CSC412: Variational Autoencodersduvenaud/courses/csc412/...• Generating Sentences from a...
34
CSC412: Variational Autoencoders David Duvenaud
Transcript of CSC412: Variational Autoencodersduvenaud/courses/csc412/...• Generating Sentences from a...
CSC412: Variational Autoencoders
David Duvenaud
Class Projects
• Focus is on research skills, providing context for results, evidence for claims
• Excuse to start larger project
Variational Inference• Directly optimize the parameters phi of an
approximate distribution q(z|x, phi) to match p(z|x, theta)
• What if there is a local latent variable per-datapoint, and some global parameters? e.g. Bayesian PCA, generative image models, topic models
• Directly optimize the parameters phi_i of each approximate distribution q(z_i|x_i, phi_i) to match p(z_i|x_i, theta)
SVI Algorithm:
• 1. Sample z from q(z|x, phi)
• 2. Return log p(z, x | theta) - log q(z | x, phi)
• In this setting, only one set of latent params z, as in a Bayesian neural net
SVI w/ per-datapoint latents:
1. Sample x_i from dataset
2. Optimize phi_i to minimize KL(q(z_i | x_i, phi_i)|| p(z_i | x_i, theta))
3. Sample z from q(z_i|x_i, phi_i)
4. Return log p(z_i, x_i | theta) - log q(z_i | x_i, phi_i)
Variational Autoencoder:
1. Sample x_i from dataset
2. Compute phi_i as a function of x, and recognition params phi_r: phi_i = f(x, phi_r)
3. Sample z from q(z_i|x_i, phi_i)
4. Return log p(z_i, x_i | theta) - log q(z_i | x_i, phi_i)
Consequences of using a recognition network
• Don’t need to re-optimize phi_i each time theta changes - much faster
• Recognition net won’t necessary give optimal phi_i
• Can have fast test-time inference (vision)
• Can train recognition net with gradient descent
• Could also differentiate through optimization
Simple but not obvious
• It took a long time get here!
• Independently developed as denoising autoencoders (Bengio et al.) and amortized inference (many others)
• Helmholtz machine - same idea in 1995 but used discrete latent variables
Autoencoder Motivation
• Want compact representation of data
• x = dec(enc(x))
• Need to prevent enc = dec = identity
• So, add noise to encoding
• Gives VAE bound but with a free parameter
Benefits of compact latent code
• http://www.dpkingma.com/sgvb_mnist_demo/demo.html
• Nearby z’s give similar x
• Recent work on ‘disentangling’ latent rep
Code examples
• Show VAE code
Variations: Decoder
• Orginally, p(x|z) = N(x | dec(z, theta), sigma I)
• Final step has independence assumption, causes blurry samples
• p(x|z) can be anything: rnn, pixelRNN, real NVP, de_convolutional net
Variations• Decoder often looks like inverse of encoder
• Encoders can come from supervised learning
Learning Deconvolution Network for Semantic Segmentation http://arxiv.org/abs/1505.04366.
Text autoencoders
• Generating Sentences from a Continuous Space. Samuel R. Bowman, Luke Vilnis, Oriol Vinyals, Andrew M. Dai, Rafal Jozefowicz, Samy Bengio
TextVAE-Interpola*on
What is a molecule?Graph SMILES string
Repurposing text autoencoders
c1ccccc1 c1ccccc1
ENCODERNeural Network
DECODERNeural Network
CONTINUOUS MOLECULARREPRESENTATION
Latent Space
Discrete Structure SMILES
Discrete Structure SMILES
Can be trained on unlabeled data
Map of 220,000 Drugs
Map of 100,000 OLEDs
Random Organic LEDs
F S S F O OO S
N N N
HS S S S SH
HO SHH2N NH2
O
FHNH2N
NH3
F F
HO F
N NH2
OHSHSO
F S FH2S
OHHO OH
FHS
NH2
OH
O
N SH
O
NH2H2N FHF
SHS
S
F
H2O
SHSHS F
CH4
H2N OHN
F F
HO S S S S S S S S S S S SH
N
S N
NO
O
N
N
N N
N
O O
OO
S O
N
N N
NN N
N
NN
N
NN N
NN
NS
N
N N
N
N
N N
N
N
N
O
O
NN
S N
NNN
N
N NN
N
O
NN N
N
N
NNO
NN
NO
NN
N
NNN
N
Standard autoencoderVariational autoencoder
Molecules near
O
NHN+
O
-O
Cl
Cl
NH
NH
N+
O
O-Cl
Cl
Br
N
N+
O
-OCl BrCl
O
NH
N+
O
O-Cl
Cl
SHCl
ON+
O
-O
NH
Cl
Cl
O
NH
N
N+
O
-OCl
F
O
ON+
O
-OCl
Cl
HN
O
NH
N+
O
O-Br
Cl O
ON+
O
O-Cl
Br O
NH
NH
N+
O
O-Cl
Cl
O
O
HN
N+
O
-O
HN Cl
Br
O
NH
N+
O
O-HO
Cl
Cl
ON+
O
O-Cl
Cl Br
ON+
O
O-Br
Cl
Cl
HN
N+
O
O-
Cl
Cl
O
O
NH
N
N+
O
O-Cl
Cl
Br
O
NH
N+
O
O-F
NH
Cl
O
O
HN
N+
O
O-HO
F
F
NHN+
O
-OF
NH
Cl
BrO
NH
N+
O
O-Br
Cl
Cl
O
O
N+
O
O-Cl
Cl Br
O
NH
N+
O
O-Cl
Br
O
ON+
O
O-Cl
Cl O
O
NH
N+
O
O-Cl
Cl
O
ON+
O
O-Cl
F
Cl
O
NH
N+
O
O-
Cl
Cl
Cl
O
NH
N+
O
O-Cl
Cl O
ON+
O
O-Cl
Cl
O
O
NH
N+
O
O-Cl
Cl
Br
O
NH
N
N+
O
O-Cl
Cl
Cl
O
NN+
O
O-
Br
Cl
Cl
O
NH
N+
O
O-Cl
O Cl
ON+
O
O-Cl
Cl
Cl
O
NH
N+
O
O-Cl
Cl Cl
O NH
N+O O-
Cl
N
O NH
N+O O-
Br
O
NH
N+
O
O-Br
Cl Cl
O
NH
N+
O
O-
NH
Cl
Br
O
N+
O
-O
N NH
Cl
O
O
NH
N+
O
O-Cl
NCl
Cl
O
O
N+
O
O-
NH
Cl Cl
O
NH
N+
O
O-Cl
Cl Br
HN
N+
O
O-
Cl
Cl
Br
O
NH
N+
O
O-Br
Cl
Br
O
NH
N+
O
O-
Cl
Cl
O
NH
N+
O
O-Cl
OCl
ON+
O
O-Cl
Cl
Br
O
NH
N
N+
O
O-Cl
Cl
Cl
O
NN+
O
O-
Cl
Cl Cl
O
NH
N+
O
O-Cl
Cl
OH
NH2 O
N ONH
O
O
HN
HN O
O
O
O
N
O
O
O
HN
O
O
O
NH NO
O
O
ONH2
NHO
HN
HN
O
O
HN N
NHO
O
O
O
NH
O
O
O
O
NO
O
N
HN O
O
O O
NH
HNN
O
O
O
N
N
OO
O
NH2
N
HNO
O
OO
N
O
NHO
S
O
N
O
O
OO
OH
NH
O
OO
O
H2NS
HNO
O
O
OHS
NH
O
O
O
HO
SHN
HN O
O
O
O
NH2
NH
N
O
OH
OO
N
HN
O
O
O
F
N
N
O
O
O N
HN
NHOO
O
N
OO
HN
N
O
N
HNN
HNO
O
N
NO
OH
OO
NH
OHN
O
OO
H2N NN
HNH2NO
O
O
NH2
N
NHO
HN
Cl
O
O
N NH
OO
NHS
OH
F
N N
O
O
HN
HN
NHO
O
O
O
NN
NO
BrO
ONH
NN
OO
NH
O
OO
HO
HN NOO
O
N
NHO
O
N
Cl
Cl
O
O
ONH
HN
O
OO
O
HN
NHO
O
O
O
O
N
HO
HN
N O
O
HN
HN
NH
O
O
NH2
O
NH
N
HN
O
O
O
O
N
OHN
O
O
F
O
N
O
OO
NN
NH2
OO
NH2
NH
N
Cl
O
O
F
N
O
O
O
O
Molecules near
No chemistry-specific design!
Gradient-based optimization
Gradient-based optimization
• Can’t necessarily start from given molecule, need to encode/decode
• Can’t go too far from start, wander into ‘holes’ or empty regions
Encoder can look at decoder
• https://www.youtube.com/watch?v=Zt-7MI9eKEo
Recent Extensions• Importance-Weighted Autoencoders (IWAE) Burda,
Grosse, Salakhudtinov
• Mixture distributions in posterior
• GAN-style ideas to avoid evaluating q(z|x), p(x|z), even p(z)
• Normalizing flows: Produce arbitrarily-complicated q(z|x)
• Incorporate HMC or local optimization to define q(z|x)
Generative Adversarial Networks
• Also a latent-variable model: x = f(z), z from N(0,I)
• Trained adversarially, can also optimize p(x)
• Recent work on adversarially-trained VAEs:
• Match p(z, x) to q(z, x)
• Sample z from p(z), x from p(x|z)
• Sample x from data, z from q(z|x)
![Abstract arXiv:2007.03629v2 [cs.LG] 8 Jul 2020Strong Generalization and Efficiency in Neural Programs Yujia Li, Felix Gimeno, Pushmeet Kohli, Oriol Vinyals DeepMind {yujiali,fgimeno,pushmeet,vinyals}@google.com](https://static.fdocuments.in/doc/165x107/5fb4e3818ec2426c9f6bf53b/abstract-arxiv200703629v2-cslg-8-jul-2020-strong-generalization-and-eficiency.jpg)
