Komo MaiResearch
Publications TeachingResearch Interests: Active learning and optimal predictions.
Most research in learning theory deals with passive learning. However, many real world learning problems are interactive, and so is animal learning.
The theoretical foundations for interactive learning and behavior are much less developed than those for passive learning. A theoretical understanding of behavioral learning lies at the heart of a new generation of machine intelligence, and is also at the core of many interesting questions about adaptation and learning in biology.
S. Still. Statistical Mechanics approach to interactive learning. 2005/2007(revised). http://arxiv.org/abs/0709.1948
Slides of a talk I gave at the NIPS workshop on Models of Behavioural Learning.
Student Projects (499, 699) and thesis projects:
Applications:
1. Control problems
This project aims at utilizing the
above mentioned theory in classical control
problems. We will compare our results
to reinforcement learning.
2. Exploration
What should a rat
or a robot do when
entering a new environment? We will
compare theoretical predictions with computer simulations and behavioural
measurements.
compare theoretical predictions with computer simulations and behavioural
measurements.
Theory:
1. Action planning
We will work on
extending the existing theory to action planning. This could
be worked out within the scope of a Master Thesis, or a PhD Thesis.
be worked out within the scope of a Master Thesis, or a PhD Thesis.
2. Convergence in the presence of non - i.i.d. data
Theorems for
convergence of the
normalized count to the actual probability
distribution rely on i.i.d. data, as do most bounds for generalization errors in
statistical learning theory. However, in the active learning scenario, the data
are not necessarily drawn i.i.d. This project is targeted towards a student with
an interest in mathematical statistics.
Psychophysics:distribution rely on i.i.d. data, as do most bounds for generalization errors in
statistical learning theory. However, in the active learning scenario, the data
are not necessarily drawn i.i.d. This project is targeted towards a student with
an interest in mathematical statistics.
1. Active learning in human behaviour
We
will design, and conduct, a
psychophysical experiment, focussing on a
very simple learning task. We will ask questions about the learning behaviour
of humans. Do they chose actions in some optimal way? Does interaction
with the world help to learn about the world?
very simple learning task. We will ask questions about the learning behaviour
of humans. Do they chose actions in some optimal way? Does interaction
with the world help to learn about the world?