nMNSD—A Spiking Neuron-Based Classifier That Combines Weight-Adjustment and Delay-Shift

Abstract

The recent “multi-neuronal spike sequence detector” (MNSD) architecture integrates the weight- and delay-adjustment methods by combining heterosynaptic plasticity with the neurocomputational feature spike latency, representing a new opportunity to understand the mechanisms underlying biological learning. Unfortunately, the range of problems to which this topology can be applied is limited because of the low cardinality of the parallel spike trains that it can process, and the lack of a visualization mechanism to understand its internal operation. We present here the nMNSD structure, which is a generalization of the MNSD to any number of inputs. The mathematical framework of the structure is introduced, together with the “trapezoid method,” that is a reduced method to analyze the recognition mechanism operated by the nMNSD in response to a specific input parallel spike train. We apply the nMNSD to a classification problem previously faced with the classical MNSD from the same authors, showing the new possibilities the nMNSD opens, with associated improvement in classification performances. Finally, we benchmark the nMNSD on the classification of static inputs (MNIST database) obtaining state-of-the-art accuracies together with advantageous aspects in terms of time- and energy-efficiency if compared to similar classification methods.

A Model of the Early Visual System Based on Parallel Spike-Sequence Detection, Showing Orientation Selectivity

“…According to [ 18 ], the LIFL neuron model has a fixed spike threshold V th , and it is able to operate in two different modes: under-threshold (or passive) mode when V m < V th , and over-threshold (active) mode when V m > V th . While in the under-threshold mode the neuron behaves as a leaky integrator, in the over-threshold mode it is characterized by the neurocomputational feature spike latency, which consists of a membrane potential-dependent time delay between the overcoming of the potential threshold and the actual spike generation (see [ 19 , 22 ]).…”

“…Among the prototypical structures developed for bio-inspired neural microcircuits, the recent multi-neuronal spike sequence detector (MNSD) [ 18 , 19 ] is able to perform online learning and recognition of parallel spike sequences (i.e., sequences of pulses belonging to different neurons/neuronal ensembles). Based on the leaky integrate and fire with latency (LIFL) neuron model [ 20 , 21 ] and neural plasticity, the MNSD combines different learning schemes (weight- and delay-based), and offers a way to understand the neural processing of information.…”

Precise Spiking Motifs in Neurobiological and Neuromorphic Data

“…In deep learning, the gradient is computed on the activation function and since spikes are not differentiable, a recent popular approach consists in using a surrogate gradient [179] to "cross-compile" a classical neural network to a spiking architecture [180]. SNNs reach in some case a similar performance as their non-spiking equivalent, for instance on the MNIST dataset for categorizing digits in a stream of events [181]. So far, this approach does not outperform classical architectures both in term of training efficiency and performances [182].…”

Precise Spiking Motifs in Neurobiological and Neuromorphic Data

“…Since the activation function of a spiking neuron is not differentiable, a recent popular approach consists in using a surrogate gradient [162] to "cross-compile" a classical Neural Network to a spiking architecture [163]. This approach is quite successful, and SNNs reach in some case a similar performance as their non spiking equivalent, for instance on the N-MNIST dataset for categorizing digits in a stream of events [164]. The main reason for adopting this approach instead of classical architectures is the possibility to benefit from dedicated neuromorphic low-energy hardware.…”

“…In other neuronal models, an efficient use or detection of these spatio-temporal patterns embedded in the spike train comes with the integration of heterogeneous delays [172,173]. The recent "multi-neuronal spike sequence detector" architecture integrates the weight-and delay-adjustment methods by combining plasticity with the modulation of spike latency emission [164]. Additional models for detection of latency patterns are presented in the extensive (graph-centric) review on synchronization in time-varying networks [174,175].…”