Application of Hybrid of ACO-BP in Convolution Neural Network for Effective Classification

Abstract

Convolution Neural Network (CNN) has been widely used in pattern recognition for various applications. Convolution neural network performs non-linear transformation on input to generate the global abstract feature vector. The resulting global feature vector is input to Fully connected Neural Network (FNN) and the activation value at the neuron in the output layer classifies the input data vector. During training of CNN on a given dataset, error at the output layer is minimized using backpropagation with stochastic gradient descent. The weights optimization using backpropagation has a drawback of local minima. Thus, in this research paper hybrid of ACO-BP has been used for initialization of CNN weights using Ant Colony Optimization (ACO) and its further optimization using Backpropagation (BP) to overcome local minima. The performance of CNN shows the improvement since the ability of deep learning architecture to generalize depends on the weight configuration during training phase. Experiment was conducted on MINST data set using k-fold cross validation method to confirm the effectiveness of CNN with hybrid of ACO-BP in pattern recognition. The results show the improvement in the classification accuracy-using hybrid of ACO-BP with CNN in comparison to CNN with BP only.