MADE: Masked Autoencoder for Distribution Estimation

MADE: Masked Autoencoder for Distribution Estimation
Reading Time: 8 minutes

These days, everyone is talking about GANs like BigGAN and StyleGAN, and their remarkable and diverse results on massive image datasets. Yeah, the results are pretty cool! However, this has led to a decline in the research of other generative models like Autoregressive models and Variational Autoencoders. So today, we are going to understand one of these unnoticed generative models: MADE.

Generative models are a big part of deep unsupervised learning. They are of two types—Explicit models, in which we can explicitly define the form of the data distribution, and Implicit models, in which we cannot explicitly define the density of data. MADE is an example of a tractable density estimation model in explicit models. Its aim is to estimate a distribution from a set of examples.

The model masks the autoencoder’s parameters to impose autoregressive constraints: each input can only be reconstructed from previous inputs in a given ordering. Autoencoder outputs can be interpreted as a set of conditional probabilities, and their product, the full joint probability.

Autoregression is a time series model that uses observations from previous time steps as input to a regression equation to predict the value at the next time-step. 

Autoencoders

Autoencoder is an unsupervised neural network that learns how to compress and encode data efficiently then learns how to reconstruct the data back from the reduced encoded representation to a representation that is as close to the original input as possible.

MADE: Masked Autoencoder for Distribution Estimation
Autoencoder for MNIST

The process of transforming input x to latent representation z is called the encoder and from latent variable z to reconstructed version of the inputMADE: Masked Autoencoder for Distribution Estimation  is referenced as the decoder.

Lower dimensional latent representation has lesser noise than input and contains essential information of the input image. So the information can be used to generate an image that is different from the input image but within the input data distribution. By computing dimensionally reduced latent representation z, we are ensuring that the model is not reconstructing the same input image.

Now we want to impose some property on autoencoders, such that its output can be used to obtain valid probabilities. By using autoregressive property, we can transform the traditional autoencoder approach into a fully probabilistic model.

Let’s suppose we are given a training set of examples MADE: Masked Autoencoder for Distribution Estimation .Here MADE: Masked Autoencoder for Distribution Estimation  and MADE: Masked Autoencoder for Distribution Estimation , because we are concentrating on binary inputs. Our motivation is to learn a latent representation, by which we can obtain the distribution of these training examples using deep neural networks.

Suppose the model contains one hidden layer and tries to learn h(x) from its input x such that from it, we can generate reconstruction MADE: Masked Autoencoder for Distribution Estimation  which is as close as possible to x. Such that,

MADE: Masked Autoencoder for Distribution Estimation

Where W and V are matrices, b and c are vectors, g is a non-linear activation function and sigm is a sigmoid function.

Cross-entropy loss of above autoencoder is,

MADE: Masked Autoencoder for Distribution Estimation

We can treat  MADE: Masked Autoencoder for Distribution Estimationas the model’s probability that MADE: Masked Autoencoder for Distribution Estimation is 1, so l(x) can be understood as a negative log-likelihood function. Now the autoencoder can be trained using a gradient descent optimization algorithm to get optimal parameters (W, V, b, c) and to estimate data distribution. But the loss function isn’t actually a proper log-likelihood function. The implied data distributionMADE: Masked Autoencoder for Distribution Estimation  isn’t normalizedMADE: Masked Autoencoder for Distribution Estimation . So outputs of the autoencoder can not be used to estimate density.

Distribution Estimation as Autoregression

Now we want to impose some property on autoencoders, such that its output can be used to obtain valid probabilities. By using autoregressive property, we can transform the traditional autoencoder approach into a fully probabilistic model.

We can write joint probability as a product of their conditional probabilities by chain rule,

MADE: Masked Autoencoder for Distribution Estimation

DefineMADE: Masked Autoencoder for Distribution EstimationandMADE: Masked Autoencoder for Distribution Estimation . So now the loss function in the previous part becomes a valid negative log-likelihood function.

MADE: Masked Autoencoder for Distribution Estimation

Here each outputMADE: Masked Autoencoder for Distribution Estimation must be a function taking as input MADE: Masked Autoencoder for Distribution Estimation only and giving output the probability of observing value MADE: Masked Autoencoder for Distribution Estimation at theMADE: Masked Autoencoder for Distribution Estimation  dimension. Computing above NLL is equivalent to sequentially predicting each dimension of input x, so we are referring to this property as an autoregressive property.

Masked Autoencoders

Since outputMADE: Masked Autoencoder for Distribution Estimation  must depend on the preceding inputsMADE: Masked Autoencoder for Distribution Estimation , it means that there must be no computational path between output unitMADE: Masked Autoencoder for Distribution Estimation  and any of the input units MADE: Masked Autoencoder for Distribution Estimation

So we want to discard the connection between these units by element-wise multiplying each weight matrix by a binary mask matrix, whose entries that are set to ‘0’ correspond to the connections we wish to remove.

MADE: Masked Autoencoder for Distribution Estimation

Where MW and MV are mask matrices of the same dimension as W and V respectively, now we want to design these masks in a way such that they satisfy the autoregressive property.

To impose the autoregressive property, we first assign each unit in the hidden layer an integer m between 1 and D-1 inclusively. The kth hidden unit’s number m(k) represents the maximum number of input units to which it can be connected. Here values 0 and D are excluded because m(k)=0 means it is a constant hidden unit, and m(k)=D means it can be connected to maximum D input units, so both the conditions violate the autoregressive property.

MADE: Masked Autoencoder for Distribution Estimation

There are few things to notice in the above figure,

  • Input 3 is not connected to any hidden unit because no output node shouldn’t depend on it.
  • Output 1 is not connected to any previous hidden unit, and it is estimated from only the bias node.
  • If you trace back from output to input units, you can clearly see that autoregressive property is maintained.

Let’s consider Multi-layer perceptron with L hidden layers,

For every layerMADE: Masked Autoencoder for Distribution Estimation, ml(k) stands for the maximum number of connected inputs of the kth unit in the lth layer. In above figure, value written in each node of MADE architecture represents ml(k).

The constraints on the maximum number of inputs to each hidden unit are encoded in the matrix masking the connections between the input and hidden units:

MADE: Masked Autoencoder for Distribution Estimation

And these constraints are encoded on output mask matrix:

MADE: Masked Autoencoder for Distribution Estimation

WhereMADE: Masked Autoencoder for Distribution Estimation , MADE: Masked Autoencoder for Distribution Estimation andMADE: Masked Autoencoder for Distribution Estimation. Note that ≥ becomes > in output mask matrix. This thing is vital as we need to shift the connections by one. The first output x2 must not be connected to any nodes as it is not conditioned by any inputs.

We set ml(k) for every layerMADE: Masked Autoencoder for Distribution Estimation by sampling from a discrete uniform distribution defined on integers from mink’ ml-1(k’) to D-1 whereas m0 is obtained by randomly permuting the ordered vector [1,2,…,D].

MV,W = MVMW1MW2…MWL represents the connectivity between inputs and outputs. Thus to demonstrate the autoregressive property, we need to show that MV,W is strictly lower diagonal, i.e. MV,Wd’,d is 0 if d'<=d.

 

Let’s look at an algorithm to implement MADE:

MADE: Masked Autoencoder for Distribution Estimation
Pseudocode of MADE

Deep NADE models require D feed-forward passes through the network to evaluate the probability p(x) of a D-dimensional test vector, but MADE only requires one pass through the autoencoder.

Inference

Essentially, the paper was written to estimate the distribution of the input data. The inference wasn’t explicitly mentioned in the paper. It turns out it’s quite easy, but a bit slow. The main idea (for binary data) is as follows:

  1. Randomly generate vector x, set i=1
  2. Feed x into autoencoder and generate outputsMADE: Masked Autoencoder for Distribution Estimation for the network, set p =MADE: Masked Autoencoder for Distribution Estimation .
  3. Sample from a Bernoulli distribution with parameter p, set input xi = Bernoulli(p).
  4. Increment i and repeat steps 2-4 until i > D.

The inference in MADE is very slow, it isn’t an issue at training because we know all x<d to predict the probability at dth dimension. But at inference, we have to predict them one by one, without any parallelization

MADE: Masked Autoencoder for Distribution Estimation

Left: Samples from a 2 hidden layer MADE. Right: Nearest neighbour in binarized MNIST.

Though MADE can generate recognizable 28X28X1 images on MNIST dataset, it is computationally expensive to generate high dimensional images from a large dataset.

Conclusion

MADE is a straightforward yet efficient approach to estimate probability distribution from a single pass through an autoencoder. It is not capable of generating comparably good images as that of state-of-the-art techniques (GANs), but it has built a very strong base for tractable density estimation models such as PixelRNN/PixelCNN and Wavenet.  Nowadays, Autoregressive models are not used in the generation of images and it is one of the less explored areas in generative models. Nevertheless, its simplicity makes room for the advancement of research in this field.

5 4 votes
Article Rating

Leave a Reply

0 Comments
Inline Feedbacks
View all comments
CEV - Handout
0
Would love your thoughts, please comment.x
()
x
%d bloggers like this: