This paper introduces a dynamic formal neuron model to address the recognition of spatio-temporal bipolar patterns. It extends traditional neural network models by incorporating convolution and temporal adaptation mechanisms. The model was applied to phoneme classification and showed robust performance in distinguishing 30 prototype phoneme patterns.
The paper highlights the significance of spatio-temporal pattern recognition in both biological and artificial systems. It critiques the limitations of static pattern models and proposes a dynamic formal neuron as a solution. Applications in fields such as speech recognition and robotics are emphasized.
The authors review foundational concepts in neural networks, including Hebbian and delta learning rules, and highlight the inadequacies of static models for dynamic tasks. Previous solutions, such as storage neurons and feedback layers, are contrasted with the proposed approach.
A dynamic formal neuron was designed by replacing weights with transfer functions and multiplication with convolution. The Hebbian and delta learning rules were generalized to accommodate temporal patterns. The model was trained and tested on phoneme patterns synthesized from speech signals.
The dynamic formal neuron demonstrated adaptability and accuracy in classifying phoneme patterns. Model 1 (Matched Filter Bank) performed well for consistent phoneme durations but struggled with variability. Model 2 (Pseudo-Inverse Filter) showed higher robustness and accuracy, even with noisy data.
Model 1 correctly classified phonemes with minor errors in both clean and noisy patterns. Model 2 achieved perfect classification on clean data and had minimal errors with noisy data. The results validate the effectiveness of dynamic formal neurons for spatio-temporal tasks.
The study successfully generalized static neural networks to dynamic ones, enabling robust spatio-temporal pattern recognition. Future work includes exploring recursive filters and refining learning rules for nonlinear systems.