Why maths, and why not too much
Employers who hire for machine learning roles often list linear algebra, probability and statistics, and skills frameworks for data science put mathematics near the top. The reason is practical. When a model behaves strangely, the person who understands what it is doing inside can find the problem. The person who only knows which buttons to press is stuck.
The good news: you do not need a maths degree to start. You need three ideas, understood well enough to explain to a friend. This section teaches them one at a time.
The three ideas
- Vectors and matrices. A model sees data as lists and grids of numbers. One farm's record (rainfall, fertiliser, planting week) becomes a list of three numbers called a vector. A whole table of farms becomes a grid called a matrix. Most of the work a model does is multiplying these grids together.
- Rates of change and gradient descent. A model starts with random settings and is badly wrong. It measures how wrong it is, works out which direction to nudge each setting to be less wrong, and repeats thousands of times. The maths of "which direction and how much" is the rate of change, also called the derivative or gradient.
- Probability. Many models do not answer yes or no. They say something like "this loan has a 0.8 chance of being repaid". Probability tells you what that number means, how to turn it into a decision, and why even a confident model can be wrong.