Brain insights improve RNNs' accuracy and robustness for hierarchical control of continually learned autonomous motor motifs

Laureline Logiaco, G Sean Escola

Keywords: neuroscience, dynamical systems, thalamocortical architecture, motor preparation, continual learning, hierarchical continuous motor control, out-of-distribution generalization, robustness

Abstract: We study the problem of learning dynamics that can produce hierarchically organized continuous outputs consisting of the flexible chaining of re-usable motor ‘motifs’ from which complex behavior is generated. Can a motif library be efficiently and extendably learned without interference between motifs, and can these motifs be chained in arbitrary orders without first learning the corresponding motif transitions during training? This requires (i) parameter updates while learning a new motif that do not interfere with the parameters used for the previously acquired ones; and (ii) successful motif generation when starting from the network states reached at the end of any of the other motifs, even if these states were not present during training (a case of out-of-distribution generalization). We meet the first requirement by designing recurrent neural networks (RNNs) with specific architectures that segregate motif-dependent parameters (as customary in continual learning works), and try a standard method to address the second by training with random initial states. We find that these standard RNNs are very unreliable during zero-shot transfer to motif chaining. We then use insights from the motor thalamocortical circuit, featuring a specific module that shapes motif transitions. We develop a method to constrain the RNNs to function similarly to the thalamocortical circuit during motif transitions, while preserving the large expressivity afforded by gradient-based training of non-analytically tractable RNNs. We then show that this thalamocortical inductive bias not only acts in synergy with gradient-descent RNN training to improve accuracy during in-training-distribution motif production, but also leads to zero-shot transfer to new motif chains with no performance cost. Besides proposing an efficient, robust and flexible RNN architecture, our results shed new light on the function of motor preparation in the brain.

One-sentence Summary: Motor preparation in nonlinear RNNs supports robust chaining of accurate continuous motor motifs in never-experienced orders.