Robust parallel decision-making in neural circuits with nonlinear inhibition.

Area of Science:

• Neuroscience
• Computational Neuroscience
• Machine Learning

Background:

• The brain performs essential max-finding computations for decision-making and action selection.
• Parallel processing in neural networks is powerful, but achieving optimal speed for max-finding remains a challenge.
• Conventional winner-take-all (WTA) networks struggle with noisy inputs and large numbers of options.

Purpose of the Study:

• To investigate if neural networks can perform max-finding operations efficiently, meeting the N-fold parallelism benchmark.
• To address the limitations of conventional WTA networks in handling noise and large-scale computations.
• To introduce a novel network architecture that overcomes existing challenges in neural max-finding.

Main Methods:

• Introduction of the nWTA (noisy WTA) network architecture with a second nonlinearity.
• Analysis of the nWTA network's performance against the parallelism benchmark for noisy candidate options.
• Evaluation of the nWTA network's ability to reproduce experimental phenomena like Hick's law.

Main Results:

• The nWTA network achieves the N-fold parallelism benchmark without parameter rescaling for varying N.
• Unlike conventional WTA networks, the nWTA network reliably produces a winner even with large N and noisy inputs.
• The nWTA network reproduces Hick's law without requiring additional readout stages or adaptive thresholds.

Conclusions:

• The nWTA network demonstrates that efficient, parallel max-finding is achievable in networks of noisy neurons.
• Cellular nonlinearities can be linked to circuit-level decision-making, bridging scales in neural computation.
• Hick's law may reflect near-optimal parallel decision-making in the presence of noisy neural inputs.