Disentangling neural network representations for improved generalization

Abstract

Despite the increasingly broad perceptual capabilities of neural networks, applying them to new tasks requires significant engineering effort in data collection and model design. Generally, inductive biases can make this process easier by leveraging knowledge about the world to guide neural network design. One such inductive bias is disentanglment, which can help preven neural networks from learning representations that capture spurious patterns that do not generalize past the training data, and instead encourage them to capture factors of variation that explain the data generally. In this thesis we identify three kinds of disentanglement, implement a strategy for enforcing disentanglement in each case, and show that more general representations result. These perspectives treat disentanglement as statistical independence of features in image classification, language compositionality in goal driven dialog, and latent intention priors in visual dialog. By increasing the generality of neural networks through disentanglement we hope to reduce the effort required to apply neural networks to new tasks and highlight the role of inductive biases like disentanglement in neural network design.