Generalized neuron: feedforward and recurrent architectures.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Computational Neuroscience

Background:

• Multilayer perceptrons (MLPs) and recurrent neural networks are standard for complex tasks but demand significant computational resources.
• Resource constraints in hardware platforms limit the practical application of traditional neural networks.

Purpose of the Study:

• To introduce and evaluate a compact neural architecture, the generalized neuron (GN) and its recurrent form (RGN).
• To demonstrate the efficacy of GN and RGN in tasks like classification, function approximation, density estimation, and time series prediction.
• To highlight the suitability of GN and RGN for resource-constrained hardware.

Main Methods:

• The study proposes feedforward and recurrent forms of the generalized neuron (GN).
• Particle Swarm Optimization (PSO) is utilized as the training algorithm for both GN and RGN.
• The performance of GN and RGN is evaluated on classification, nonlinear function approximation, density estimation, and chaotic time series prediction tasks.

Main Results:

• Generalized Neuron (GN) and Recurrent Generalized Neuron (RGN) architectures demonstrate effective performance in classification and time series prediction.
• GN and RGN exhibit resilience to complex nonlinearities due to their dual aggregation and activation functions.
• The compact nature of GN and RGN, with fewer trainable parameters, leads to reduced memory and computational requirements.

Conclusions:

• Generalized Neuron (GN) and Recurrent Generalized Neuron (RGN) offer a computationally efficient alternative to traditional neural networks.
• The reduced resource demands make GN and RGN highly suitable for implementation on resource-constrained hardware.
• These compact neural architectures present an attractive option for fast and efficient computational tasks.