How to Make Hardware Reservoirs Robust to Noise and Variability: A Conceptor-Based Approach
The increasing demand for low-power and energy-efficient computing, particularly in edge and embedded systems, has motivated the development of brain-inspired computing paradigms such as Reservoir Computing (RC), which is especially attractive for hardware implementation. However, physical reservoirs are inherently affected by non-idealities such as internal noise, device variability, and hardware degradation, which can alter their dynamics and compromise performance over time.
In this talk, we explore the use of conceptors as a mechanism to improve the robustness and adaptability of hardware reservoir computing systems. Conceptors provide a framework to characterize and control reservoir dynamics by identifying the state-space regions associated with specific dynamical patterns. However, conceptors have not yet been adapted to physical hardware, where internal states are inevitably affected by noise and device imperfections. To address this limitation, we propose a hardware-compatible adaptation based on cross-trial correlations (CTC), designed to mitigate the effects of noise during conceptor computation. Numerical simulations under realistic conditions demonstrate that this approach enhances robustness, preserves internal dynamics, and improves performance compared to both standard conceptors and unconstrained reservoirs.
This Annual PhD student seminar will be broadcasted in the following zoom link: https://us06web.zoom.us/j/89466064429?pwd=po9p99eAEYVPaNI8xIIGoOIz0hOqaF.1
Detalles de contacto:
Miguel C. Soriano Contact form