Pattern Recognition using Adaptive Convolutional Neural Network

Deep learning is an appreciate framework to deal with high-level data in order to find meaningful relationships between the features of the data. One of the well-known architecture of deep learning scheme is Convolutional Neural Network (CNN) which includes one or more convolutional layers with fully connected layers. Despite of the abilities of CNN, it suffers from parameter setting where the parameters with different values have high impacts on the performance of CNN. In this paper, parameter adaptation of CNN using new metaheuristic algorithm is proposed. The used metaheuristic algorithm is Harmony Search (HS) which is improved in this paper. Finally, the proposed method which is name Adaptive Convolutional Neural Network (ACNN) is applied on handwritten digit recognition field. Experimental results on MNIST dataset prove the superiority of the proposed ACNN to CNN.

کلیدواژه ها:

deep learning ، CNN ، pattern recognition ACNN ، convolutional neural network