Jure Taslak: Logic for learning

Datum objave: 14. 6. 2022
Seminar za temelje matematike in teoretično računalništvo
četrtek
16
junij
Ura:
10.00 - 12.00

The typical approach to machine learning is to encode examples as vectors in $\mathbb{R}^n$ and then use classic algorithms such as linear regression, support vector machines, neural networks, etc. With this approach much dependens on the way data is encoded as vectors – an inappropriate encoding may result in information loss.

In the talk I will describe an alternative approach, where we encode data as terms in a higher-order logic and learn directly from their structure. The focus throughout is on learning comprehensible theories.

I will define the logic used to encode examples, the class of terms used to represent them and develop an approach of constructing predicates which serve as the hypothesis language. In the end I will describe an algorithm that one can apply to the learning problem to learn a hypothesis.