Year: 1989
Type: article
Abstract: Although there are straightforward methods for training neural nets to show certain types of sequential behavior, the general problem of structural and temporal credit assignment (appropriate weight changes) in networks existing in a changing environment remains unsolved.In this paper we argue that a learning mechanism for nding temporal input/output relations ought to depend solely on computations local in both space and time, and that no teacher should be required to indicate the starts and ends of relevant sequences to the network.We ask whether there are learning rules which do not depend on such external hints yet still can deal with `hidden units' and with units whose past activations are `hidden in time'.(Back propagation, for instance, is not even local in time.)We propose a discrete time version and a continuous time version of a simple on-line method local in space and time which is designed to deal with hidden states.The approach is inspired by Holland's idea of the bucket brigade for classi er systems, which is transformed to run on a neural network with xed topology.The result is a recurrent `neural' dissipative system which is consuming `weight-substance' and permanently trying to distribute this substance onto its connections in an appropriate way.Simple experiments demonstrating the feasability of the algorithm are reported.
Author: J. Schmidhuber
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
Funder: Siemens
OpenAlex ID: W41338955