain classification is done with machine
learning. Using this and a rapidly-exploring random graph, the rover can suggest courses which
minimize terrain risks. Inputs are both camera images and a digital elevation map, already
constructed. Learning from a random forest algorithm, terrain classification is done for each
pixel of the image. The algorithm differentiates five types of terrain: sand, bedrock, loose rock,
embedded angular rock, and embedded round rock. Terrain types and obstacles recognized, as
well as a cost map generated from data extracted from the DEM, are used in the path planner
algorithm. The rover aims to avoid driving over angular embedded rocks or rocks larger than 30
cm, and aims to minimize turning, which helps protect the wheels from unnecessary wear.
However, as there are many criteria to choose a path, such as length, tilt, or number of turns, the
algorithm submits multiple paths and the operator chooses one to execute. In the situation of
densely populated small yet sharp rocks, the planner runs into the issue that ensuring it keeps a
rover-length away from all of them, it has no path through. However, here the algorithm would
check that just the wheel track areas are free, predicting a greater number of feasible paths as the
small objects are allowed to pass between the wheels.

Annotation:
The majority of authors are associated with NASA’s Jet Propulsion Laboratory, and their
expertise covers robotic control, computer vision, wheel damage, and navigation. Written in
2015, this article is still current on machine learning techniques to minimize risk associated with
terrain navigation. The information in this article can be corroborated. In the article, “No GPS?
No problem! Exploring the dunes of Titan with Dragonfly using visual odometry,” the authors
agree that it is necessary to create tools which automate navigation by taking local terrain into
​ he coverage of this article is both
account to determine a feasible path (Witte et al., 2019).​ T
broad and deep. It begins with an overview of the inputs and output data and their formats, which
would be needed for path planning. Then, the authors proceed to discuss in detail each specific
algorithm within the whole tool, including how the terrain classification works, and the steps to
path planning itself. The purpose of this article is to present the authors’ developed and tested
tool for risk-relative terrain navigation, in order to support its application to actual rovers on
Mars. Thus, their audience is aerospace engineers working on such rovers, who would think this
design a candidate for implementation.

Citation:
Park, T. H., Sharma, S., & D’Amico, S. (2019). Towards robust learning-based pose estimation
of noncooperative spacecraft. ​2nd RPI Space Imaging Workshop​.
https://arxiv.org/abs/1909.00392

Summary:
Tracking spacecraft pose on-board is key to being able to automatically update controls. Current
methods involve recognizing features from the image, matched onto features of a known model,
in order to determine attitude information. However, machine learning does not use this
matching technique and is therefore more robust and less computationally expensive. The
authors propose instead using convolutional neural networks. They mention that neural networks
are prone to overfitting, and much testing must be done to meet accuracy requirements of a
real-world problem. Especially since acquisition of real space images is expensive, networks
must train on little variation of brightness, textures, or other potentially confounding image
features. The authors’ lab has developed the Testbed for Rendezvous and Optical Navigation
facility to test computer vision-based navigation algorithms. The model explained in this article
aims to extract keypoints and then estimating pose with a regression of the key points.

Annotation:
All authors are affiliated with the Space Rendezvous Laboratory (SLAB) at Stanford University,
the group which developed the Testbed for Rendezvous and Optical Navigation (TRON) Written
in 2019, this article is current on the applications of machine learning for on-board spacecraft
pose estimation. The information in this article can be corroborated. The authors of
“​Convolutional neural network-based spacecraft attitude control for docking port alignment”
agree that one challenge with machine learning is that it requires a lot of expensive space data to
be accurate (Kim et al. 2017). ​The coverage of this article is both broad and deep. It includes a
broad explanation of a proposed convolutional neural network architecture, and the steps it will
take to estimate the pose of a spacecraft. It explores deeper the challenges which are posed with a
neural network’s need for large quantities of representatively varied data and the many design
elements which went into TRON which aim to recreate a space environment. The purpose of this
article was to inform readers of the general work being done at SLAB, advocating for the
continued development of the TRON facility. The audience of the article is researchers at the
Space Imaging Workshop where this work was presented, though it is written at a level that
college students studying the topic could easily understand.

Citation:
Pathak, A. R., Pandey, M., & Rautaray, S. (2018). Application of deep learning for object
detection. ​Procedia Computer Science​, ​132​, 1706-1717.
https://doi.org/10.1016/j.procs.2018.05.144

Summary:
This article focuses on object detection using convolutional neural networks, which train by
themselves from pixels to objects, rather than with human-created feature extractors. Object
detection can either be single-class or multi-class. Challenges arise for occluded objects, difficult
lighting conditions, or unique poses. In convolutional neural network, convolutions of the image
and the given activation function are used to create feature maps. Eventually, feature maps are
flattened and processed in fully connected layers, to get an array of confidence scores for each
predicted class. CNNs also use a variety of pooling layers for subsampling feature maps. The
article provides a list of different databases which house images of various objects and deep
learning frameworks such as TensorFlow, which one might be able to use to create and train a
neural network. Object detection has the potential to track entities in surveillance, either
retroactively or in real time, and apply the ability to a variety of fields, such as security,
self-driving vehicles, or drones. The convolutional neural network is one of the more
state-of-the-art methods for object detection. With these, bounding-box regression determines the
location of objects, aiming to be accurate even with lighting changes, deformations, occlusions,
or variations in physical appearance of the same object. Shallow CNNs exist as faster
alternatives, and a popular framework of this type is YOLO: “You Only Look Once”. The future
of convolutional neural networks requires improved efficiency of algorithms or of GPUs in order
to better work in real time with videos 30-60 frames per second.

Annotation:
All authors are affiliated with the School of Computer Engineering at the Kalinga Institute of
Industrial Technology, and Pandey and Rautaray are assistant professors there. Written in 2018,
this article is still current on the development and architecture of convolutional neural networks
for object detection. The information in this article can be corroborated. The authors of “Scalable
Object Detection using Deep Neural Networks,” agree in their explanation of object detection,
writing that it can be done through regression: finding the coordinates of bounding boxes around
an identified object (Erhan et al. 2014).​ ​The coverage of this article is both broad and deep. The
article starts with a broad description of how convolutional neural networks work, as this
architecture differs greatly from shallow learning methods as well as other neural networks. It
continues with a deeper review of different available datasets which one could use to train an
object detection algorithm, as well as the types of algorithms supported by different deep
learning frameworks one could use. The audience of this article is machine learning model
developers, particularly those working where surveillance of some sort is needed. The authors
start from the basic description of a CNN and include different datasets and development
environments one could use, as their purpose is to help others create models for object detection.
They also stress the real-life applications of object detection in tracking animals, vehicles, or
people.

Citation:
Shangguan, Z., Wan, L., Zhang, J., & Dong, W. (2019). Vision-based object recognition and
precise localization for space body control. ​International Journal of Aerospace
Engineering.​ ​https://doi.org/10.1155/2019/7050915

Summary:
To control the motion of a spacecraft, one must first measure inertia, position, attitude, and
velocity, for which various sensors exist. In all applications of autonomous machines, computer
vision systems offer the ability of object localization, though edge detection and feature
matching. Current positioning systems fail in certain backgrounds or lighting conditions, though
deep learning has the potential to improve control speed and precision. For close procedures,
such as docking, precision is of utmost importance. To recognize an object in an image
successfully on a complex background, the authors’ system uses a convolutional neural network.
In rendezvous, the computer would recognize if the target is in view of the camera, and adjust
accordingly. On the training data, target recognition for the authors’ model had an accuracy of
99.976%. On the testing data, the accuracy was 99.847%, better than traditional object
recognition’s methods. The authors use a new algorithm for when the object is less than 1.5
meters away.

Annotation:
All authors are in both the Key Laboratory of Space Utilization and the Technology and
Engineering Center for Space Utilization, affiliated with the Chinese Academy of Sciences in
Beijing. Written in 2019, this article is still current on the use of neural networks on object
recognition in image analysis for docking spacecraft. The information in this article can be
corroborated. The article, “Machine learning in space: extending our reach,” agrees that with
machine learning “…features of interest…can be automatically identified and targeted for
additional observations” (McGovern & Wagstaff, 2011). The coverage of this article is both
broad and deep. The authors explain broadly what the need is for a model to detect objects from
an image. They explain why certain sensors work or don’t work at different distances and
conditions, and why machine learning offers a valid solution in their place. They also explain
deeply the process of creating and testing their model. Details are given on its structure and
layers, and data is provided for its training accuracy, testing accuracy, and verification methods.
The audience of this paper is other aerospace engineers who want to develop an algorithm for a
similar function, and the purpose is to provide expertise and resources found in the authors’
development of their model. Throughout their paper, technical language, specific to the field of
machine language, is used. The authors detail what activator functions they used, the number of
convolutional layers and fully connected layers in their neural network, the size of their training
data, and the number of iterations done in training. In addition, the authors have added their code
and data to the Scientific Data’s List of Recommended Repositories for readers to access and
use.

Citation:
Trainer, M. G., Brinckerhoff, W. B., Freissinet, C., Lawrence, D. J., Peplowski, P. N., Parsons, a.
M., … Stickle, A. M. (2018). Dragonfly: Investigating the surface composition of Titan.
NASA Technical Reports Server.​ Retrieved from
https://ntrs.nasa.gov/archive/nasa/casi.ntrs.nasa.gov/20180003047

Summary:
Dragonfly, part of NASA’s New Frontiers Program, aims to determine the material composition
of various locations on the surface of Titan. In particular, the rotorcraft would be investigating
prebiotic chemistry indicative of the development of extraterrestrial water-based or
hydrocarbon-based life. Titan is an ocean world, dominated by ice. Some of the locations that the
rotorcraft will explore were chosen for evidence of past liquid water. At each location, it will
measure the abundance of water ice, organic material, and minor elements pertaining to
biochemistry, without needing to extract a sample. Instead, it uses the Dragonfly Gamma-Ray
and Neutron Spectrometer (DraGNS). Based on this, a surface sample may be acquired for a
more detailed analysis with the Dragonfly Mass Spectrometer (DraMS).

Annotation:
The first listed author, M. G. Trainer, is affiliated with the Goddard Space Flight Center, and all
of her colleagues are associated with either a university or another physics laboratory. Written in
2018, this article is still current on the goals of the NASA’s Dragonfly mission. The information
in this article can be corroborated. In “Dragonfly: A rotorcraft lander concept for scientific
exploration at Titan,” the authors agree that DraMS is an instrument used to analyze samples
taken into the lander, whereas DraGNS detects elemental composition without any sampling
through a pulsed neutron generator to excite the gamma-ray signature (Lorenz et al., 2018).​ T ​ he
coverage of this article is both broad and deep. It begins with an explanation of Dragonfly’s
general goal of assessing the presence of prebiotic compounds on Titan. Then, it provides deeper
information on two main instruments which the probe will use to collect information, a
gamma-ray and neutron spectrometer and a mass spectrometer. This article is a conference
paper, thus its purpose is to present the plan of a specific topic, in this case what Dragonfly will
measure to achieve its function, in order to be invited to present more in depth research. Though
the language is not overly technical, the document type indicates it is intended for an audience
well-educated in space exploration fields.

Citation:
Westerhoff, J., Lessmann, S., Meuter, M., Siegemund, J., & Kummert, A. (2016). Development
and comparison of homography based estimation techniques for camera to road surface
orientation. ​IEEE Intelligent Vehicles Symposium (IV)1​ 034-1040.
https://doi.org/10.1109/IVS.2016.7535516​.

Summary:
Like in space, automated driving requires information about a vehicle's pose relative to the road
surface, to calculate distance, orientation, and roll movement. The position of a camera should
not be assumed to be static, as acceleration and road surface irregularities cause it rather to be
dynamic. To determine homography, generally either image registration (IR) or optical flow
(OF) is used. This paper offers a direct comparison of the two. While IR based methods tend to
be more accurate, OF methods are more commonly used. The authors test their models on both
rendered and real-world data sets. Image registration outputs the best roll and pitch angle
estimation results. However, optical flow, particularly the pyramidal "Lucase-Kanade" (PLK)
algorithm, performs with only slightly less accuracy but is much more computationally efficient.
While both are sufficient, the authors conclude OF is more desirable. However, the authors note
more research is needed into the two types of models' performance on data sets with occlusions.

Annotation:
Authors Westerhoff and Kummert are affiliated with the School of Electrical, Information and
Media Engineering, Lessmann is with the Intelligent Systems Group at the University of
Duisburg-Essen, and Meuter and Siegemund are affiliated with Delphi Electronics and Safety
Advanced Engineering. Written in 2016, this article is current on road surface orientation as one
real-world application of image homography estimation. The information in this article can be
corroborated. Authors of "Applying Kalman filtering to road homography estimation" agree that
Kalman filters can be used to estimate the final pitch and roll angle of the camera to the road
surface between two images (Klappstein et al. 2007). The coverage of this article is both broad
and deep. The authors start with a qualitative overview of what other researchers have concluded
about IR vs OF. Then, they dive into their own investigation, detailing the formulas and vector
diagrams throughout which they use to estimate pitch and roll angles. The purpose of this article
is to provide readers with a comparison of image registration and optical flow for homography
which they could not get elsewhere -- the authors state that theirs, to the best of their knowledge,
is the first study of its kind. Thus, their authors are engineers who would be applying some
method of homography for road surface pose estimation.

Citation:
Witte, I. R., Bekker, D. L., Chen, M. H., Criss, T. B., Jenkins, S. N., Mehta, N. L., … Thomas, J.
R. (2019). No GPS? No problem! Exploring the dunes of Titan with Dragonfly using
visual odometry. ​AIAA SciTech Forum​. ​https://doi.org/10.2514/6.2019-1177

Summary:
Past flybys, made by ​Voyager​ and ​Cassini​, and the ​Huygens​ probe, have collected data on
Titan’s atmosphere, imaged its surface at low-quality, and collected information about surface
composition at a singular landing site. The design is a dual-quadcopter that will be able to
explore different sites on the surface, not just the site at which the probe lands. On Titan, no high
resolution map exists, and there is low contrast in the terrain due to its thick, hazy atmosphere
diffusing light. Terrain relative navigation (TRN) has been applied to guided weaponry, and is
now being implemented in space exploration landers. A TRN algorithm would continuously
estimate the relative lateral position change of the rotorcraft using onboard imagery, then these
measurements would be used to determine velocity and to help navigate when flying. The
algorithm transforms two images to be in the same reference frame, and from this, difference in
lateral position can be determined.

Annotation:
All authors are members of staff in the space exploration sector at Johns Hopkins University
Applied Physics Laboratory. Their expertise includes space and missile defense applications,
embedded applications, and imaging systems. Written in 2019, this article is current on the uses
of and potential algorithms for terrain relative navigation on Titan. The information in this article
can be corroborated. In the article, “Risk-aware planetary rover operation: Autonomous terrain
classification and path planning,” the authors agree that, it is necessary to be able to have a
​ he
relatively self-reliant rover plan a safe path based on images it collects (Ono et al., 2015).​ T
coverage of this article is both broad and deep. It begins with a broad explanation of the
navigation problem, along with the factors which make certain methods impossible. Then, the
authors explain their proposed algorithm in depth, from the theory to its testing. The purpose of
this article is to provide its target audience, aerospace engineers who might be implementing a
similar algorithm, with an idea of how to do so. The majority of the article focuses on how one
would design, implement, and test the system, and in depth information as to how the algorithm
was created is provided.

Citation:
Yu, L., Yu, Z., & Gong, Y. (2015). An improved ORB algorithm of extracting and matching
features. ​International Journal of Signal Processing, Image Processing, and Pattern
Recognition, 8​(5), 117-126. ​https://doi.org/10.14257/ijsip.2015.8.5.12​.

Summary:
Image registration can be done either through pixel value, transform domain, or image features.
Of the three, image features are the most robust against differences in scale, rotation, brightness,
or noise. SIFT was the first mainstream algorithm of this category. SURF was created later,
differing itself from its predecessors with its faster computing time. FAST had the same
improvement in speed, but at the expense of precision. Thus ORB was developed in 2001 to
combine the feature detection of FAST and the feature description of BRIEF. In this paper, ORB
is used to extract features, and RANSAC is used to weed out wrong matches. To improve ORB,
the authors removed texture in certain areas of the image to reduce noise. Another model was
used to screen for false matching points, given a certain threshold. Finally, the authors used a
perspective transformation to match two images. They compare how SIFT, SURF, and ORB, and
ORB with RANSAC find points, and then compare all of those to their improved ORB method.
The improved ORB model had 80-90 more matches than the normal ORB, both of which took
far less time than SIFT or SURF.

Annotation:
All authors are affiliated with the College of Information and Communication Engineering at
Harbin Engineering University, where one is an associate professor focusing on image
processing. Written in 2015, this article is still current on what the ORB model does and how it
can be improved. The information in this article can be corroborated. The authors of
"Unsupervised deep homography: A fast and robust homography estimation model" agree that
ORB performs faster, though finds fewer matches, than SIFT (Nguyen et al. 2018). The coverage
of this article is both broad and deep. The authors discuss in depth the history and evolution of
feature matching algorithms, explaining the qualitative differences in function and performance
between SIFT, SURF, FAST, BRIEF, and ORB. The authors also explain the mathematics
behind both components of the ORB algorithm particularly in depth. This article was intended
for researchers who intend to use ORB to match features of an image. In it, the authors convey a
number of different elements which one can add to the original model which will make it more
effective at finding good matches.

Potrebbero piacerti anche