Neural Network for Image Classification

Abstract

As observed machine learning, computer vision techniques and other computer science algorithms cannot compete the human level of intelligence in pattern recognition such as hand written digits and traffic signs. But here we have reviewed a biologically plausible deep neural network architecture which can make it possible using a fully parameterizable GPU implementation deep neural network independent of the pre-wired feature extractors designing, which are rather learned in a supervised way. In this method tiny fields of winner neurons gives sparsely connected neural layers which leads to huge network depth as found in human like species between retina and visual cortex. The winning neurons are trained on many columns of deep neurons to attain expertise on pre-processed inputs in many different ways after which their predictions are averaged. Also GPU used, enables the models to be trained faster than usual. Upon testing the proposed method over MNIST handwriting data it achieves a near-human performance. Upon considering traffic sign recognition, our architecture has an upper hand by a factor of two. We also tried to improve the state-of-theart on a huge amount of common image classification benchmarks.

Keywords: Neural network, Machine learning, Computer vision