A new synthesis approach for feedback neural networks based on the perceptron training algorithm

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Machine Learning

Background:

• Associative memories are crucial for information retrieval and pattern recognition.
• Designing feedback neural networks for associative memory poses significant challenges.
• Existing synthesis methods often lack guaranteed convergence or flexibility with constraints.

Purpose of the Study:

• To develop a new synthesis approach for associative memories based on perceptron training.
• To formulate the neural network design problem as linear inequalities solvable by perceptron training.
• To investigate and establish properties and existence results for networks with specific connection matrix constraints.

Main Methods:

• Formulation of the feedback neural network synthesis problem as a set of linear inequalities.
• Application of the perceptron training algorithm to solve these inequalities.
• Development of design algorithms for networks with sparsity and/or symmetry constraints on the connection matrix.

Main Results:

• Guaranteed convergence of perceptron training for unconstrained connection matrices.
• Established properties and existence conditions for networks with diagonal element constraints.
• Presented design algorithms for sparse and/or symmetric connection matrices.

Conclusions:

• The proposed perceptron-based synthesis approach offers a robust method for designing associative memories.
• The method is applicable to various feedback neural network models and handles diverse constraints.
• Demonstrated applicability and superiority over existing methods through specific examples.