gaudiano@retina.bu.edu (Paolo Gaudiano) (03/05/90)
>Note that babies bootstrap, so that "knowing what to look for" becomes >increasingly sophisticated. >Note that networks are especially good at feature extraction. >So, combining my two short comments, why not build a network that bootstraps >on an increasingly complex environment? - in other words, build two networks. >There are a couple of ways you might do this; one way could be hierarchical; >one on top of another. The top network is the "knowing what to look for" unit, >which would provide a parallel "vigilance" input to the lower unit which >is actually performing the task in hand. I have been working on an adaptive model for autonomous control of arm trajectories. The non-adaptive model for arm trajectory formation is the VITE model of Bullock & Grossberg[1]. Grossberg and I have extended this to include adaptability, and also an additional circuit that autonomously generates vectors to the Adaptive VITE (AVITE) [2,3]. This works as a bootstrapping procedure that makes use EXCLUSIVELY of an internal measure of error to generate correct trajectory learning. We call the bootstrapping a "circular reaction" after Piaget [e.g.,4], who basically observed eye-hand coordination in children and suggested that during a circular reaction children "spontaneously" move their hand to some position, and as the eyes automatically follow the hand, a transformation is learned between motor and sensory domains. This is an oversimplified summary, but we have suggested that the Endogenous Generator (EG) circuit can be used universally. The babbling phase of speech acquisition is another example of this circular reaction. Also, funny that you should use the term "vigilance". Carpenter and Grossberg's Adaptive Resonance Theory (ART) [5,6,7] makes use of what they call a "vigilance parameter" to control the level of discrimination that must be met if two objects are to be categorized together. In his reply, Lawrence says: >The ultimate neural network would probably allow direct modification of >its weights (like in artificial models) and also "automatic" mode (like >in the brain). You'd get the best of both worlds. This is a convincing >argument (among others) that if an artificial brain is ever devised with >the capabilities of the human one, then there also exists one of the >former that is superior to any of the latter. well, ART is part of such a network. The ART model constantly compares incoming inputs with the model's internal representation of what that input should be (if such an internal representation already exists). ART has been embedded within (among other things) circuits that have explained large amounts of data on classical and instrumental conditioning [8,9,10]. Here ART acts at different levels: (1) it performs "object recognition" at the sensory input level, and (2) it is also embodied in the interactions between internal representations of the sensory inputs, and the drives that "motivate" behavior in the organism. Note that the concept of "Adaptive resonance" was first introduced in 1976 as a general concept [11], which has since been implicated in a number of different contexts (vision, speech, conditioning). It is only unfortunate that so much of Grossberg's work is so difficult to read. The unfortunate side is that--while *sometimes* his writing style is not the clearest--the reason for this difficulty is usually that the work is so overwhelming in scope. It takes a true interdisciplinary scientist to be able to follow through even the simplest of his papers. Without guidance, it is almost impossible to know which papers to read in which order, and what are the important points, etc. Oh well. Maybe in some years this stuff will be taught at the undergraduate level! Paolo REFERENCES: [1] D. Bullock, and S. Grossberg (1988a), ``Neural Dynamics of Planned Arm Movements: Emergent Invariants and Speed-Accuracy Properties During Trajectory Formation.'' {\em Psychological Review}, {\bf 95} (1), 49-90. [2] P. Gaudiano and S. Grossberg (1990) ``A Self-Regulating Endogenous Generator of Sample-and-Hold Random Training Vectors.'' In M. Caudill (Ed.) {\em International Joint Conference on Neural Networks. Washington, DC, January 1990.}\/ Hillsdale, NJ: Earlbaum. [3] P. Gaudiano and S. Grossberg (1990), ``A Self-Organizing Neural Circuit for Control of Planned Movement Trajectories.'' In preparation. [4] Piaget, J. (1963). {\em The origins of intelligence in children}. New York: Norton. {\bf 95(1)}, 49-90. [5] G. Carpenter and S. Grossberg (1986) ``A Massively Parallel Architecture for a Self-Organizing Neural Pattern Recognition Machine.'' {\em Computer Vision, Graphics, and Image Processing.}\/ {\bf 37}, 54-115. [6] G. Carpenter and S. Grossberg (1987) ``ART 2: self-organization of stable category recognition codes for analog input patterns.'' {\em Applied Optics},\/ {\bf 26}, (23), 4919-4930. [7] G. Carpenter and S. Grossberg (1990) ``ART 3: Hierarchical Search Using Chemical Transmitter in Self-Organizing Pattern Recognition Architectures'' {\em Neural Networks,} In Press. [8] S. Grossberg (1986) {\bf The Adaptive Brain 1: Cognition, Learning, Reinforcement, and Rhythm.} Amsterdam: Elsvier/North-Holland. [9] S. Grossberg and D. Levine (1987) ``Neural dynamics of attentionally modulated Pavlovian conditioning: blocking, interstimulus interval, and secondary reinforcement.'' {\em Applied Optics},\/ {\bf 26}, (23), 5015-5030. [10] S. Grossberg and N. Schmajuk (1987) ``Neural dynamics of attentionally modulated Pavlovian conditioning: conditioned reinforcement, inhibition, and opponent processing.'' {\em Psychobiology,}\/ {\bf 15} (3), 195-240. [11] S. Grossberg (1976) "Adaptive Pattern classification and universal recoding, II: Feedback, expectation, olfaction, and illusion". {\em Biological Cybernetics},\/ {\bf 23}, 187-202