ONLINE GIBBS LEARNING

We propose a new model of on-line learning which is appropriate for le arning realizable and unrealizable, smooth as well as threshold, funct ions. Following each presentation of an example the new weights are ch osen from a Gibbs distribution with an on-line energy that balances th e need to minimize the instantaneous error against the need to minimiz e the change in the weights. We show that this algorithm finds the wei ghts that minimize the generalization error in the limit of an infinit e number of examples. The asymptotic rate of convergence is similar to that of batch learning.