Files in this item
|(no description provided)|
|Title:||Neural networks for signal processing and control|
|Author(s):||Hesselroth, Ted Daniel|
|Doctoral Committee Chair(s):||Schulten, Klaus J.|
|Department / Program:||Physics|
|Degree Granting Institution:||University of Illinois at Urbana-Champaign|
|Abstract:||Neural networks are developed for controlling a robot-arm and camera system and for processing images. The networks are based upon computational schemes that may be found in the brain.
In the first network, a neural map algorithm is employed to control a five-joint pneumatic robot arm and gripper through feedback from two video cameras. The pneumatically driven robot arm employed shares essential mechanical characteristics with skeletal muscle systems. To control the position of the arm, 200 neurons formed a network representing the three-dimensional workspace embedded in a four-dimensional system of coordinates from the two cameras, and learned a set of pressures corresponding to the end effector positions, as well as a set of Jacobian matrices for interpolating between these positions. Because of the properties of the rubber-tube actuators of the arm, the position as a function of supplied pressure is nonlinear, nonseparable, and exhibits hysteresis. Nevertheless, through the neural network learning algorithm the position could be controlled to an accuracy of about one pixel ($\sim$3 mm) after two hundred learning steps. Applications of repeated corrections in each step via the Jacobian matrices leads to a very robust control algorithm since the Jacobians learned by the network have to satisfy the weak requirement that they yield a reduction of the distance between gripper and target.
The second network is proposed as a model for the mammalian vision system in which backward connections from the primary visual cortex (V1) to the lateral geniculate nucleus play a key role. The application of hebbian learning to the forward and backward connections causes the formation of receptive fields which are sensitive to edges, bars, and spatial frequencies of preferred orientations. The receptive fields are learned in such a way as to maximize the rate of transfer of information from the LGN to V1. Orientational preferences are organized into a feature map in the primary visual cortex by the application of lateral interactions during the learning phase. The organization of the mature network is compared to that found in the macaque monkey by several analytical tests.
The capacity of the network to process images is investigated. By a method of reconstructing the input images in terms of V1 activities, the simulations show that images can be faithfully represented in V1 by the proposed network. The signal-to-noise ratio of the image is improved by the representation, and compression ratios of well over two-hundred are possible. Lateral interactions between V1 neurons sharpen their orientational tuning. We further study the dynamics of the processing, showing that the rate of decrease of the error of the reconstruction is maximized for the receptive fields used.
Lastly, we employ a Fokker-Planck equation for a more detailed prediction of the error value vs. time. The Fokker-Planck equation for an underdamped system with a driving force is derived, yielding an energy-dependent diffusion coefficient which is the integral of the spectral densities of the force and the velocity of the system. The theory is applied to correlated noise activation and resonant activation. Simulation results for the error of the network vs time are compared to the solution of the Fokker-Planck equation.
|Rights Information:||Copyright 1994 Hesselroth, Ted Daniel|
|Date Available in IDEALS:||2011-05-07|
|Identifier in Online Catalog:||AAI9512392|