Skill · Machine learning · Frame, equip, train and reproduce machine learning models
Train a small neural network on CPU with PyTorch and read its decision boundary
Explain the layers, non-linear activation, loss and backpropagation of a neural network, install PyTorch for CPU in an isolated environment, define a network with torch.nn, write a complete training loop (gradient reset, loss, backpropagation, weight update with Adam), choose CPU or GPU with .to(device), then plot the learned decision boundary on two classes that no straight line separates.
Start lesson 1The Machine learning domain
- 2 lessons
- ~50 min of reading
- 2 quizzes
- 0 proof lab
The path
Where it comes from, where it leads
What you need to know first, where you stand on this skill, and what it opens next. Your status is computed in your browser, from what you have read and passed.
Before · 1 prerequisite
Recommended before · 1
This skill, for you
Train a small neural network on CPU with PyTorch and read its decision boundary
Computing
It unlocks · nothing yet
No other skill requires it: it is an end-of-path skill.
The path through the foundations From the most upstream foundation to this skill: what is in place, what remains to read.
Lessons
How to learn it
2 lessons, about ~50 min, in program order. Each lesson ends with its quiz.
The program
2 lessons · ~50 min
Beyond the lessons
Assess, practise, pass the proof
Reading a lesson confirms nothing. A skill is confirmed by a passed quiz, practised in a guided lab, and put to the test by a challenge or a mission; an imported lab result is not verified by the site.
Assess yourself
2 quizzes
At the end of the lessons: passed at 70%, they confirm the skill.
Practice
No guided lab
No guided lab practices it yet.
Proof
No proof
Neither a challenge nor a mission with a proof yet.
Placement test
Preparing the questions…
Your answers in detail
For each question asked: your answer, the correct answer, its explanation and the page to reread.
Around it
Who requires it, who teaches it
Required by
No published job requires it yet.
Taught by
Same block · Frame, equip, train and reproduce machine learning models
- Frame a machine learning problem: choose the learning family and place the work in the project lifecycle
- Set up a reproducible data science workstation: a pinned conda environment and a Jupyter notebook that runs end to end
- Train a linear regression and evaluate it on a test set (scikit-learn, R², RMSE)
- Group unlabelled data with k-means and choose the number of clusters (elbow, silhouette)
- Version a dataset outside Git and get back an exact version of it (DVC)
- Make a training run reproducible and describe it as a parameterised pipeline replayed identically (uv, DVC)
- Track every training run on a tracking server and find the run that produced a model (MLflow)
- Register and version a model and refer to it by alias with its lineage (MLflow)
- Test data and model before promotion: schema, ranges, threshold and non-regression against the champion (pytest)