Where to start with machine learning: a four-step order
If you don't know where to start with machine learning, follow four steps in order: Python, data tools, the math and statistics a model uses, then classical models evaluated honestly. Each step ends in a checkpoint.
Start machine learning with Python, not with a model. Learn in four steps: (1) core Python, (2) the data tools NumPy, pandas, Matplotlib and scikit-learn, (3) the math and statistics that explain what a model does, and (4) classical machine learning, beginning with linear regression and gradient descent and moving to classification and evaluation. Steps 3 and 4 overlap: learn each piece of math right before the model that needs it. Move to the next step when you can pass its checkpoint without copying code, not when you finish a set number of videos.
This post is only about the order of learning machine learning itself. If you are still deciding between AI fields, including generative AI applications, read how to learn AI from scratch. If you already know you want the job, and want to know what comes after the learning (deployment, a portfolio, applying), read the machine learning engineer roadmap.
Why does the order matter?
Most people who stall in machine learning start in the wrong place. Some start with a deep learning tutorial and copy code they cannot change. Others spend months on linear algebra textbooks and never train a model. Both problems have the same cause: each step depends on the one before it, and skipping a step leaves you unable to tell whether your results are right.
| Step | What you learn | Checkpoint |
|---|---|---|
| 1. Python | The language, files, errors, tests | Count the ten most common words in a text file, with a test for the counting function |
| 2. Data tools | NumPy, pandas, Matplotlib, the scikit-learn API | Load a table, average one column per group, plot it, and fit a model against a baseline |
| 3. Math and statistics | Vectors, matrices, derivatives, distributions, probability | Read a model's formula aloud and explain what each symbol means |
| 4. Classical ML | Regression, gradient descent, classification, evaluation | Implement linear regression from scratch and match scikit-learn's coefficients |
The checkpoints come from the SchoolWhool learning path, which orders every lesson across the five course tracks into stages, each ending with a task to complete.
Step 1: Learn Python, the way ML code uses it
Learn variables and types, if and loops, lists and dictionaries, strings, functions, modules, reading files, handling errors, and writing a simple test. Learn classes well enough to read model.fit(X, y).
You can skip this step if you can already write a function that loops over a list and returns a result. Otherwise, the free Python track starts from zero; its first lesson, Why Python?, explains why machine learning runs on Python and where code is written and run. For a detailed list of what to learn and what to skip, see Python for machine learning.
Checkpoint: read a text file, count how often each word appears using a dictionary, and print the ten most common words. Write a pytest test for the counting function.
Step 2: Learn the four data tools
Machine learning is mostly data handling. Before a model sees anything, the data has to be loaded, cleaned, reshaped and looked at. The Data Tools track covers the four libraries in order:
- NumPy for arrays: shapes, indexing, boolean masks and broadcasting.
- pandas for tables: selecting, filtering, cleaning missing values and groupby.
- Matplotlib for charts, because you should look at data before modeling it.
- scikit-learn for models: every model shares the same
fit,predictandscoreinterface, as Meet scikit-learn shows by scoring a first model against a do-nothing baseline.
This is also the right time to run a complete model end to end, even before you understand its insides. Machine learning code in Python walks through a full 31-line example. Seeing the whole workflow early gives the math in step 3 something to attach to.
Checkpoint: load a CSV with pandas, compute the average of one column per group, plot it, and fit a scikit-learn model that beats a baseline on a held-out test set.
Step 3: Learn the math and statistics a model uses
You need a specific, limited set of ideas, set out in detail in prerequisites for machine learning:
- Math: functions and logarithms, reading ML notation, vectors and the dot product, matrices and the matrix–vector product, derivatives and the gradient. The Maths track covers these, each shown on a picture first. Its lesson on derivatives has a free video; more lessons open as they are recorded.
- Statistics: describing a distribution, spread, correlation and confounding, who is (and is not) in a dataset, probability and Bayes' theorem. The Statistics track covers these, starting with why statistics matters for ML.
Do not try to finish step 3 before starting step 4. Learn enough to read a formula like ŷᵢ = Σⱼ wⱼ xᵢⱼ in words, then start models and come back for each new idea as a model needs it: derivatives right before gradient descent, Bayes' theorem right before Naive Bayes, log loss right before logistic regression.
Checkpoint: explain, in plain words, what a dot product computes, what the sign of a derivative tells you, and why the median can describe a skewed feature better than the mean.
Step 4: Learn classical machine learning, starting with regression
Begin with the model that teaches the most per hour: linear regression trained by gradient descent. It brings together everything from steps 1 to 3: a prediction is a dot product, the error is a function of the weights, and the derivative tells you which way to move them. The Machine Learning course starts its regression module with three free video lessons:
- Linear regression, explained geometrically.
- Gradient descent, the method that trains most models, including neural networks.
- Gradient descent for linear regression, step by step.
After regression, the order that works is: a first classifier (k-nearest neighbors), then how to measure a model honestly (overfitting, the test set, cross-validation, precision and recall), then linear classifiers such as logistic regression, then preparing real data, then trees and ensembles. The course outline lists every lesson in that order; lessons that are still being recorded are labeled.
Checkpoint: implement linear regression with gradient descent from scratch in NumPy, and match the coefficients scikit-learn's LinearRegression finds on the same data.
What can I leave until later?
Deep learning, large language models, reinforcement learning, and deployment tools such as Docker can all wait until you can train, evaluate and explain a classical model. They build on the same foundations: a neural network is trained by the same gradient descent, and evaluating one uses the same test-set discipline. Starting there first usually means copying code you cannot debug.
Also leave out, at first: comparing ten algorithms on one dataset, tuning hyperparameters before you have a baseline, and chasing benchmark scores. One model you understand fully is worth more than ten you called.
How long does it take to learn machine learning?
It depends on your starting point and the hours you put in, so any fixed number would be a guess. A better measure is the checkpoints: if you can pass step 4's checkpoint without looking up each line, you have the foundation of machine learning. Keep a short log of what you built at each checkpoint and where you got stuck; it shows your progress more honestly than counting hours or videos watched.
Can I learn machine learning without a degree?
Yes, you can learn it. You can learn every topic in the four steps from free lessons, textbooks and library documentation. Whether a particular employer requires a degree is a separate question that depends on the role; the machine learning engineer roadmap covers how to show what you can do through projects.
Your next step
Open the learning path and start at the first stage you cannot yet pass. If that is step 1, begin with the Python track. Lesson parts and key terms are free to read, and so are the free videos. A free account adds the quizzes and saves your progress across all five tracks, with no payment.