ABSTRACT
In the mid-2010s, a new type of computation appeared: graphics processing unit (GPU) computation. This was a genuine revolution in the world of machine learning. A computer is made of numerous electronic components: the processor that does the computation, the memory that stores the data, but also the keyboard, screen, mouse, hard disk, and so on. All these components work with binary numbers computation. Graphics cards are an essential component in video game consoles and, just as often, computers. Each new generation offers better image quality for a better gaming experience: it is the graphics cards that make this possible. The performance of these components has continuously increased since the 1980s, driven by the formidable economic clout of the video game market. A graphics card is made up of hundreds of small mini-processors: “graphics processing units” or GPUs. In a deep network, the “right” values for each layer’s parameters depend on the preceding layer’s work.
