J. Blanc-Talon et al. (Eds.): ACIVS 2006, LNCS 4179, pp. 1133 – 1142, 2006. © Springer-Verlag Berlin Heidelberg 2006 Adaptive Learning Procedure for a Network of Spiking Neurons and Visual Pattern Recognition Simei Gomes Wysoski, Lubica Benuskova, and Nikola Kasabov Knowledge Engineering and Discovery Research Institute, Auckland University of Technology, 581-585 Great South Rd, Auckland, New Zealand {swysoski, lbenusko, nkasabov}@aut.ac.nz http:www.kedri.info Abstract. This paper presents a novel on-line learning procedure to be used in biologically realistic networks of integrate-and-fire neurons. The on-line adap- tation is based on synaptic plasticity and changes in the network structure. Event driven computation optimizes processing speed in order to simulate net- works with large number of neurons. The learning method is demonstrated on a visual recognition task and can be expanded to other data types. Preliminary experiments on face image data show the same performance as the optimized off-line method and promising generalization properties. 1 Introduction The human brain has been modelled in numerous ways, but these models are far from reaching comparable performance. These models are still not as general and accurate as the human brain despite that outstanding performances have been reported [1] [2] [3]. Of particular interest to this research are the models for visual pattern recognition. Visual pattern recognition models can be divided in two groups according to the con- nectionist technique applied. Most of the works deal with the visual pattern recogni- tion using neural networks comprised of linear/non-linear processing elements based on the neural rate-based code [4] [5]. Here we refer to these methods as traditional methods. In another direction, a visual pattern recognition system can be constructed through the use of brain-like neural networks. Brain-like neural networks are networks that have a closer association with what is known about the way brains process information. The definition of brain-like net- works is intrinsically associated with the computation of neuronal units that use pulses. The use of pulses brings together the definitions of time varying postsynaptic potential (PSP), firing threshold (ϑ), and spike latencies (Δ), as depicted in Figure 1 [6]. Brain-like neural networks, despite being more biologically accurate, have been considered too complex and cumbersome for modeling the proposed task. Table 1 shows a general classification of neural models according to the biological accuracy. However recent discoveries on the information processing capabilities of the brain and technical advances related to massive parallel processing, are bringing back the idea of using biologically realistic networks for pattern recognition. A recent pioneer-