Reservoir Computing Seminar Group​

Abstract: Reservoir computing has emerged as a ubiquitous paradigm for performing complex temporal computations. One particularly impressive feature of reservoir computers (RCs) is inference—the ability to accurately forecast new time-series data—thereby demonstrating an understanding of the underlying equations that generated the data. How do we make exact this notion of “understanding” in the language of dynamical systems? Can we engineer this understanding to design RCs that run custom algorithms, thereby bringing reservoir computing closer to real computers? In this talk, I will first demonstrate an analytical mechanism used by RCs to model the generative equations of data. Then, I will demonstrate how to decompose RCs to decode this model from trained RCs, and how to simply design the interactions between RC units to program our own dynamical models and algorithms into the subspace of RCs. Finally, I will discuss some current work on traversing the model manifolds of RCs—the nonlinear geometry of RC parameters that perform the same computations with different weights—to define families of functionally equivalent RCs, and RC compositionality—how to construct RCs with complex functions from those with elementary functions—to compose complex RC programs. The eventual goal is to transform RCs from computational devices into fully-fledged analog computers.