Practical Time-Series Analysis

2D convolutions

Let's start by describing the 2D CNNs and we will derive 1D CNNs as a special case. CNNs take advantage of the 2D structure of images. Images have a rectangular dimension of w, where n is the height and h x w x n is the width of the image. The color value of every pixel would be an input feature to the model. Using a fully-connected dense layer having 28 x 28 neurons, the number of trainable weights would be 28 x 28 x 100 = 78400. For images of handwritten digits 32 x 32 from the MNIST dataset, the number of trainable weights in the first dense layer with 100 neurons would be c = 3. The CIFAR-10 dataset is popularly used to train object recognition models. Colored images in this dataset are 32 x 32 x 3 x 100 = 307200 and have ...