Machine learning fundamentals

What the algorithm is given to learn from: labelled examples, unlabelled data, or an environment to act in. Answer that and you have named the type.

Supervised learns from labelled examples to predict the label of something new. Unsupervised is given no labels and finds groups or structure the data already contains — but cannot say what those groups mean.

An agent acts in an environment, receives a reward or penalty for what happens, and over thousands of attempts learns which actions earn the most reward. It needs no dataset, but must be allowed to fail repeatedly.

It uses many layers, each building on what the last found — edges, then shapes, then whole objects — so it works out its own features. The cost is needing a great deal of data and being hard to explain.

Starting from a model already trained on a related task and retraining only its final layers on your own data. It is used when you have too few examples to train from scratch — hundreds rather than millions.

Training repeats the same calculation over millions of examples for hours or days, needing massive parallel processing. Using the trained model is one small calculation per question, which a phone or camera can often do.

Both provide the massive parallel processing training needs. A GPU has thousands of general simple cores; a TPU is a chip designed specifically for machine-learning calculations.

An ASIC is built for one task and can never do another — the fastest and most power-efficient option, but useless if the task changes. An FPGA can be reconfigured after manufacture, so it is slower but adaptable.

When a trained model must answer immediately on the spot — in a camera or a vehicle. It gives millisecond responses, keeps data off the network, and works even when the connection drops.

Training datasets run to terabytes and must be read fast enough to keep the processors busy. Slow storage leaves very expensive hardware sitting idle and waiting.