4.0 Convolutional Neural Network
CNNs are a much better approach to solving image classification problems. CNNs are based on the visual cortex of mammals and therefore find their origins in biological research just like simple perceptrons. It is based on the realization that neurons in the visual cortex have a small local receptive field which can overlap to create a large robust visual system. This inspired an ANN architecture that later became the CNN. Famously implemented by Yann LeCun et al (See Section 2.0). The LeNet-5 architecture was first used to classify the MNIST dataset. Other famous CNN architectures include AlexNet by Alex Krizhevsjy, GoogLeNet by Szegedy, and ResNet by Kaiming He et. al. Compared to Densely Connected Neural Networks, the CNN uses lesser parameters and thus scales well over larger input data and is easier to deploy as applications. The CNN's architecture gives its strength. There are basically two main activities: (1) feature extraction via convolutions, and (2)classification. Input data (Images) are represented as tensors of dimensions: H-Height of image in pixels, W-Width of image in pixels, Color channel dimensions - 1 for Gray color and 3 for RGB color. Figure 4 below is an outlook of a CNN model.