A pulsed neural network capable of universal approximation

Area of Science:

• * Computational neuroscience
• * Artificial intelligence
• * Neural network modeling

Background:

• * The Cerebellar Model Articulation Controller (CMAC) is a widely recognized adaptive control system.
• * Traditional CMAC models often rely on discrete or continuous value representations.
• * There is a need for alternative CMAC architectures that can process temporal information efficiently.

Purpose of the Study:

• * To introduce and describe a novel pulsed network version of the Cerebellar Model Articulation Controller (CMAC).
• * To investigate the functional approximation capabilities of this pulsed CMAC network.
• * To demonstrate the effectiveness of training the pulsed CMAC using a least mean square algorithm.

Main Methods:

• * Development of a pulsed neural network architecture based on CMAC principles.
• * Mathematical formulation of the network's output pulse timing as a function of input intervals.
• * Application of a least mean square (LMS) algorithm for network training.

Main Results:

• * The pulsed CMAC network produces output pulses whose timing is dependent on input pulse intervals.
• * The network demonstrated the ability to approximate bounded measurable functions within causality constraints.
• * Simulation results confirmed the successful training of the pulsed CMAC using the LMS algorithm.

Conclusions:

• * The pulsed CMAC network offers a viable alternative for adaptive control and function approximation.
• * The temporal processing capability of the pulsed network expands the applicability of CMAC models.
• * The study validates the use of LMS for training this novel pulsed CMAC architecture.