Learning Multiple Tasks with Deep Relationship Networks

Deep networks trained on large-scale data can learn transferable features to promote learning multiple tasks. As deep features eventually transition from general to specific along deep networks, a fundamental problem is how to exploit the relationship across different tasks and improve the feature transferability in the task-specific layers. In this paper, we propose Deep Relationship Networks (DRN) that discover the task relationship based on novel tensor normal priors over the parameter tensors of multiple task-specific layers in deep convolutional networks. By jointly learning transferable features and task relationships, DRN is able to alleviate the dilemma of negative-transfer in the feature layers and under-transfer in the classifier layer. Extensive experiments show that DRN yields state-of-the-art results on standard multi-task learning benchmarks.