Introduction to Deep Learning

For forty years, getting a machine to recognize a handwritten digit meant describing the digit by hand. You wrote code to count closed loops, measure the height-to-width ratio, and detect stroke intersections. These measurements were called features, and they were the real work: the classifier behind them, whether a logistic regression or an SVM, was a mere formality. Changing problems meant starting over. An engineer who knew how to describe digits knew nothing about how to describe faces.

Deep learning reverses this relationship. You no longer describe the features. You describe an architecture capable of inventing them, and let it discover them from examples. The first layers of a network trained on images end up detecting edges and color blobs; the next ones detect textures; the last ones detect parts of objects. Nobody wrote "detect an edge." It fell out of the optimization.

This shift has a price, and this course will not hide it. A deep network needs lots of examples and lots of compute, and it fails in baffling ways: it learns nothing even though the code is correct, it memorizes instead of generalizing, or it slides into a region where gradients vanish and freeze. The difference between someone who stacks layers and someone who knows how to build a model comes down entirely to the ability to diagnose these failures.