hudak@siemens.UUCP (Michael J. Hudak) (12/03/87)
SEMINAR ANNOUNCEMENT Professor David Rumelhart Department of Psychology Stanford University Palo Alto, CA Title: Learning and Generalization in PDP Networks Location: Siemens Corporate Research & Support, Inc Princeton Forrestal Center 105 College Road East Princeton, NJ 08540-6668 (609/734-3373) 3rd floor Multi-Purpose Room Date: Wednesday December 9, 1987 Time: 10:00 am (refreshments: 9:45)
hudak@siemens.UUCP (Michael J. Hudak) (01/07/88)
SEMINAR ANNOUNCEMENT Speaker: Peter Cariani Systems Science Dept., Thomas J. Watson School of Engineering State University of New York at Binghamton Title: Structural Preconditions for Open-Ended Learning through Machine Evolution Location: Siemens Corporate Research & Support, Inc. 3rd floor Multi-Purpose Room Princeton Forrestal Center 105 College Road East Princeton, NJ 08540-6668 Date: Thursday, 14 January 1988 Time: 10:00 am (refreshments: 9:45) For more information call Mike Hudak: 609/734-3373 Abstract One of the basic problems confronting artificial life simulations is the apparent open-ended nature of structural evolution, classically known as the problem of emergence. Were it possible to construct devices with open-ended behaviors and capabilities, fundamentally new learning tech- nologies would become possible. At present, none of our devices or models are open-ended, due to the nature of their design and construction. The best devices we have, in the form of trainable machines, neural net simulations, Boltzmann machines and Holland-type adaptive machines, exhibit learning within the categories fixed by their feature spaces. Learning occurs through the performance dependent optimization of alter- native I/O functions. Within the adaptive machine paradigm of these devices, the measuring devices, feature spaces, and hence the real world semantics of such devices are stable. Such machines cannot create new primitive categories; they will not expand their feature and behavior spaces. Over phylogenetic time spans, however, organisms have evolved new sensors and effectors, allowing them to perceive more and more aspects of their environments and to act in more and more ways upon those environments. This involves a whole new level of learning: the learning of new primitive cognitive and behavioral categories. In terms of constructible devices, this level of learning encompasses machines which construct and select their own sensors and effectors, based upon their real world performance. The semantics of the feature and behavior spaces of such devices thus changes so as to optimize their effectiveness as categories of perception and action. Such devices construct their own primitive categories, their own primitive concepts. Evolutionary devices could be combined with adaptive ones to both optimize primitive categories and I/O mappings within those categories. Evolutionary machines cannot be constructed through computations alone. New primitive category construction necessitates that new physical measuring structures and controls come into being. While the behavior of such devices can be represented to a limited degree by formal models, those models cannot themselves create new categories vis-a-vis the real world, and hence are insufficient as category-creating devices in their own right. Computations must be augmented by the physical construction of new sensors and effectors implementing processes of measurement and control respectively. This construction process must be inheritable and replicable, hence encodable into symbolic form, yet involving the autono- mous, unencoded dynamics of the matter itself. The paradigmatic example of a natural construction process is protein folding. A one-dimensional string of nucleotides, itself a discrete, rate-independent symbolic structure, is transformed into continuous, rate dependent dynamics having biological function through the action of the physical properties inherent in the protein chain itself. The functional properties of speed, specificity, and reliability of action are thus achieved with symbolic constraints but without the explicit direction of rules.