How to learn machine learning from scratch

What you need before your first model#

Python at a confident level: functions, collections, working with files, plus the NumPy (arrays and vectorised operations) and pandas (tables) libraries. Without them, all your time goes into fighting the code instead of understanding the models. If you are not there yet, start with how to learn Python from scratch.

Math — enough to understand, not enough to prove:

• linear algebra — vectors, matrices, matrix multiplication;
• calculus — derivatives, the gradient, the idea of optimisation;
• probability and statistics — distributions, mean and variance, conditional probability.

You do not have to cover all of it in advance: it is convenient to study a math topic at the moment it first shows up in an algorithm.

Your working environment#

Jupyter notebooks locally or in a free cloud notebook service such as Google Colab. Libraries: numpy, pandas, matplotlib, scikit-learn. A separate virtual environment for your studies saves you from version conflicts.

Classic machine learning (2–3 months)#

Problem types: regression (predict a number — a price, demand), classification (pick a class — spam or not), clustering (split objects into groups without ready answers). Algorithms in order: linear and logistic regression, k-nearest neighbours, decision trees, random forest, gradient boosting.

The main idea to grasp right away is evaluating a model on data it has not seen during training:

from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split
from sklearn.ensemble import RandomForestClassifier
from sklearn.metrics import accuracy_score

X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(
    X, y, test_size=0.25, random_state=42)

model = RandomForestClassifier(random_state=42).fit(X_train, y_train)
print(accuracy_score(y_test, model.predict(X_test)))

Next come cross-validation, hyperparameter tuning, overfitting and regularisation, and feature preparation: missing values, categorical features, scaling.

Metrics: how not to fool yourself#

Accuracy is misleading on imbalanced data: if 95% of emails are not spam, a model that says "nothing is spam" scores 0.95 and is useless. Learn precision, recall, F1 and ROC AUC for classification, and MAE and RMSE for regression. Always compare your model with the simplest baseline — for example, predicting the most frequent class or the mean value.

Neural networks (the next 2–3 months)#

After the classics — fully connected networks, backpropagation, activation functions, optimisers. Then convolutional networks for images and the transformer architecture for text. Framework: PyTorch or TensorFlow/Keras. Start by training a small network to recognise handwritten digits and getting it not to overfit.

Projects and self-checks#

Your portfolio is two or three projects with the full cycle: problem statement, exploratory data analysis, a baseline, several models, a comparison by metrics, conclusions. Ideas: predicting house prices from public listings, classifying reviews as positive or negative, predicting customer churn on an open dataset. Keep the code in Git and fix random_state so results can be reproduced.

How to check yourself:

• take part in practice competitions on platforms like Kaggle — scoring on hidden data shows your model's real quality;
• reproduce a result from a textbook or a paper from scratch without peeking at the code;
• implement linear regression and gradient descent in NumPy yourself — after that, library models stop being a black box;
• explain why the model got specific examples wrong.

Step-by-step plan

1. Months 1–2 — Python and mathNumPy, pandas, linear algebra basics, derivatives, probability.
2. Months 3–4 — classic algorithmsRegression, classification, trees, ensembles in scikit-learn; train/test splits.
3. Month 5 — model qualityMetrics, cross-validation, overfitting, feature work, a baseline.
4. Months 6–8 — neural networksPyTorch or Keras, fully connected and convolutional networks, transformer basics.
5. PortfolioTwo or three full-cycle projects and practice competitions.

Check yourself

Sources

• scikit-learnDocumentation and user guide for classic algorithms
  free
• Kaggle LearnFree short courses on pandas and machine learning
  free
• PyTorchA neural network framework with official tutorials
  free
• Mathematics for Machine LearningA free book covering the math behind ML
  free

Was this helpful?

Yes No