AG 1, AG 2, AG 3, AG 4, AG 5, RG1, SWS, MMCI   Info

Public Audience

English

60 Minutes

Saarbrücken

For most currently successful machine learning techniques performance guarantees are only available when assuming an "i.i.d." scenario: one is given a training set consisting of independent and identically distributed samples, and one aims at building a model that will work well on future samples from the same distribution. Unfortunately, in many real-world situations the i.i.d. property does not hold, causing existing techniques to become unreliable or outright fail.

In my talk, I will discuss recent results from our group for situations in which the "future" data distribution differs from the distribution of the available training data. In particular, I will highlight new theoretical guarantees we were able to obtain for the situation of multi-task learning, and I will show how such theoretical results can give rise to new and principled machine learning algorithms.

Connie Balzert

--email hidden

passcode not visible

logged in users only