The challenge of (machine) learning

The challenge of (machine) learning

I'm often asked to define machine learning, and to explain how it works. Here's a simple way to think about it.

The goal of machine learning is to make predictions based on the lessons of past experience. There are 3 components - model, training and evaluation.

A model is the result of training.  Over time a race car driver develops a “model” that finds the fastest line through a corner, given an infinite set of starting conditions as they enter the turn. The more they practice and the more diversified the conditions, the better the model. At peak performance, the model will find the best line nearly every time. Imagine taking that acquired skill - the model the driver has in her head - and storing it in a file. That's a machine learning model.

Training is the process of developing the model. Imagine we could snap our fingers and put our driver instantly at the start of the turn, wait for them to navigate it, and then snap our fingers to start again. For every iteration, our driver needs context - speed, condition of the car, debris on the track, worn tires, etc - all factors necessary to navigate the turn. If we could run this thousands of times, each time with a slightly different context, their model will continue to improve as they learn how things work out.

Evaluation is when the model is applied to new data. It’s race day, a sunny Sunday afternoon, and our trained driver enters the turn. The effort to predict the best line is low… just apply the model. It’s instant and effortless.