Multi-task and multi-domain learning with tensor networks

Abstract

We propose a tensor network that can learn to perform multiple tasks by adjusting the factors of each layer. Most of the existing methods for multi-task learning train a single network to extract task-specific features and subsequent prediction. We propose to use a single network with task-specific transformations that can extract task-specific features and perform task inference with small memory overhead. In particular, we transform features using low-rank updates in the convolution kernels. We present experiments on different datasets for multi-task and multi-domain learning and demonstrate that our method achieves state-of-the-art performance with minimal memory overhead compared to existing methods.