Richard.Wallstein@A.CS.CMU.EDU.UUCP (06/12/86)
The CMU Summer Research Seminar Series continues this Friday, June 13 at 2:30 PM, 7500 WeH with a talk by Geoffrey Hinton on his new research: A New Algorithm for Learning by Selection Imagine a complicated non-linear process that contains specific steps that are controlled by switches which can be on or off. Each switch has a particular stored probability of being on. Using these probabilities, we generate a random combination of switch settings and then run the process and decide whether the result is good or bad. I shall describe a new learning algorithm that uses information about the goodness of the outcomes to revise the stored probabilities associated with the switches. The algorithm is guaranteed to change the switch probabilites in such a way that future random combinations of switch settings are more likely to produce good outcomes. It can be applied to stochastic processes of arbitrary complexity. If each switch is a synapse, it suggests a new model of learning in the cortex. If each switch is an enzyme and its stored probability is the relative frequency of the relevant gene in the gene pool, the learning algorithm is an efficient way of using the information provided by survival to optimize gene frequencies. The extension to optimizing frequencies of gene combinations appears to be feasible.