Task structure shapes underlying dynamical systems

that implement computation

The talk will be in two parts. 1) First, I will present published work on multitasking artificial recurrent neural networks that revealed "dynamical motifs". Dynamical motifs are recurring patterns of network activity that implement specific computations through dynamics, such as attractors, decision boundaries and rotations. Motifs are reused across tasks and reflect the modular subtask structure of commonly studied cognitive tasks. We believe this compositional structure to be a feature of most complex cognitive tasks. This work establishes dynamical motifs as a fundamental unit of compositional computation, intermediate between neuron and network. 2) Building on these insights, we are now investigating whether shared dynamical motifs can explain the effectiveness of curriculum learning in animal behavior. Also known as animal shaping, curriculum learning is a training approach where complex tasks are learned in stages through a series of subtasks. Our work provides a novel framework for analysis of neural and behavioral data and has the potential to guide the design of optimal training protocols for both artificial systems and for animals.