LEARNING AND PATTERN RECOGNITION

Main lecturer

Mail address

Phone number

Pierre Charbonnier, Directeur de Recherche MEDDE

p.charbonnier@unistra.fr, Pierre.charbonnier@cerema.fr

+33 (0)3 68 77 46 44

Other instructor(s)

“N/A”

APOGEE code

Track - Year - Option - Semester

Coefficient = ECTS

Duration

EP013M62

Engineer - 3Y G ISSD - S9

Master - 2Y ID G + HCI / Topo - S3

2.25 / 5 / 2 (Topo)

15,75h CM, 15,75h TP (Computer Exercises)

EXAMS

Duration

Authorized documents

      If yes, which ones :

School calculator authorized

Session 1

1h30

Yes

lecture notes

Matlab

Session 2

1h30

Yes

lecture notes

Matlab

Prerequisites

Basic knowledge in statistics, probabilities and image processing.

Lecture goals

The goals of this lecture are to introduce the main statistical and neuromimetic approaches of machine learning and pattern recognition.

Detailed outline

• Introduction
• Statistical approaches: Bayesian decision theory, study of the Gaussian case,
• Probability density estimation, parametric methods, non-parametric methods : Parzen windows, k-nearest neighbours
• Dimensionality reduction techniques: feature extraction using Principal Component Analysis, Fisher linear discriminant analysis, MDS and ISOMAP. Feature selection methods
• Unsupervised classification, or clustering: mixture densities and the EM algorithm, the k-means algorithm, Meanshift, hierarchical clustering techniques, HDBSCAN  
• Linear and non-linear discriminant functions: neural approaches, the multi-layer Perceptron,
  Radial Basis Functions, Support Vector Machines
• Ensemble learning, decision trees, random forests, AdaBoost
• Evaluation of classifiers
• Conclusion

Applications

• 1D Bayesian classification
• Parametric density estimation, 2D Bayesian classification
• Naive Bayes classifier
• Non-parametric probability density estimation methods (Parzen, k-NN), Bayesian classifiers based on non parametric density estimators
• Principal Component Analysis, snapshot PCA, application in pattern recognition
• Fisher linear discriminant and supervised classification
• Unsupervised classification using the k-means algorithm
• Linear discriminants : least squares, logistic regression, Perceptron
• Non-linear extensions : kernel logistic regression, kernel Perceptron
• Evaluation, ROC curves

Acquired skills

After this course, the student will know the basics of the main machine learning and pattern recognition algorithms and will be able to apply them.