Self organizing neural network architectures for adaptive pattern recognition and robotics

Author(s): Bullock, D. | Carpenter, G.A. | Grossberg, S. |

Year: 1991

Citation: In P. Antognetti & V. Milutinovic (Eds.), Neural Networks: Concepts, Applications, and Implementations, I, Englewood Cliffs, NJ: Prentice Hall, 33-53.

Abstract: In this chapter, we discuss some recent results in neural networks relevant to adaptive pattern recognition and sensory-motor control problems. IN biologically-oriented neural networks, whose fast dynamics are governed by slowly changing network transmission weights as well as by rapidly fluctuating external inputs, two major foci of research are (1) how to ensure that short-term dynamics are pattern preserving and (2) how to automatically regulate learning (transmission weight modification) in such a way that the system is guaranteed to develop along an adaptive trajectory. TO exemplify these issues, we will quickly move through a series of network constructs. ON the perceptual-cognitive end, we will discuss how recurrent competitive fields allow invariant pattern registration despite large fluctuations in input energies, and how an adaptive resonance module can learn a stable categorization-recognition code without an external teacher. ON the motor-control side, results on variable speed trajectory formation, sensory updating, learning anticipatory error compensation, and length-tension factorization illustrate networks that are applicable in several performance domains (planned arm and speech movements, ballistic eye-movements) and that help explain data on several distinct but cooperative neural regions (pre-central motor cortex, globus pallidus, cerebellum, spinal cord).

Topics: Biological Learning, Biological Vision, Machine Learning, Robotics,

PDF download

Cross References