Analysis of an Adversarial Approach to Blind Source Separation

In this work, we analyze the adversarial network proposed by Brakel and Bengio at [5] to solve the problem of independent component analysis (ICA). Guided by a discriminator of independence, a linear autoencoder learns to codify a set of samples into estimates of their independent components. This is achieved by training the autoencoder to "fool the discriminator" and generate a code of latent variables. The present study focuses on linear mixtures, having the JADE and FastICA algorithms as benchmarks, but the paradigm has a significant potential of extension to more general scenarios.