P-CRITICAL: a reservoir autoregulation plasticity rule for neuromorphic hardware

Abstract

Backpropagation algorithms on recurrent artificial neural networks require an unfolding of accumulated states over time. These states must be kept in memory for an undefined period of time which is task-dependent and costly for edge devices. This paper uses the reservoir computing paradigm where an untrained recurrent pool of neurons is used as a preprocessor for temporally structured inputs and with a limited number of training data samples. These so-called reservoirs usually require either extensive fine-tuning or neuroplasticity. We propose a new local and unsupervised plasticity rule named P-CRITICAL designed for automatic reservoir tuning that translates well to physical and digital neuromorphic processors. The spiking neuronal architecture implementation is simulated on the Loihi research chip from Intel and on a conventional CPU. Comparisons on state-of-the-art machine learning datasets are given. Improved performance on visual and auditory tasks are observed. There is no need to a priori tune the reservoir when switching between tasks, making this approach suitable for physical implementations. Furthermore, such plastic behaviour of the reservoir is a key to end-to-end energy-efficient neuromorphic-based machine learning on edge devices.

Excitatory/inhibitory balance emerges as a key factor for RBN performance, overriding attractor dynamics

“…However, in the case of discontinuous activation functions, such as the one we have with the RBN, it has been shown that the Echo State Property (ESP) cannot be achieved. The spectral radius alone fails to characterize the dynamics and performance of these reservoirs (Oztuik et al, 2007 ; Alexandre et al, 2009 ; Tieck et al, 2018 ; Balafrej et al, 2022 ). In Supplementary material S8.1 , we explicitly discuss the link between ρ and the mean and variance of the weight matrix and show ρ is of no particular interest in the study of the dynamics.…”

Extended liquid state machines for speech recognition

“…The resulting E/I balance is related to an optimized dynamical network chaos (Van Vreeswijk and Sompolinsky, 1996 ), called the edge of chaos. In previous research (Ivanov and Michmizos, 2021 ; Balafrej et al, 2022 ), edge-of-chaos dynamics were shown to improve the neural coding in an LSM in comparison with a randomly initiated liquids. The balance of excitatory and inhibitory currents can thus be seen as an often neglected heuristic for optimized coding in LSMs.…”

“…Recently, several approaches to improving the performance of LSMs have been proposed. Task-agnostic, data-driven training of the recurrent liquid weights (Jin and Li, 2016 ; Ivanov and Michmizos, 2021 ), continuous neuronal adaptation based on intrinsic neuronal plasticity (Zhang and Li, 2019a ), reservoir autoregulation (Balafrej et al, 2022 ), liquid ensembles (Wijesinghe et al, 2019 ), and evolutionary optimization (Zhou et al, 2020 ) have all been shown to improve the basic LSM design, keeping its sparse properties. However, these enhancements come at the cost of increased (training) complexity and data-dependent tuning of the LSM parameters, eliminating some of its inherent advantages.…”