A Neuromorphic Normalization Algorithm for Stabilizing Synaptic Weights with Application to Dictionary Learning in LCA

Instabilities in neuromorphic machine learning can occur when synaptic updates meant to encode matrix transforms are not normalized. This phenomenon is encountered in Hebbian learning [5], where, as a synapse’s strength grows, post-synaptic activity increases, further enhancing synaptic strength, leading to a runaway condition, where synaptic strength becomes saturated [3, 7]. A number of mechanisms have been suggested for regulating and stabilizing this phenomenon [1, 2, 4, 6]. Here, we present a new neuromorphic algorithm to directly normalize a synaptic connectivity. The algorithm is based on a synfire-gated synfire chain-based information control network in concert with Hebbian synapses [8, 10]. The algorithm is designed to directly normalize synaptic weights via the minimization of a cost function. We demonstrate its effectiveness as a component of a Locally Competitive Algorithm (LCA) [9] with dictionary learning, which exhibits runaway in the absence of an effective normalization procedure [15]. LA-UR-21-31884