Neural Classifiers with Limited Connectivity and Recurrent Readouts.

Area of Science:

• Computational neuroscience
• Artificial neural networks
• Machine learning

Background:

• Neural network models often face limitations in classifying inputs due to sparse connectivity, posing a challenge for brain-inspired architectures.
• Existing solutions like committee machines shift connectivity demands to downstream neurons, creating a new bottleneck.

Purpose of the Study:

• To propose a novel neural network architecture that overcomes connectivity limitations in pattern classification.
• To demonstrate how recurrent networks can achieve scalable performance with sparse connectivity, mirroring biological neural circuits.

Main Methods:

• Analytical proof of performance scaling in a recurrent attractor neural network model.
• Modeling readout mechanisms using interconnected perceptrons within a recurrent framework.

Main Results:

• The number of classifiable random patterns grows unboundedly with the number of perceptrons, irrespective of finite individual perceptron connectivity.
• Both recurrent and downstream readout connectivity remain finite, avoiding the bottlenecks of feedforward models.

Conclusions:

• Recurrent networks with sparse connectivity can effectively replace feedforward classifiers with extensive long-range connections.
• This approach offers a scalable network design strategy that better reflects the recurrent nature of biological neural circuits.