2022/2023
Statistical Learning Theory
Type:
Mago-Lego
Delivered by:
Big Data and Information Retrieval School
When:
1, 2 module
Open to:
students of all HSE University campuses
Language:
English
ECTS credits:
8
Contact hours:
54
Course Syllabus
Abstract
We study a theory that inspired the development of two important families of machine learning algorithms: support vector machines and boosting. More generally, in a typical classification task we are given (1) a dataset for training and (2) a familyof classifiers, indexed by some parameters that need to be tuned. A learning algorithm uses the training set to select one of the classifiers, in other words, tune the parameters. For example, given a neural network, every choice of the weights specifies a classifier and the learning algorithm assigns values to the weights. On one side, we want to have a large set of classifiers to model all structure in the training data. This might lead to a large number of parameters to train. On the other hand a too large set of classifiers might lead to a decrease of accuracy on unseen examples. This is called overfitting. We study a theory that quantifies overfitting theoretically. Moreover, support vector machines and boosting algorithms can be seen as algorithms that optimize the trade-off discussed above under some additional assumptions. Moreover, the theory can determine good values for meta-parameters in machine learning algorithms that otherwise need to be tuned using cross-validation. We also study some recent deep boosting algorithms that were developed using the theory. These algorithms are currently among the best for classification tasks when the number of classes is high. Finally, we study the online mistake model. This model is more general but its mathematical analysis has many similarities with the theories above.