A First Year in Machine Learning

Most people learning machine learning do not lack material. They lack an order. Without one, a year of effort produces a large collection of half-understood techniques and very little judgement.

This is one workable order. It is not the only one.

Months one to three — the ground#

Learn to program properly before learning to model. Then the three pieces of mathematics that everything rests on:

• Linear algebra — vectors, matrices, and what a matrix multiplication does geometrically
• Probability — distributions, conditional probability, expectation
• Calculus — derivatives and the chain rule, which is all that backpropagation is

Do not try to complete these subjects. Learn enough to read a paper's notation without stopping, and return when something specific is unclear.

Months four to six — classical models#

Before neural networks, work through linear regression, logistic regression, decision trees and ensembles. They are interpretable, they train in seconds, and they teach the habits that matter far more than architecture: splitting data honestly, recognising overfitting, and choosing a metric that reflects what you actually care about.

Build three small projects end to end. Finishing something mediocre teaches more than abandoning something ambitious.

Months seven to nine — neural networks#

Now the depth. Feed-forward networks, then convolutional networks, then attention and transformers. Implement a small one from scratch once — not because you will ever do it again, but because it converts the diagram into something you actually understand.

Months ten to twelve — a real problem#

Pick a problem you personally care about, with messy data you have to clean yourself. This is where the gap between tutorial competence and real competence closes, because nothing in a real problem arrives pre-formatted.

The goal of the first year is not to know every method. It is to be able to read something new and place it correctly against what you already know.