Universal perceptron and DNA-like learning algorithm for binary neural networks: LSBF and PBF implementations

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Machine Learning

Background:

• The Universal Perceptron (UP) generalizes Rosenblatt's perceptron, capable of implementing all Boolean functions (BFs).
• Boolean functions are classified into linearly separable (LSBF), parity (PBF), and other complex classes.
• Existing algorithms for training perceptrons can be computationally intensive and complex.

Purpose of the Study:

• To introduce a novel Universal Perceptron (UP) architecture and a DNA-like learning algorithm.
• To efficiently implement various Boolean functions (BFs), with a focus on LSBFs and PBFs.
• To develop a new measure, nonlinearly separable degree (NLSD), for classifying BF complexity.

Main Methods:

• A Universal Perceptron (UP) with minimal hidden layers and neurons is proposed.
• A DNA-like learning algorithm, inspired by biological DNA sequences, is developed for rapid network training.
• Criteria for LSBF and PBF implementation are established, alongside the NLSD measure for BF complexity.

Main Results:

• The DNA-like learning algorithm demonstrates fast training speeds for implementing BFs, outperforming error-correction (EC) and backpropagation (BP) algorithms.
• The proposed UP and algorithm effectively handle LSBFs and PBFs using single-layer perceptrons (SLPs).
• The nonlinearly separable degree (NLSD) quantifies BF complexity, identifying PBFs as the most complex.

Conclusions:

• The Universal Perceptron (UP) combined with the DNA-like learning algorithm provides an efficient and robust method for implementing all Boolean functions (BFs).
• This approach offers significant advantages in speed and computational requirements compared to traditional algorithms.
• The derived synaptic weights and thresholds are directly applicable to designing cellular neural networks (CNNs), a novel computing paradigm.