Neural Networks, Supercomputers and Computer Vision
A neural network is a crude attempt to reproduce the functions of a brain by imitating its anatomical structure. Like the brain, these machines are composed of many highly interconnected units. Neural networks can be simulated by software on traditional computers or by building hardware neural computers. The software approach, when applied to large memory supercomputers offers an interesting avenue to experimentation with larger networks and more globally oriented learning algorithms.
Even though such supercomputers only parallelize by vectoring, the PDP model vectorizes very well. The large memory (which is shared by all nodes) allows the storage of large networks and the great speed of of such machines allows learning algorithms to continue through many cycles. In addition it is possible to think in terms of global learning algorithms in which all learning patterns are presented to the algorithm in advance and which stay resident in memory during all the learning phase. A PDP program written in Fortran at NASA Johnson Space Center was modified, vectorized and optimized for the NEC SX-2. A typical learning run involving 100 input nodes, 100 middle nodes and less than 10 output nodes required only one CPU second for computation and completed in less than three or four wall-clock seconds.
We apply this program to problems in machine vision. The DARPA benchmark computer vision program involving a 2.5 dimensional partially obscured mobile is a recent example of hypothesis generation based on a data base of models. The PDP program avoids explicit pattern matching with reference model segments as well as the creation of hypotheses. The idea is to utilize the neural network’s ability to perform pattern matching with distorted and incomplete data. We consider the problem of recognizing simple four sided polygons in a 2D scene of straight lines. The scene may or may not contain the given polygons and if it does, there may be multiple occurrences of them. Further, the given polygonal shapes may be partially present and one may be “behind” another.
KeywordsNeural Network Input Pattern Hide Unit Input Node Output Pattern
Unable to display preview. Download preview PDF.
- Ballard D. and Brown C, Computer Vision, Prentice Hall (1982).Google Scholar
- Crimson W.E.L., From Images to Surfaces, MIT Press (1981).Google Scholar
- Horn B.K.P., Robot Vision, MIT Press (1986).Google Scholar
- Hough P.V.C. “Method and Means for Recognizing Complex Patterns”, U.S. Patent. 3,069,654 (1962).Google Scholar
- Minsky M. and Papert S., Perceptrons, MIT Press (1968), Expanded Edition (1988).Google Scholar
- Rosenblatt F., Principles of Neurodynamics: Perceptrons and the Theory of Brain Mechanisms, Washington D.C., Spartan (1961).Google Scholar
- Rosenfeld A. and Kak A., Digital Picture Processing, Academic Press (1982).Google Scholar
- Rumelhart D.E. and McClelland J.L., Parallel Distributed Processing, MIT Press (1986).Google Scholar
- Waltz D.L., “Generating Semantic Descriptions from Drawings of Scenes with Shadows”, The Psychology of Computer Vision P.H. Winston ed., McGraw-Hill (1972).Google Scholar
- McClelland J.L. and Rummelhart D.E., “Explorations in Parallel Distributed Processing”, MIT Press, (1987).Google Scholar
- Rumelhart D.E., Hinton and Williams, “Learning Internal Representation by Error Propagation”, Chapter 8, PDP Volume 1, MIT Press, (1986).Google Scholar
- Rumelhart D.E., Hinton and McClelland J.L., “A General Framework for PDP”, Chapter 2, PDP Volume 1, MIT Press, (1986).Google Scholar
- Stone, “An Analysis of the Delta Rule and the Learning of Statistical Associations”, Chapter 11, PDP Volume 1, MIT Press, (1986)Google Scholar
- Sejnowski,T.J. and Rosenberg, C.R.,“NETtalk: A Parallel Network that Learns to Read Aloud”, Technical Report JHU/EECS-86/01, John Hopkins University, Baltimore, Md. (1986).Google Scholar
- Weems, C., “The DARPA Benchmark SUN-3 Cartridge Tape”, University of Massachusetts, Amherst, Ma. (1988).Google Scholar