√Čtienne Simon

Learning in heterogeneous network

A master's degree project on learning latent representations of heterogeneously typed input supervised by Ludovic Denoyer.

The goal was to learn pairs of encoder and decoder between a "view" of a concept and its representation in a latent space. The encoders and decoders were simple multi-layer perceptrons. It was trained with pairs of data representing the same concept (such as the word "cat" and the picture of a cat).

The code is in C++ and mostly use an interface used by the LIP6 at the time named NMLP, you can still look at it.

In French: Operational plan, Specifications.