Neural networks in spatiotemporal pattern recognition

Mohimm Daniel Tom, Purdue University

Abstract

Neural network research, long focused on static pattern recognition, is now extended to spatiotemporal pattern recognition. A simple windowing technique extends the use of a feedforward network as a static pattern associator to a stored temporal pattern associator. Experiments with real-valued inputs show that the capabilities of the feedforward network is not limited to the binary domain. The combination of the above two techniques produces a pattern recognizer for a speech pattern. In order to perform spatiotemporal pattern recognition in real-time, the storage or memory problem must be solved. A study of the accommodation processes of the neuron gives rise to the hystery unit, a computational neuron model having short term memory. Theorems show that the hystery unit converges under repeated pattern presentation, and therefore help to explain the transformation of short term memory into long term synaptic memory. The hystery unit encodes its entire history of bipolar temporal inputs in a single final output value, from which the input sequence can be regenerated. This perfect memory property of the hystery unit enables it to perform delta modulation sequence recognition in real-time. With the help of the MIND units, hystery units discriminate between noise embedded spatiotemporal patterns in real-time with very high accuracy. The contribution of the hystery unit extends beyond that of pattern recognition. It introduces short term memory characteristics into otherwise memoryless computational neuron architectures, thus enabling future neural networks to be more powerful computational models of the brain.

Degree

Ph.D.

Advisors

Tenorio, Purdue University.

Subject Area

Electrical engineering|Artificial intelligence

COinS