In this work we study the distributed representations learnt by generative neural network models. In particular, we investigate the properties of redundant and synergistic information that groups of hidden neurons contain about the target variable. To this end, we use an emerging branch of information theory called partial information decomposition (PID) and track the informational properties of the neurons through training. We find two differentiated phases during the training process: a first short phase in which the neurons learn redundant information about the target, and a second phase in which neurons start specialising and each of them learns unique information about the target. We also find that in smaller networks individual neurons learn more specific information about certain features of the input, suggesting that learning pressure can encourage disentangled representations.

Recent advances in self-supervised learning through contrastive training have shown that it is possible to learn a competitive speech recognition system with as little as 10 minutes of labeled data. However, these systems are computationally expensive since they require pre-training followed by fine-tuning in a large parameter space. We explore the performance of such systems without fine-tuning by training a stateof-the-art speech recognizer on the fixed representations from the computationally demanding wav2vec 2.0 framework. We find performance to decrease without fine-tuning and, in the extreme low-resource setting, wav2vec 2.0 is inferior to its predecessor. In addition, we find that wav2vec 2.0 representations live in a low dimensional subspace and that decorrelating the features of the representations can stabilize training of the automatic speech recognizer. Finally, we propose a bidirectional extension to the original wav2vec framework that consistently improves performance.

No abstract

No abstract