Seminario di Informatica: Model Selection and Error Estimation in Learning from Empirical Data

Venerdì 26 Giugno 2015, ore 14:30 - Aula 1BC50 - Luca Oneto



Venerdì 26 Giugno 2015 alle ore 14:30 in Aula 1BC50, Luca Oneto (SmartLab, DITEN, University of Genoa) terrà un seminario dal titolo "Model Selection and Error Estimation in Learning from Empirical Data".

In the Supervised Learning framework, a model is built by exploiting the available observations through a Learning Algorithm that is able to capture the information hidden in the data. Model Selection addresses the problem of tuning a Learning Algorithm to the available data in order to reduce the Generalization (True) Error of the final model. This problem affects most of the algorithms because, in general, their effectiveness is controlled by one or more hyperparameters which must be tuned during the learning process for achieving optimal performances. Associated to the issues of Model Selection we find the problem of estimating the True Error of a classifier: in fact, the main objective of building an optimal classifier is to choose the parameters and hyperparameters that minimize its True Error and compute an estimate of this value for predicting the classification performance on future data. Unfortunately, despite the large amount of work done on this important topic, the problem of Model Selection and Error Estimation for a Learning Algorithm is still open and the focus of extensive research. The purpose of this seminar is to give an overview of the problem of Model Selection and Error Estimation. We will start from the seminal works of the 80s until the most recent results on this topic. Finally we will discuss future directions of this multidisciplinary field of research.

Download Seminari di Informatica