Figure 5.

The figure shows a simple CNN meant for classifying images (in this case, images of digits). Most CNN architectures include (1) convolutional layers, (2) pooling layers, and (3) dense (fully connected) layers. A convolutional layer typically has multiple filters (similar to the image filters), wherein the filter weights are allowed to change and learn from the data. Each of these filters is moved across the length and breadth of the entire image as it is convolved with the image pixel values. It should be noted that these filters act like feature extractors, and the output (feature maps) obtained after performing the convolution operation is used as input to the next layer. The pooling layer provides an approach to down sample the feature maps while summarizing the presence of features, either locally or globally. Also, the pooling layer acts like a feature detector that helps identify important features and to a certain degree helps in providing rotational and translational invariance. The dense layer is a fully connected network wherein each neuron receives input from each neuron of the previous layer. Typically, the dense layer contributes to the greatest number of learnable parameters (weights and biases) and helps reduce the training error. The sharing of filters in convolutional layers helps the CNN to avoid overfitting. The network as a whole thus attempts to achieve low training error and high generalization ability. CNN indicates convolutional neural network.