Online Meta-neuron based Learning Algorithm for a spiking neural classifier

Abstract

This paper presents a new spiking neural network architecture with a meta-neuron which envelopes all the pre- and postsynaptic neurons in the network. The concept of the meta-neuron is inspired by the role of astrocytes in modulating synaptic plasticity in biological neural networks. The meta-neuron utilizes the global information stored in the network (synaptic weights) and the local information present in the input spike pattern to determine a weight sensitivity modulation factor for a given synapse. Based on the weight sensitivity modulation factor and the postsynaptic potential of a neuron, the meta-neuron based learning rule updates the synaptic weights in the network to produce precise shifts in the spike times of the postsynaptic neurons. Using this learning rule, an Online Meta-neuron based Learning Algorithm (OMLA) is presented for an evolving spiking neural classifier. The learning algorithm employs heuristic learning strategies for learning each input spike pattern. It can choose to add a neuron, update the network parameters or delete a spike pattern depending on the spike times of the output neurons. OMLA employs a meta-neuron with memory that stores only those spike patterns which are used to add a neuron to the network. These spike patterns (spike patterns in meta-neuron memory) are used as representative of past information stored in the network during subsequent neuron additions. The performance of OMLA has been compared with both the existing online learning and batch learning algorithms for spiking neural networks using the UCI machine learning benchmark data sets. The statistical comparison clearly indicates that the OMLA performs better than other existing online learning algorithms for spiking neural networks. Since, OMLA uses both, the global as well as the local information in the network, it is also able to perform better than other batch learning algorithms.