A Torch autoencoder example

Extracting features from MNIST digits

Posted by Andrea Ferretti on November 6, 2015

In the previous example, we have seen how to set up a simple Torch application to recognize digits from the MNIST set.

In this iTorch notebook, we will see how neural networks can be used to extract features out of a dataset in a completely unsupervised way.

A simple way to do this is with an autoencoder, which is just a neural network with just a hidden layer, where the output and the input layer are the same. So we are asking of a way to to factorize the identity through a nonlinear layer. The nonlinearity guarantees that the network will not just learn a factorization of the identity matrix.

Moreover, one can force the hidden layer to be smaller than the other one, so that the factorization has to pass through a lower dimension, further enhancing the robustness of the learned model. Also, dropout can be used to guarantee that hidden neurons will learn independent features.

The features that we learn in this way are nonlinear functions of the input. We can see them explicitly by taking a basis vector in the hidden layer and applying the transformation to the output layer. In other words, features correspond to column vectors of the hidden/output matrix. This is particularly convenient when working with image datasets, because such vectors live in the same space as the original images, and as such we can plot them to see visually what we have extracted.

The features that we get here are not very good! Try to play around with the network architecture and parameters to extract more significant traits.

download notebook