Optimal binary encoding through convolutional autoencoders

The widespread adoption of machine learning (ML) tools deployed at large scale has created an interest in new hardware for ML. LightOn, a French start up based in Paris, has created an optical coprocessor, the OPU (optical processing unit), that can accelerate such computations at large scale while using less power. The input of the OPU must be binary. It is hard to design a binary encoder that works well on natural data, so we investigate if is possible to learn it from them, and to transfer the encoder that we learned to other datasets. The aim of this work is to explore a deep learning model called autoencoder (AE) to build a binary AE skim which is cheap, scale invariant and input independent. In particular we focus on convolutional AE, developing its structure while choosing harder and harder datasets, proving it as a powerful tool in the binary encoding task. In the last part of this work we discuss on how to engineer a better loss function and on how to use a Generative Adversarial Network (GAN) to train the AE.