Statistical Mechanics and Neural Networks

Informal Workshop

Wednesday, 13-Jun-2001, 2 pm - 5 pm

Time Speaker Title Abstract
14:00 Michael J. Barber Neural belief propagation without multiplication
Neural network models can be derived from the hypothesis that populations of neurons perform statistical inference. Such networks can be generated from a broad class of probabilistic models, but often function through the multiplication of neural firing rates. By introducing additional assumptions about the nature of the probabilistic models, we derive a class of neural networks that function only through weighted sums of neural activities.
14:30 Arnaud Buhot Strong and fragile glassy behaviour in kinetically constrained systems
15:15   Coffee break
15:45 Michael J. Barber Noise-induced signal enhancement in heterogeneous neural networks
Neural networks can represent complex functions, but are often constructed of very simple units. We investigate the limitations imposed by such a simple unit, the McCulloch-Pitts neuron. We explore the role of stochastic resonance in units of finite precision and show how to construct neural networks that overcome the limitations of single units.
16:15 Sebastian Risau-Gusman Typical properties of Soft Margin Classifiers
16:45   Free discussion