CS236: Deep Generative Models
Last tended to on May 19, 2022

Generation: High level description => raw sensory outputs
Inference: raw sensory outputs => high level description

are learned from data. (Priors are necessary, but not as strict as in Graphics)

Data = samples
Priors = parametric (e.g. Gaussian prior), loss function, optimization algo, etc.

Image $x$ => [probability distribution $p$] => $p(x)$

Sampling from $p$ produces realistic samples

Discriminative model: input $X$ is given. learns $P(Y|X)$ (e.g., probability of bedroom given image)

Generative model: input $X$ is not given. learns $P(Y,X)$

They blur the line between generative and discriminative, because they also condition on some input features.

$P(X|Y=Bedroom)$

Superresolution: p(high-res signal | low-res signal)
Inpainting: p(full image | mask)
Colorization: p(color image | greyscale)
Translation: p(English text | Chinese text)
Text-to-Image: p(image | caption)

We are given a dataset of examples, e.g. images of cats. $P_{data}$ is the underlying distribution that generates the samples. We want to get a $P_{\theta }$ that is pretty close to $P_{data}$. $\theta$ is learned by a model, e.g. a neural net.

Generation: Then, if we sample $x_{new}\sim p_{data}(x)$, it should look like a cat.

Density estimation: $p(x)$ should be high if $x$ looks like a dog, and low otherwise.

Unsupervised: We should be able to learn what cats have in common, e.g. ears, tail, etc. (features!)

Consider an input with several components $X_1,...,X_n$ (these could be pixels in an image.) If $X_1,...,X_n$ are independent, then
$p(x_1,\dots ,x_n)=p\left(x_1\right)p\left(x_2\right)\cdots p\left(x_n\right)$

However, this assumption is too strong – oftentimes, components are highly correlated (like pixels in an image.)

Chain rule – fully general, no assumption on the joint. but the conditionals toward the end become large and intractable. Way too many parameters.
$p(S_1\cap S_2\cap \cdots \cap S_n)=p\left(S_1\right)p(S_2\mid S_1)\cdots p(S_n\mid S_1\cap \dots \cap S_{n-1})$

Need a better simplifying assumption in the middle…

p(y,x) = p(x|y)p(y) = p(y|x)p(x)

Generative: need to learn/specify both p(y), p(x|y)
Discriminative: just need to learn p(y|x) (X is always given)

Discriminative assumes that p(y|x;a) = f(x;a) (assumes that the probability distribution takes a certain functional form.)

E.g. logistic regression. Modeling p(y|x) as a linear combination of the inputs => squeeze with softmax. Decision boundaries are straight lines (assumption of logistic regression.) Logistic does not assume conditional independence like Naive Bayes does.

Using a conditional model is only possible when X is always observed. when some Xi are unobserved, the generative model allows us to compute p(Y|X_evidence) by marginalizing over unseen.

\(p\left(X_{i}=1 \mid x_{

simple: model as Bernoulli
more classes: model as Categorical
RNADE: continuous- model as mixture of Gaussians
Like FVSBN, but use a 1-hidden-layer neural net:
\(\mathrm{h}_{i}=\sigma\left(A_{i} \mathrm{x} $p(x_i\mid x_1,\cdots ,x_{i-1};\underset{\text{parameters }}{\underbrace{A_i,{\mathrm{c}}_i,{\mathit{\alpha }}_i,b_i}})=\sigma ({\mathit{\alpha }}_i{\mathrm{h}}_i+b_i)$

Problem: lots of redundant parameters. Solution: “tie” the weights:
Tie weights to reduce the number of parameters and speed up computation (see blue dots in the figure):
$$

\begin{array}{r} \mathrm{h}_{i}=\sigma\left(W_{\cdot, $$

$p(x_i\mid x_1,\cdots ,x_{i-1})=\sum _{j=1}^K\frac{1}{K}\mathcal{N}(x_i;{\mu }_i^j,{\sigma }_i^j)$
\(\mathrm{h}_{i}=\sigma\left(W_{\cdot, ${\hat{\mathit{x}}}_i=({\mu }_i^1,\cdots ,{\mu }_i^K,{\sigma }_i^1,\cdots ,{\sigma }_i^K)=f\left({\mathrm{h}}_i\right)$

Use masks to disallow certain paths in an autoencoder to make it autoregressive.

Solution: use masks to disallow certain paths (Germain et al., 2015). Suppose ordering is $x_2,x_3,x_1$.

1. The unit producing the parameters for $p\left(x_2\right)$ is not allowed to depend on any input. Unit for $p(x_3\mid x_2)$ only on $x_2$. And so on…
   
2. For each unit in a hidden layer, pick a random integer $i$ in $[1,n-1]$. That unit is allowed to depend only on the first $i$ inputs (according to the chosen ordering).
   
3. Add mask to preserve this invariant: connect to all units in previous layer with smaller or equal assigned number (strictly $<$ in final layer)
   

Challenge: the history for these autoregressive models keeps getting longer and longer. Ideally we’d just have a fixed-size “summary” of the history.

masked self-attention preserves autoregressive structure.

Masked convolutions preserve raster scan order.

Lol u can use these for adversarial attacks –

“distance” between two distributions, $p$ and $q$.

$D(p\Vert q)=\sum _{\mathrm{x}}p(\mathrm{x})\log \frac{p(\mathrm{x})}{q(\mathrm{x})}$
However, it’s not quite a “distance,” because it’s asymmetric. $D(p\Vert q)\ne D(q\Vert p)$

Intuition: we have a known distribution $q$, and we’re trying to minimize the distance to a target distribution $p$.

Bias limitation: If the hypothesis space of functions is very limited, we might not be able to represent the data distribution.

Variance limitation: If the hypothesis space is too expressive, it will overfit to the data.

How to prevent overfitting? Prefer “simpler” models (Occam’s razor.) Regularization in the objective function. Evaluate on validation set while training.

There are lots of variability in images due to high-level semantic factors: gender, eye color, pose, etc.

Idea: model these factors using latent variables $\mathbf{z}$.

If you choose $\mathbf{z}$ properly, p(x|z) could be a lot simpler than p(x).

We could identify the factors of variation using these generative models – e.g. p(eye color = blue | x)

$z\sim \mathcal{N}(0,I)$

$p(x\mid z)=\mathcal{N}({\mu }_{\theta }(z),{\mathrm{\Sigma }}_{\theta }(z))$ where ${\mu }_{\theta },{\mathrm{\Sigma }}_{\theta }$ are neural networks

We hope that z will capture useful factors of variation in an unsupervised manner. Training a classifier on top of z could be a lot easier.

Features computer via $p(z|x)$

A mixture of $k$ gaussians:
$z\sim \text{Categorical}(1,...,K)$
$p(x|z=k)=\mathcal{N}({\mu }_k,{\mathrm{\Sigma }}_k)$

A mixture of an infinite number of gaussians (since z is continuous):
$z\sim \mathcal{N}(0,I)$

$p(x\mid z)=\mathcal{N}({\mu }_{\theta }(z),{\mathrm{\Sigma }}_{\theta }(z))$ where ${\mu }_{\theta },{\mathrm{\Sigma }}_{\theta }$ are neural networks

Simple example:
${\mu }_{\theta }(z)=\sigma (Az+c)=(\sigma (a_1\mathbf{z}+c_1),\sigma (a_2\mathbf{z}+c_2))=({\mu }_1(z),{\mu }_2(z))$
${\mathrm{\Sigma }}_{\theta }(z)=\mathrm{diag}(\exp (\sigma (B\mathbf{z}+d)))=\left(\begin{array}{cc}\exp \left(\sigma (b_{1}z+d_1)\right)& 0\\ 0& \exp \left(\sigma (b_2\mathbf{z}+d_2)\right)\end{array}\right)$
$\theta =A,B,c,d$

Change of variables (1D case): If $X=f(Z)$ and $f(\cdot )$ is monotone with inverse $Z=f^{-1}(X)=h(X)$, then:
$$p_X(x)=p_Z(h(x))|h^{\prime }(x)|$$

This result comes from chain rule on the PDF.

This allows us to change the distribution of $X$ in interesting ways – start with a simple $Z$, transform with change-of-variables, and potentially get something much more complex than the prior.

$p_X(x)=p_Z(z)\frac{1}{f^{\prime }(z)}$

Intuition: if you’re expanding in one direction, you’re contracting in the other.

All this intuition carries over to random vectors (not just random variables.) See slides for more.

Change of variables (General case): The mapping between $Z$ and $X$, given by $f:{\mathbb{R}}^n\mapsto {\mathbb{R}}^n$, is invertible such that $X=\mathrm{f}(Z)$ and $Z=f^{-1}(X)$
$$p_X(\mathrm{x})=p_Z({\mathrm{f}}^{-1}(\mathrm{x}))\left|\det \left(\frac{\partial {\mathrm{f}}^{-1}(\mathrm{x})}{\partial \mathrm{x}}\right)\right|.$$

Equivalently, since $\det \left(A^{-1}\right)=\det (A{)}^{-1}$ for any invertible matrix $A$,
$$p_X(\mathrm{x})=p_Z(\mathrm{z}){\left|\det \left(\frac{\partial \mathrm{f}(\mathrm{z})}{\partial \mathrm{z}}\right)\right|}^{-1}.$$

Note: $Z$ has to have the same dimensionality as x (so that the mapping is invertible.)

It’s kinda like a VAE, but $p(x|z)$ is deterministic. And, crucially, we can directly evaluate the marginal likelihood $p(x)$; no integration

$$p_X(\mathrm{x}|\theta )=p_Z({\mathrm{f}}_{\theta }^{-1}(\mathrm{x}))\left|\det \left(\frac{\partial {\mathrm{f}}_{\theta }^{-1}(\mathrm{x})}{\partial \mathrm{x}}\right)\right|.$$

Note: $x,z$ need to be continuous and have the same dimension.

A flow of transformations: invertible transformations can be composed together.
$${\mathrm{z}}_m={\mathrm{f}}_{\theta }^m\circ \cdots \circ {\mathrm{f}}_{\theta }^1\left({\mathrm{z}}_0\right)={\mathrm{f}}_{\theta }^m\left({\mathrm{f}}_{\theta }^{m-1}(\cdots \left({\mathrm{f}}_{\theta }^1\left({\mathrm{z}}_0\right)\right))\right)\triangleq {\mathrm{f}}_{\theta }\left({\mathrm{z}}_0\right)$$

By change of variables
$$p_X(\mathrm{x};\theta )=p_Z({\mathrm{f}}_{\theta }^{-1}(\mathrm{x}))\prod _{m=1}^M\left|\det \left(\frac{\partial {\left({\mathrm{f}}_{\theta }^m\right)}^{-1}\left({\mathrm{z}}_m\right)}{\partial {\mathrm{z}}_m}\right)\right|.$$

By adding more “layers” in the transformation (i.e. a deeper neural net,) we get something increasingly complexified from the prior.

The prior $p(z)$ should be simple+efficient; e.g. isotropic Gaussian.

The transformations should have tractable evaluation in both directions.

Computing the likelihoods $p(x)$ and $p(z)$ require you to evaluate the determinant of an $n\times n$ Jacobian matrix; this is $O(n^3)$, and way to expensive to do in a learning loop.

Key idea: Choose transformations so that their Jacobians have a “special” structure; e.g. the determinant of a triangular matrix is the product of the diagonals; this is $O(n)$.

^how do we get that to happen? Some possibilities:

• Make $x_i=f_i(z)$ only depend on $z_{\le i}$.
  
• More efficient ways of computing Jacobians that are “close” to the identity matrix (Planar flows paper.)
  

Partition the variables $z$ into two disjoint subsets: $z_{1:d}$ and $z_{d+1:n}$

Forward mapping (z=>x):
$x_{1:d}=z_{1:d}$
$x_{d+1:n}=z_{d+1:n}+m_{\theta }(z_{1:d})$ (Where $m_{\theta }$ is a neural net)

Reverse mapping (x=>z):
${\mathrm{z}}_{1:d}={\mathrm{x}}_{1:d}$ (identity transformation)
${\mathrm{z}}_{d+1:n}={\mathrm{x}}_{d+1:n}-m_{\theta }\left({\mathrm{x}}_{1:d}\right)$

Jacobian:
$J=\frac{\partial \mathrm{x}}{\partial \mathrm{z}}=\left(\begin{array}{cc}{I}_d& 0\\ \frac{\partial {\mathrm{x}}_{d+1:n}}{\partial {\mathrm{z}}_{1:d}}& I_{n-d}\end{array}\right)$
$\det (J)=1$

Since the determinant is 1, it is a volume preserving transformation. (No expanding/contracting)

• Invertible
  
• Easy to compute
  
• Tractable marginal likelihood
  

Additive coupling layers can be composed together.

Final layer of NICE applies a rescaling transformation (so we can change the volume.)
Forward mapping $z\mapsto x:$
$$x_i=s_i{z}_i$$
where $s_i>0$ is the scaling factor for the $i$-th dimension.
Inverse mapping $x\mapsto z$ :
$$z_i=\frac{x_i}{s_i}$$
Jacobian of forward mapping:
$$\begin{array}{c}J=\mathrm{diag}(\mathrm{s})\\ \det (J)=\prod _{i=1}^n{s}_i\end{array}$$

Same as NICE, but rescaling happens at each layer.

Forward mapping $z\mapsto x:$

• ${\mathrm{x}}_{1:d}={\mathrm{z}}_{1:d}$ (identity transformation)
  
• ${\mathrm{x}}_{d+1:n}={\mathrm{z}}_{d+1:n}\odot \exp \left({\alpha }_{\theta }\left({\mathrm{z}}_{1:d}\right)\right)+{\mu }_{\theta }\left({\mathrm{z}}_{1:d}\right)$
  
• ${\mu }_{\theta }(\cdot )$ and ${\alpha }_{\theta }(\cdot )$ are both neural networks with parameters $\theta ,d$ input units, and $n-d$ output units $[\odot$ denotes elementwise product $]$
  

Inverse mapping $x\mapsto z$ :

• ${\mathrm{z}}_{1:d}={\mathrm{x}}_{1:d}$ (identity transformation)
  
• ${\mathrm{z}}_{d+1:n}=({\mathrm{x}}_{d+1:n}-{\mu }_{\theta }\left({\mathrm{x}}_{1:d}\right))\odot (\exp (-{\alpha }_{\theta }\left({\mathrm{x}}_{1:d}\right)))$
  

Jacobian of forward mapping:
$$\begin{array}{c}J=\frac{\partial \mathrm{x}}{\partial \mathrm{z}}=\left(\begin{array}{cc}{I}_d& 0\\ \frac{\partial {\mathrm{x}}_{d+1:n}}{\partial {\mathrm{z}}_{1:d}}& \mathrm{diag}(\exp \left({\alpha }_{\theta }\left({\mathrm{z}}_{1:d}\right)\right))\end{array}\right)\\ \det (J)=\prod _{i=d+1}^n\exp \left({\alpha }_{\theta }{\left({\mathrm{z}}_{1:d}\right)}_i\right)=\exp \left(\sum _{i=d+1}^n{\alpha }_{\theta }{\left({\mathrm{z}}_{1:d}\right)}_i\right)\end{array}$$
Non-volume preserving transformation in general since determinant can be less than or greater than 1

We can view autoregressive models as flow models.

Consider a Gaussian autoregressive model:
\[ p(\mathrm{x})=\prod_{i=1}^{n} p\left(x_{i} \mid \mathrm{x} such that \(p\left(x_{i} \mid \mathrm{x}_{1\) and constants for $i=1$
Sampler for this model:

• Sample $z_i\sim \mathcal{N}(0,1)$ for $i=1,\cdots ,n$
  
• Let $x_1=\exp \left({\alpha }_1\right)z_1+{\mu }_1$. Compute ${\mu }_2\left(x_1\right),{\alpha }_2\left(x_1\right)$
  
• Let $x_2=\exp \left({\alpha }_2\right)z_2+{\mu }_2$. Compute ${\mu }_3(x_1,x_2),{\alpha }_3(x_1,x_2)$
  
• Let $x_3=\exp \left({\alpha }_3\right)z_3+{\mu }_3\dots$
  

Flow interpretation: transforms samples from the standard Gaussian $(z_1,z_2,\dots ,z_n)$ to those generated from the model $(x_1,x_2,\dots ,x_n)$ via invertible transformations (parameterized by ${\mu }_i(\cdot ),{\alpha }_i(\cdot ))$

It’s possible to change a convolutional architecture to become invertible.

We can use masked convolutions to enforce ordering => Jacobian is lower triangular + easy to compute. If all the diagonal elements of Jacobian are positive, the transformation is invertible.

The point is, you can train a ResNet normally, then invert + use as a flow model.

Let $X=f_{\theta }(Z)$ be a flow model with Gaussian prior $Z\sim \mathcal{N}(0,I)=p_Z$, and let $\tilde{X}\sim p_{\text{data }}$ be a random vector distributed according to the true data distribution.

Flow models are trained with maximum likelihood to minimize the $\mathrm{KL}$ divergence $D_{\mathrm{KL}}(p_{\text{data }}\Vert p_{\theta }(x))=D_{\mathrm{KL}}(p_{\tilde{X}}\Vert p_X).$ Gaussian samples transformed through $f_{\theta }$ should be distributed as the data.

It can be shown that $D_{\mathrm{KL}}(p_{\tilde{X}}\Vert p_X)=D_{\mathrm{KL}}(p_{f_{\theta }^{-1}(\tilde{X})}\Vert p_{f_{\theta }^{-1}(X)})=D_{\mathrm{KL}}(p_{f_{\theta }^{-1}(\tilde{X})}\Vert p_Z).$Ideally, data samples transformed through $f_{\theta }^{-1}$ should be distributed as Gaussian (Hence “Gaussianizing.”) Then, we can easily turn it around and efficiently generate new data samples from Gaussian prior. So, how can we achieve this?

A neural net that tries to distinguish “real” from “fake” samples.

Maximize the two-sample test objective (in support of the hypotehsis $p_{\text{data }}\ne p_{\theta }$)
Training objective for discriminator:

$$\underset{D}{max}V(G,D)=E_{\mathbf{x}\sim p_{\text{data }}}[\log D(\mathbf{x})]+E_{\mathbf{x}\sim p_G}[\log (1-D(\mathbf{x}))]$$
For a fixed generator $G$, the discriminator is performing binary classification with the cross entropy objective

• Assign probability 1 to true data points $\mathbf{x}\sim p_{\text{data }}$
  
• Assing probability 0 to fake samples $\mathbf{x}\sim p_G$
  

Optimal discriminator
$$D_G^{\ast }(\mathbf{x})=\frac{p_{\mathrm{data}}(\mathbf{x})}{p_{\mathrm{data}}(\mathbf{x})+p_G(\mathbf{x})}$$

(We don’t want to use likelihoods, though.)

Directed, latent variable model with a deterministic mapping between $z$ and $x$, $G_{\theta }$.
Training objective for generator:
$$\underset{G}{min}\underset{D}{max}V(G,D)=E_{\mathbf{x}\sim p_{\text{data }}}[\log D(\mathbf{x})]+E_{\mathbf{x}\sim p_G}[\log (1-D(\mathbf{x}))]$$
For the optimal discriminator $D_G^{\ast }(\cdot )$, we have
$$

$$\begin{array}{c}V(G,D_G^{\ast }(\mathbf{x}))\\ =E_{\mathbf{x}\sim p_{\text{data }}}[\log \frac{p_{\text{data }}(\mathbf{x})}{p_{\text{data }}(\mathbf{x})+p_G(\mathbf{x})}]+E_{\mathbf{x}\sim p_G}[\log \frac{p_G(\mathbf{x})}{p_{\text{data }}(\mathbf{x})+p_G(\mathbf{x})}]\\ =E_{\mathbf{x}\sim p_{\text{data }}}[\log \frac{p_{\text{data }}(\mathbf{x})}{\frac{p_{\text{data }}(\mathbf{x})+p_G(\mathbf{x})}{2}}]+E_{\mathbf{x}\sim p_G}[\log \frac{p_G(\mathbf{x})}{\frac{p_{\text{data }}(\mathbf{x})+p_G(\mathbf{x})}{2}}]-\log 4\\ =\underset{2\times \text{ Jenson-Shannon Divergence }(\text{ JSD })}{\underbrace{D_{KL}[p_{\text{data }},\frac{p_{\text{data }}+p_G}{2}]+D_{KL}[p_G,\frac{p_{\text{data }}+p_G}{2}]}}-\log 4\\ =2D_{JSD}[p_{\text{data }},p_G]-\log 4\end{array}$$

$$

Btw, there are other divergences that we can use than Jenson-Shannon Divergence.

Sample minibatch of $m$ training points ${\mathbf{x}}^{(1)},{\mathbf{x}}^{(2)},\dots ,{\mathbf{x}}^{(m)}$ from $\mathcal{D}$ Sample minibatch of $m$ noise vectors ${\mathbf{z}}^{(1)},{\mathbf{z}}^{(2)},\dots ,{\mathbf{z}}^{(m)}$ from $p_z$ Update the discriminator parameters $\varphi$ by stochastic gradient ascent
$${\mathrm{\nabla }}_{\varphi }V(G_{\theta },D_{\varphi })=\frac{1}{m}{\mathrm{\nabla }}_{\varphi }\sum _{i=1}^m[\log D_{\varphi }\left({\mathbf{x}}^{(i)}\right)+\log (1-D_{\varphi }\left(G_{\theta }\left({\mathbf{z}}^{(i)}\right)\right))]$$
Update the generator parameters $\theta$ by stochastic gradient descent
$${\mathrm{\nabla }}_{\theta }V(G_{\theta },D_{\varphi })=\frac{1}{m}{\mathrm{\nabla }}_{\theta }\sum _{i=1}^m\log (1-D_{\varphi }\left(G_{\theta }\left({\mathbf{z}}^{(i)}\right)\right))$$
Repeat for fixed number of epochs…or until samples look good, lol.

GANs can be very challenging to train in practice.