Learn to build, evaluate, and distrust machine learning models properly.
Read the guide: How to learn machine learning from scratch: a roadmap
Direct Answer: Learn to build, evaluate, and distrust machine learning models properly — from linear regression to gradient boosting, with scikit-learn in your hands the whole way.
The honest definition, the two families of problems, and the split that makes any of it trustworthy.
The two workhorses you should always try first, and the overfitting problem they teach you to see.
The models that win on tabular data, and why combining weak learners beats tuning one strong one.
Why accuracy lies, which metric matches your decision, and how to tune without fooling yourself.
The engineering that actually moves scores, learning without labels, and the failures no metric will show you.
Getting a fitted pipeline out of the notebook, behind an API, and under watch — plus the skew that quietly ruins good models.
Error analysis that finds the real defect, explanations you can defend, and the tests that stop a pipeline breaking in silence.
What changes once rows arrive in an order, from the backtest that replaces cross-validation to the features you are actually allowed to compute at prediction time.
Making a working model survive real volume, from data that will not fit in RAM to the price of a thousand predictions.
Where a network genuinely beats a boosted tree, and how to borrow representations somebody else spent millions of dollars learning.
Closing the gap between a metric that improved offline and a business that measurably got better because of it.
The obligations that arrive once decisions are automated, and what it takes to run many models for years instead of one for a quarter.
Topic checks
Each one opens the next topic
Chapter quizzes
Each one opens the next chapter
50-question final
50 of 100 marks to pass
Certificate
Issued instantly, verifiable
One attempt, 60 minutes. If you don't pass, you can retake once after 30 days.
To learn machine learning from scratch, get comfortable with Python and pandas first, then learn one small loop and repeat it until it is boring: split your data, score a dumb baseline, fit a simple model, measure it honestly, and read the mistakes it makes. Linear models, then trees and boosting, then evaluation, then everything else. Neural networks come late, not first. Most people who give up did it the other way round.
That is the short answer. The rest of this post is the order we would actually follow, what each stage is for, the traps that cost beginners months, and how our free AI & Machine Learning Essentials course maps onto it chapter by chapter.
Less maths than you fear, more programming.
| Before machine learning | How much you need | Why |
|---|---|---|
| Python | Comfortable: functions, loops, lists, dictionaries, reading an error | Every tool you will use is a Python library |
| pandas | Load a CSV, filter rows, group, join, handle missing values | Most real ML time is spent shaping data |
| SQL | Basic SELECT, WHERE, GROUP BY, JOIN | Data at work lives in databases, not CSV files |
| Statistics | Mean, spread, what a distribution looks like | You will be asked whether a score difference is real |
| Linear algebra and calculus | Intuition, not proofs | You can learn the pieces when a lesson needs them |
If the first two rows are shaky, fix them before anything else. Our Python development course covers the language, and SQL for data analysis covers the querying. A week here saves a month later.
On hardware: classic machine learning on tabular data runs fine on an ordinary laptop. If yours is old, a free hosted notebook works too.
Each stage below builds on the one before it. Skipping ahead is the most common reason people feel lost.
Machine learning means writing a program that finds the rules from examples instead of you writing the rules. Learn the vocabulary once (feature, label, sample, training, inference), the difference between supervised and unsupervised problems, and above all the train, validation and test split. The test set is the only honest exam your model gets. If you look at it while building, you have spent it.
Then learn baselines. Before any real model, score something stupid: predict the most common class, or the average value. Every model you build afterwards has to beat that number, and you will be surprised how often a first attempt does not.
Linear regression for numbers, logistic regression for yes or no. They are fast, explainable, and often good enough. This is also where you meet overfitting, the bias and variance trade off, and regularisation, which are ideas you will use for the rest of your career.
On spreadsheet style data, tree ensembles are usually the strongest tool you have. Learn a single decision tree, then random forests, then gradient boosting. scikit-learn ships its own fast boosting models, and XGBoost, LightGBM and CatBoost are the specialist versions you will see in job posts.
This is the stage most tutorials rush, and it is the one that separates people who can be trusted with a model from people who cannot. Accuracy lies whenever one class is rare. Learn the confusion matrix, precision, recall, F1, ROC and precision recall curves, cross validation, hyperparameter search, and how to put error bars on a score so you can say whether model B is actually better than model A.
Feature engineering usually moves your score more than switching algorithms does. Wrap every preprocessing step in a scikit-learn pipeline so the same transformations run in training and in production. Learn data leakage here: information from the future or from the answer sneaking into your features, which produces a model that looks brilliant until it meets real data.
Then learn to save a model, serve it behind an API, and watch it for drift, followed by error analysis, time series, scaling, neural networks, and proving a model helped a business at all. This is what turns a notebook into something an employer can use.
It depends on your starting point and your hours, so treat this as a shape rather than a promise.
| Your starting point | Realistic first goal | What it takes |
|---|---|---|
| No programming yet | Python and pandas first | Plan for a few months before ML itself |
| Comfortable with Python | Build and evaluate models on tabular data | Roughly two months at a steady part time pace |
| Already doing data analysis | Add ML to your existing work | Weeks, since the data skills transfer |
Our course is listed at eight weeks, with around eleven hours of reading across 48 topics. Reading is not the slow part. Running the code, breaking it and fixing it is, and that is where the learning happens.
Starting with deep learning. On tabular data a boosted tree usually beats a network, trains in seconds, and teaches you more about evaluation. Our course gets to networks in chapter 10 on purpose.
Collecting courses instead of building. Three half finished courses teach less than one finished course plus one project you can explain. Pick one path and finish it.
Trusting accuracy. If 2 in every 100 transactions are fraud, a model that says "never fraud" is 98 percent accurate and completely useless. Choose the metric that matches the decision someone will make with the prediction.
Touching the test set early. Each peek makes it a less fair exam. Use validation data or cross validation for decisions and look at the test set once.
Leakage you do not notice. A feature that is only known after the event you are predicting will make your score look wonderful. If a result seems too good, assume leakage and go hunting.
The course uses scikit-learn throughout and is written as text with code you run yourself, not as videos. It is aimed at people who already write some Python and want to go from beginner to a solid intermediate level. Here is what each of its 12 chapters covers.
| Chapter | What you learn |
|---|---|
| 1. What Machine Learning Actually Is | Learning from data, supervised versus unsupervised, the train, validation and test split, baselines |
| 2. Linear Models | Linear and logistic regression, choosing a threshold, overfitting and regularisation, gradient descent |
| 3. Trees, Forests, and Boosting | Decision trees, random forests and bagging, gradient boosting, stacking |
| 4. Measuring a Model Honestly | Confusion matrix, precision, recall, F1, ROC-AUC and PR curves, cross validation, error bars |
| 5. Features, Clusters, and the Limits of Models | Feature engineering and pipelines, clustering, bias and leakage, PCA |
| 6. Shipping It | Saving a model safely, a prediction service, monitoring and drift, packaging |
| 7. Debugging Models and Testing ML Code | Error analysis, explanations you can defend, testing ML code, triage when a model will not learn |
| 8. Time Series | Walk forward backtesting, lag and rolling features, forecast horizons, regime change |
| 9. Scaling It | Data that does not fit in memory, training time, latency budgets, the cost of a prediction |
| 10. Neural Networks, Embeddings, and Pretrained Models | Multilayer networks, training without wasting time, embeddings, transfer learning and when to call an API |
| 11. Proving It Helped | Why offline gains do not always show up, A/B testing a model, uplift, impact without randomisation |
| 12. Governance, Lineage, and a Portfolio of Models | Model cards, reproducing old decisions, regulated decisions, running many models over years |
Chapters 1 to 5 are the core roadmap above. Chapters 6 to 12 cover the work after a model exists, which most free material leaves out.
Every chapter has four topics. Each topic has its own short quiz, and you need 70 percent on it to unlock the next topic. Each topic also has a minimum reading time of 90 seconds before its quiz opens, so you cannot click straight through. Each chapter then closes with a chapter quiz that pulls the four topics together.
The certificate comes from the final exam: 50 questions, 60 minutes, one timed attempt, and you pass with 50 out of 100 marks. If you fail, a retake opens 30 days later with the questions and options reshuffled. Pass, and you get a free certificate with a verification code and a QR code that anyone can check. Be realistic about what that paper is worth; we wrote about it honestly in are free online certificates worth anything.
Pick one public dataset with a clear thing to predict. Split it, score a baseline, fit a logistic or linear regression, and write down both scores. Then start AI & Machine Learning Essentials at one topic a day. If you later want to build products on large language models instead, AI engineering is the natural next course, and our AI engineer roadmap lays out that path.
Start with chapter 1 of AI & Machine Learning Essentials: learning from data, the split that makes a score trustworthy, and the baseline every model has to beat. It is free, and you can finish the first topic today.
05 questions
Yes. What you need is working Python, comfort with data in pandas, and patience with evaluation. A degree helps with some employers, but the skills are learnable on your own.
Less than you fear to start. Basic statistics and an intuition for what a slope and an average mean will carry you through the first several stages. Pick up the linear algebra and calculus when a lesson needs them.
Machine learning first. Linear models, trees, boosting and honest evaluation are the foundation deep learning sits on, and on spreadsheet style business data they often perform better anyway. Neural networks make much more sense once you can already measure a model properly.
Yes. The lessons, quizzes, final exam and certificate are all free. You need an account to track progress and sit the exam.
If you already write Python, around two months of steady part time study gets you building and evaluating models on tabular data. If you are starting without programming, add a few months for Python and pandas first.
