Recipients of this mailing list may be interested to learn about a newly discovered neuro-computational mechanism. This mechanism, which I named "conceptors", suggests a common computational principle underneath phenomena as diverse as - incremental learning of dynamical patterns, - attentional focussing, - neural noise suppression, - morphing and modulating motor patterns, - learning pattern classes from a few prototype examples, - top-down bias control in hierarchical architectures, - Boolean combination of evidence in pattern classification. Conceptor-based algorithms are simple, cheap, and robust against noise and parameter variation. A number of fundamental neuro-computational problems become efficiently solved: - conceptors make it possible in the first place to govern a diversity of processing modes in a single recurrent neural network, - they enable incremental learning of temporal patterns with an option to quantify and monitor claimed neural memory capacity (coping with the catastrophic interference problem), - they lead to an auto-associative memory for dynamical patterns (a generalization of Hopfield networks to temporal patterns), - Boolean concept combinations and conceptual abstraction can be rigorously interpreted and implemented by the dynamics of recurrent neural networks (a robust solution for core parts of the neuro-symbolic integration problem). More information: 1. A video of a complex human motion sequence governed by a single RNN whose modes are governed by conceptors: http://youtu.be/DkS_Yw1ldD4 2. An informal introduction with many pictures (11 pages): http://arxiv.org/abs/1406.2671 3. A comprehensive technical report, including a link to Matlab code (195 pages): http://arxiv.org/abs/1403.3369 ------------------------------------------------------------------ Dr. Herbert Jaeger Professor for Computational Science Jacobs University Bremen gGmbH Campus Ring 28759 Bremen, Germany Phone (+49) 421 200 3215 Fax (+49) 421 200 49 3215 email h.jaeger@jacobs-university.de http://minds.jacobs-university.de ------------------------------------------------------------------