Multiaffine representations mediate tradeoff between generalization and parallel processing capacity in networks trained to multitask

Gregory Henselman-Petrusek, Tyler Giallanza, Sebastian Musslick, Jonathan Cohen

Keywords: Multitask learning, shared representations, generalization

Abstract: Artificial systems currently outperform humans in diverse computational domains, but none has achieved parity in speed and overall versatility of mastering novel tasks. A critical component to human success, in this regard, is the ability to redeploy and redirect data passed between cognitive subsystems (via abstract feature representations) in response to changing task demands. The present work formalizes this coordination procedure in terms of multiaffine functions of stimulus and control states. In experiments, the resulting model robustly predicts behavior and performance of multitasking networks on natural language data (MNIST) using common deep network architectures. Consistent with existing theory in cognitive control, representation structure varies in response to (a) environmental pressures for representation sharing, (b) demands for parallel processing capacity, and (c) tolerance for crosstalk. Implications for geometric (dimension, curvature), functional (automaticity, generalizability, modularity), and applied aspects of representation learning are discussed.

One-sentence Summary: We show that deep neural networks trained to perform multiple tasks can control the flow of feature information between layers via a multiaffine encoding scheme.

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics

Reviewed Version (pdf): /references/pdf?id=jNWwxRLsW5