LEARNING OUTCOMES

The goal of the class is to present Artificial Neural Networks and other well-known Machine Learning techniques as systems for solving supervised and unsupervised learning problems, with a specific emphasis on Robotics applications. Such learning systems can be applied to pattern recognition, function approximation, time-series prediction and clustering problems. Some mention will be made to the use of ANNs as static systems for information coding, and dynamical systems for optimization and identification.

AIMS AND LEARNING OUTCOMES

After successfully attending this course, students will have an exposure to many topics that underlie the field of machine learning, so that they will be able to autonomously apply the methods presented as well as other methods to concrete problems. During practical activities, students will both implement several methods from scratch, and use existing machine learning libraries, thus gaining a hands-on experience backed up by the theoretical concepts.

PREREQUISITES

• Basic multi-dimensional calculus
• Continuous optimization
• Probability and some information theory
• Discrete proficiency in programming (one of Matlab or Python, or ability to quickly catch up if coming from different programming backgrounds)

TEACHING METHODS

• Lectures
• Practical assignments, formatted as homeworks but also worked out with assistance by the teacher during lab hours, to be handed in every 2 weeks

Assignments are used for continuous assessment whose weight is 50% of the final marks, the rest being obtained with a final exam and discussion.

Due to the teaching style and to the continuous assessment, attendance is mandatory

SYLLABUS/CONTENT

1. Introduction
2. Perceptual problems
3. The decision problem in the presence of complete deterministic information: Representation problems
4. The decision problem in the presence of complete probabilistic information: Bayes decision theory
5. The decision problem in the presence of incomplete samples (data): Statistics and the learning problem. Inductive bias, the bias-variance dilemma
6. Parametric methods and maximum likelihood estimation
7. Non-parametric methods, some popular classification and clustering methods
8. Evaluating learning: Indexes and resampling methods.
9. Neural networks: Historical methods, shallow networks
10. The learning problem as optimization. Algorithms and strategies.
11. Data mapping: Dimensionality reduction and kernel methods 
12. Deep neural networks
13. Learning from sequential data

RECOMMENDED READING/BIBLIOGRAPHY

Course slides and assignments are available on the official study portal.

A selection of suggested readings (journal articles and textbooks) will be provided during lectures.