ML 1

Machine learning 1

Description: This course sets out the general framework of machine learning, allowing you to situate the different approaches in the field. It covers the notions of data pre-processing, an introduction to statistical learning theory (risks, overlearning, convex proxies, regularization), the difference between frequentist and Bayesian approaches, supervised, unsupervised, semi-supervised and reinforcement learning paradigms. Some approaches are detailed (Kernel methods, SVM, Boosting, Bagging, Decision trees…).

Learning outcomes: At the end of this course, students will be able to recognize the different classes of algorithms in the landscape of the many methods available on the shelf. They will also have the statistical notions that will enable them to make reasoned use of these methods, thus avoiding a black-box approach with blind parameter testing.

Means: The courses and practical work are given by Hervé Frezza-Buet, Arthur Hoarau, Jérémy Fix. The courses present theoretical aspects, mathematical proofs, but are also illustrated by demonstrations of algorithms. The practical work will be done in Python, using sickit-learn, in pairs.

Evaluation methods: 2h written test, can be retaken.

Evaluated skills:

  • Research and Development
  • Development

Course supervisor: Hervé Frezza-Buet

Geode ID: 3MD1540

CM:

  1. Datasets and learning (1.5 h)
  2. Frequentist, Bayesian, evaluation (1.5 h)
  3. Risks (1.5 h)
  4. C-SVC, Lagrange formulation (1.5 h)
  5. Kernels, numerical resolution (1.5 h)
  6. SVMs for regression, unsupervized learning, nu-versions of SVMs. (1.5 h)
  7. Arbres de décision (1.5 h)
  8. Bagging (1.5 h)
  9. Boosting (1.5 h)

TP:

  1. Data Science en Python (3.0 h)
  2. Arbres de décision (3.0 h)
  3. Bagging (3.0 h)
  4. Forêts aléatoires (3.0 h)
  5. TP+ 1/2 (4.0 h)
  6. TP+ 2/2 (4.0 h)