Sei sulla pagina 1di 29

Disentangling Factors of Variation

Adarsh k
Sc17m001
MTECH-MACHINE LEANRNIGN
 Jensen's Inequality

 KL Divergence – Measure of dissimilarity of 2 probability


distribution
 Data – Example-Image

Matrix of N*N ,We flatten the matrix and represent it as a vector in RN^2
 Similarly for different images there will be
different vector representation,(Suppose we
have a collection 200 images of handwritten
3),We define a Probability measure over
Random Variable(Random Vector).
 Primary objective - Learn this distribution
 Why?
 1) Can Generate new Images
 2)Detect Anomalies
 Simple example
Hard to optimize, lets look at alternative approach to make optimization
easier
 Latent Variables-set of variables that
correspond to some feature of my
image/data but not directly observable.
 Eg-given a handwritten 3 ,latent variable
might be [identity of digit,shape,size,….]
 These are not directly measureable,hence
hidden/latent.
Introduction of latent Variables made the model simpler to optimize, now
using this model we find the parameters theta.(EM algorithm)
Let q be distribution over the bounds
 We need to learn distribution of set of
images, we represent that distribution my a
mixture of several Gaussians
 We need to find parameters of these
gaussians,we use lower bound and use the
EM algorithm to find these parameters
 How to model p(x)?
 Mixture of several Gaussians(Restrictive).
 Mixture of infinitely many Gaussians.
Suppose we somehow know the distribution over latent variable t.How to
generate new images?
CNN = A Function which has parameters w ,that takes t as its input and outputs
mean and std deviation of p(x/t,w)
 How to find w of that function approximator?
 EM algorithm!!
 How to remove the hypothetical assumption
that t is given?
 Find t .
 Disentangling factors of variation and using it
for few shots learning.
 What is disentangling?
 Given an image say handwritten 3,the current
methods give highly entangled
representation(t),my objective is to learn
disentangled representation(t).
 Example – given a handwritten 3 ,current
representation might give latent variables t
which may not make sense to a human being.
 While in my representation each component
of t correspond to separate factors
 Eg – [size of digit, tilt of digit ,thickness of
stroke]
 EX-2 ,Given a image of cat,
 Latent variable = [size,breed,color,……..]
 Say we learnt handwritten 3 ,we get a latent
variable representation say
 [identity of digit ,orientation,size,thickness]

 Now to learn a new digit(say 5) we might be


able to reuse some of the previously learnt
distributions for example in this case
 [..,orienation,size ,thickness]

Potrebbero piacerti anche