logo-polimi
Loading...
Risorse bibliografiche
Risorsa bibliografica obbligatoria
Risorsa bibliografica facoltativa
Scheda Riassuntiva
Anno Accademico 2019/2020
Tipo incarico Dottorato
Insegnamento 055047 - ONLINE LEARNING AND MONITORING
Docente Boracchi Giacomo
Cfu 5.00 Tipo insegnamento Monodisciplinare

Corso di Dottorato Da (compreso) A (escluso) Insegnamento
MI (1380) - INGEGNERIA DELL'INFORMAZIONE / INFORMATION TECHNOLOGYAZZZZ055047 - ONLINE LEARNING AND MONITORING
055210 - ONLINE LEARNING AND MONITORING

Programma dettagliato e risultati di apprendimento attesi

AIMS AND SCOPE

This course provides an overview of Machine Learning (ML) methods that are meant for streaming data and that force the learner to operate in an online or incremental manner. These settings are often encountered in real-world applications, e.g., to select sponsored links for Internet advertising, or to detect frauds in credit card transaction. The online setting poses relevant challenges to classical data-driven solutions since i) the model has to integrate new pieces of information as soon as they become available, ii) the learning algorithm has to adapt to the current operating conditions, iii) the learning algorithms have to be computationally efficient, to be executed in real-time.

 

GOALS

Provide the student with a general overview of problems rising when using ML techniques on streaming data, understand the most prominent algorithm from the literature and their use to solve real-world problems.

 

PROGRAM

At first, we provide an overall introduction and a unifying formulation of the addressed learning problems and an overview of the involved research challenges (2 hours approximately). Then, we illustrate different types of models and landmark algorithms for incremental learning in evolving environments, together with change-detection tests, which are meant to detect distribution changes and can be used to steer model adaptation (6 hours approximately). We cover sequential decision-making problems in the framework of online learning, considering both the bandit and the expert settings (6 hours approximately). After that, we present techniques to detect anomalous samples in the incoming data  (outliers), as well as domain adaptation solutions (5 hours approximately).

Finally, we cover some landmark applications of the above described techniques, which combines online/incremental learning algorithms with those techniques explicitly modeling the change in the environment (6 hours approximately). In the specific, an online advertising selection method, a bank fraud detection framework, and an automatic ECG monitoring.

In particular the course presents:

  • Incremental learning algorithms, which can be seen as a natural extension of classical ML techniques (supervised and unsupervised) to the settings where samples are provided over time and/or the data distribution is changing over time;
  • Change detection methods, which can be used to steer model adaptation;
  • Online learning algorithms, which have been developed in the reinforcement learning literature, and are specifically crafted to deal with sequential decision problems;
  • Domain adaptation algorithms, which can be used to adapt a pre-trained model to operate in different settings;
  • Anomaly Detection methods.

 

TEACHING ORGANIZATION 

Traditional lectures for presenting theory and the most important algorithms. Guided computer-laboratory hours to illustrate applications and getting familiar with the problem. Student will fill in simple Matlab code snippets to implement the key algorithms of the field.

 

PREREQUISITES

Students that are proficient in machine learning and statistics.

 


Note Sulla Modalità di valutazione

Students have to prove proficiency in dealing with online learning settings and the landmark algorithms presented. Final assessment will be based on class participation and a discussion with teachers. Projects are available upon request.


Intervallo di svolgimento dell'attività didattica
Data inizio
Data termine

Calendario testuale dell'attività didattica
  • May, 6 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 8 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 13 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 15 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 20 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 22 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 27 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20
  • May, 29 2020 from 14:30 till 17:30, Sala Seminari - Ed. 20

 


Bibliografia
Risorsa bibliografica obbligatoriaShalev-Shwartz, Shai, Online learning and online convex optimization, Editore: Foundations and Trends in Machine Learning 4.2 : 107-194, Anno edizione: 2012
Risorsa bibliografica obbligatoriaCesa-Bianchi, Nicolo, and Gabor Lugosi., Prediction, learning, and games, Editore: Cambridge university press, Anno edizione: 2006
Risorsa bibliografica obbligatoriaBasseville, Michèle, and Igor V. Nikiforov., Detection of abrupt changes: theory and application, Editore: . Vol. 104. Englewood Cliffs: Prentice Hall., Anno edizione: 1993
Risorsa bibliografica obbligatoriaPan, Sinno Jialin, and Qiang Yang., A survey on transfer learning, Editore: IEEE Transactions on knowledge and data engineering 22.10 (2009): 1345-1359., Anno edizione: 2009

Mix Forme Didattiche
Tipo Forma Didattica Ore didattiche
lezione
25.0
esercitazione
0.0
laboratorio informatico
0.0
laboratorio sperimentale
0.0
progetto
0.0
laboratorio di progetto
0.0

Informazioni in lingua inglese a supporto dell'internazionalizzazione
Insegnamento erogato in lingua Inglese
Disponibilità di materiale didattico/slides in lingua inglese
Possibilità di sostenere l'esame in lingua inglese

Note Docente
schedaincarico v. 1.6.1 / 1.6.1
Area Servizi ICT
08/12/2019