Centrale Méditerranée
Sitraka FORLERCentrale Méditerranée · AMSE

Machine Learning, DDEFI track

Deep learning

Deep learning, the essentials: key concepts, why they matter, the maths on the board, and a mini tutorial

Reference for the whole course, to read by track, maths or code, then the other, reference deck
Centrale Méditerranée, Château-Gombert technology park, Marseille

θ₀θ*θ ← θ − η ∇f(θ), η = 0.12, f = θ₁² + 7 θ₂²
Figure 1: gradient descent, each step is perpendicular to the level curves, hence the zigzag when the problem is ill-conditioned.

Open the session

Each session has its own password, given in class. Capitals, spaces and hyphens do not matter. The slides are encrypted: nothing can be read without it.

In this session

  1. Computation graph, gradient step and Hessian, normalisation and batch norm
  2. Depth and ResNet, convolution, RNN and LSTM, attention and transformers
  3. Autoencoders, VAE, GAN: the formulas alone, then in words, then in ten lines
  4. A mini tutorial of six exercises, maths track and code track, with answers
Introduction14 October 2026Tools14 October 2026Session 114 October 2026Session 214 October 2026Session 315 October 2026Session 415 October 2026Homework 1Milestone 1Session 515 December 2026Session 616 December 2026Session 716 December 2026Homework 2Milestone 2Session 812 February 2027Homework 3Final deliverySession 912 March 2027ThreadReferenceExampleReferenceThe codeReferenceProjectsReferenceHandbookReferenceInfraReferencePanoramaReferenceDeep learningReference
Figure 2: the thread of the course; this session is highlighted.
SessionTitleDate
IntroductionAI at a crossroads: history, troughs, jobs, and you14 October 2026Open
ToolsKaggle, GitHub, Python: your tools, your accounts, your repository14 October 2026Open
Session 1Frame before you model, then least squares14 October 2026Open
Session 2Likelihood, generalisation, project kick-off14 October 2026Open
Session 3Convex optimisation and regularisation15 October 2026Open
Session 4Decide under cost, ship a reproducible repository15 October 2026Open
Homework 1Homework after block 1: a defensible baselineMilestone 1Open
Session 5Back-to-class test, kernels, SVM and decision trees15 December 2026Open
Session 6Forests, boosting and the honest comparison16 December 2026Open
Session 7Explain, audit fairness, discover without labels16 December 2026Open
Homework 2Homework between blocks 2 and 3: a documented champion modelMilestone 2Open
Session 8Neural networks, drift, conformal prediction, going live12 February 2027Open
Homework 3Homework after block 3: the project deliveredFinal deliveryOpen
Session 9Final presentation day: the defences12 March 2027Open
ThreadThe running thread, step by step: the default-score repository, in screenshotsReferenceOpen
ExamplePredicting an asset price: time series, XGBoost and news sentimentReferenceOpen
The codeThe code of every model: import, download by API, fit, predictReferenceOpen
ProjectsOpen data and machine learning projects by sector: finance, insurance, logistics, tech, sportReferenceOpen
HandbookThe data and AI expert handbook: scope, tool up, model, deploy, observeReferenceOpen
InfraData engineering, SRE and cloud: the commands, the principles, the jobs, from junior to expertReferenceOpen
PanoramaA panorama of models: which ones, what for, in which sectorsReferenceOpen
Deep learningDeep learning, the essentials: key concepts, why they matter, the maths on the board, and a mini tutorialReference

Table 1: the sessions of the course, in order; each one opens with its own password.