Deep Multi-View Learning with Stochastic Decorrelation Loss
share
Multi-view learning aims to learn an embedding space where multiple views are either maximally correlated for cross-view recognition, or decorrelated for latent factor disentanglement. A key challenge for deep multi-view representation learning is scalability. To correlate or decorrelate multi-view signals, the covariance of the whole training set should be computed which does not fit well with the mini-batch based training strategy, and moreover (de)correlation should be done in a way that is free of SVD-based computation in order to scale to contemporary layer sizes. In this work, a unified approach is proposed for efficient and scalable deep multi-view learning. Specifically, a mini-batch based Stochastic Decorrelation Loss (SDL) is proposed which can be applied to any network layer to provide soft decorrelation of the layer's activations. This reveals the connection between deep multi-view learning models such as Deep Canonical Correlation Analysis (DCCA) and Factorisation Autoencoder (FAE), and allows them to be easily implemented. We further show that SDL is superior to other decorrelation losses in terms of efficacy and scalability.
READ FULL TEXT
