Spiking LCA in a Neural Circuit with Dictionary Learning and Synaptic Normalization

Abstract

The Locally Competitive Algorithm (LCA) [17, 18] was put forward as a model of primary visual cortex [14, 17] and has been used extensively as a sparse coding algorithm for multivariate data. LCA has seen implementations on neuromorphic processors, including IBM’s TrueNorth processor [10], and Intel’s neuromorphic research processor, Loihi, which show that it can be very efficient with respect to the power resources it consumes [8]. When combined with dictionary learning [13], the LCA algorithm encounters synaptic instability [24], where, as a synapse’s strength grows, its activity increases, further enhancing synaptic strength, leading to a runaway condition, where synapses become saturated [3, 15]. A number of approaches have been suggested to stabilize this phenomenon [1, 2, 5, 7, 12]. Previous work demonstrated that, by extending the cost function used to generate LCA updates, synaptic normalization could be achieved, eliminating synaptic runaway [7]. It was also shown that the resulting algorithm could be implemented in a firing rate model [7]. Here, we implement a probabilistic approximation to this firing rate model as a spiking LCA algorithm that includes dictionary learning and synaptic normalization. The algorithm is based on a synfire-gated synfire chain-based information control network in concert with Hebbian synapses [16, 19]. We show that this algorithm results in correct classification on numeric data taken from the MNIST dataset.