Intuitively Understanding Convolutions for Deep Learning

The advent of powerful and versatile deep learning frameworks in recent years has made it possible to implement convolution layers into a deep learning model an extremely simple task, often achievable in a single line of code. However, understanding convolutions, especially for the first time can often feel a bit unnerving, with terms like kernels, filters, channels and so on all stacked onto each other. Yet, convolutions as a concept are fascinatingly powerful and highly extensible, and in this post, we'll break down the mechanics of the convolution operation, step-by-step, relate it to the standard fully connected network, and explore just how they build up a strong visual hierarchy, making them powerful feature extractors for images. The 2D convolution is a fairly simple operation at heart: you start with a kernel, which is simply a small matrix of weights. This kernel "slides" over the 2D input data, performing an elementwise multiplication with the part of the input it is currently on, and then summing up the results into a single output pixel.