Impact of Convolutional Neural Network Input Parameters on Classification Performance
Fourth International Conference for Convergence in Technology, I2CT 2018
Deep Convolutional Neural Networks have shown impressive capabilities for solving complex image classification problems. There are numerous input parameters that decide the architecture of the network such as the number of convolutional layers, convolution kernel size, number of convolution filters in one layer, type of activation function, pooling window size, stride etc. This paper is an attempt to understand the impact of some of the input parameters on the classification performance of the network. The work is performed for a five class problem using a widely used color fundus retinal image dataset to classify stages of diabetic retinopathy. CNN input parameters such as the number of convolutional layers, number of filters in one layer, size of the convolution kernel and activation function is considered. The impact of these input parameters on the accuracy of classification and the runtime for training the network is analyzed. It has been observed that both the classification accuracy and the runtime for training the network is more heavily dependent on the number of convolution filters in one layer and size of the convolution kernels than on the number of convolutional layers or the depth of the network. It is also found the type of activation function is actually having no impact on the accuracy. This preliminary work helps to understand the functioning of CNN, identify the crucial parameters which will finally lead to explanation of the reason behind their impact on the performance on one hand, and possibly the justification to correlate these to real biological neuronal networks, on the other.
Maitra, Sanjit; Ojha, Rahul Kumar; and Ghosh, Kuntal, "Impact of Convolutional Neural Network Input Parameters on Classification Performance" (2018). Conference Articles. 51.