Reinforcement learning and conditioning . . The activity of neurons in the basal ganglia code for unpredictable rewards and is probably involved in reinforcement learning and conditioning. How is this signal generated? What are the temporal conditions? What is the formal relation to mathematical theories of reinforcement learning? Robot learning.. How can bio-inspired evolutionary algorithms help robot learning? How do genetic changes over generations interact with learning on the level of the individual? The project is headed by Dario Floreano who has joined in spring 2000 the Department of Microtechnique at the EPFL.

Saccadic Pattern Recognition and Support Vector Machines.. The human saccadic system allows efficient selection of task relevant information. Can the same principles be applied to pattern recognition tasks such as face authentication? How can classification be improved by modern machine learning methods like Support Vector Machines?

Modeling temporal processing in the auditory system.. The auditory system enables us to recognize rapid sound transients ('clicks') which last for less than thirty milliseconds. How can we understand the processing steps at the level of the cochlea, the auditory nerve, and the auditory nuclei? Can auditory models be useful in the context of speech recognition?

Applications of neural networks. Standard neural networks are powerful and user-friendly tools for several practical applications, such as data mining, time series analysis, forecast, production control, image analysis, etc. Since its creation, the Mantra Centre has collaborated with industrial partners, governmental offices, and other research units for the development of neural network tools in specific application domains, such as weather analysis, avalanche risk forecast , coin recognition, forecast of powerline charge, etc. Some of these activities are carried out in collaboration with Prof. Alain Germond from the Electrical Engineering Department and Prof. Popovich from the Micro-Engineering Department, and Dr. Alain Sc horderet from the Mechanical Engineering Department.

In collaboration with the IDIAP in Martigny, we study Mixtures of Experts.. Modular neural network models, known as mixtures of experts, consist of a gating networks that partition the input space and expert networks that treat the resulting subproblems.