A Non-fully-Connected Spiking Neural Network with STDP for Solving a Classification Task

Abstract

This paper presents a concept of a non-fully-connected spiking network capable of solving a classification task by means of the local bio-inspired learning rule of Spike-Timing-Dependent Plasticity. The network comprises one layer of neurons, each neuron receiving a subset of the input vector components. Input vectors are encoded by mean rates of Poisson input sequences. After training several networks each on its own class, the output spiking rates contain the information on the classes, which can be extracted with a conventional learning algorithm. We demonstrate that the STDP-based classification algorithm proposed achieves competitive accuracy on both discrete-data task of handwritten digits recognition (96% ± 1%) and classification of real-valued vectors of Fisher’s Iris (93% ± 3%). The attractive feature of the algorithm is the simplicity of the network structure without much loss in classification accuracy. This property gives the possibility to implement classifiers based on the proposed algorithm in robotic devices with limited resources.