Alexandra Carpentier, Yann Ollivier
Achtung: geänderte Uhrzeiten !
14:30 Uhr Yann Ollivier (Paris-Saclay)
Artificial intelligence and inductive reasoning: from information theory to neural networks
In order to one day build an artificial intelligence, one must face inductive reasoning problems such as prediction ("guess what comes next after this sequence of numbers"), generalization, or inference (finding "explanations" for given observations). It seems at first that these problems are not rigorously defined from a mathematical point of view. Actually, there is a well-defined mathematical theory of inductive reasoning, based on information theory. But this theory, though extremely elegant, is hard to apply in practice.
Today, neural networks (a.k.a. "deep learning") are a technique of choice in a series of practical problems (image recognition, speech recognition, Go games, self-driving cars...). In this talk we will present the general approach of neural networks, and make the link with the mathematical framework of information theory.
15:30 Uhr coffee break
16:00 Uhr Alexandra Carpentier (Universität Potsdam)
Sequential learning : the bandit approach to decision-making
A pillar of modern Artificial Intelligence (AI) - or Machine Learning - is automatic and sequential learning for decision making. More and more decisions - ranging from trade decisions on high frequency markets or ad placement in online advertisement, to decisions in specific medical trials problems - have nowadays stopped being made by humans and are made by AIs of some kind. These AIs need first to explore the environment they are in (e.g. collecting informations on placements, ads, or treatments), so that they can eventually reach their goal (e.g. maximizing the profit, the number of clicks, or the number of cured patients).
Balancing between these two tasks, namely exploration of the environment for acquiring knowledge, and exploitation of this knowledge toward the global aim, is key for designing efficient AIs. A simple mathematical formulation of this problem is called the bandit problem. Its theoretical study enables to better understand how to balance the exploration and exploitation tasks, ultimately leading to the design of efficient AIs.
