A noise tolerant auto resonance network for image recognition

Abstract

Auto Resonance Network (ARN) is a biologically inspired generic non-linear data classifier with controllable noise tolerance useful in control applications like robotic path planning. It appears that multi-layer ARN can be useful in image recognition but such capabilities have not been studied. This paper presents a multi-layer ARN implementation for an image classification system for handwritten characters from MNIST data set. ARN nodes learn by tuning their coverage in response to statistical properties of the training data. This paper describes the effect of adjusting the resonance parameters like coverage and threshold on accuracy of recognition. ARN systems can be trained with a small training set. It has been possible to attain accuracy up to 93% with as few as 50 training samples. ARN nodes can spawn secondary nodes by perturbing the input and output similar to a mutated cell in a biological system. Use of input transformations to achieve required perturbation of the network is discussed in the paper. Results show that the network is able to learn quickly without any stability issues. The present work shows that reasonable level of image classification can be obtained using small sample training sets using resonance networks.