Learning and inference with correlated neural variability.

Area of Science:

• Computational Neuroscience
• Artificial Intelligence
• Deep Learning

Background:

• The brain's inherent noise suggests stochasticity is crucial for neural computation.
• Learning in spiking neural networks (SNN) under correlated noise remains a challenge.

Purpose of the Study:

• To develop a theory for gradient-based learning in SNN within a noise-driven regime.
• To introduce a novel deep learning architecture for SNN.

Main Methods:

• Proposed stochastic neural computing (SNC) theory using a moment closure approach.
• Introduced moment neural networks (MNN) generalizing rate-based networks to second-order moments.
• Demonstrated direct parameter transfer from MNN to SNN without fine-tuning.

Main Results:

• Trained MNNs capture realistic biological neuron firing statistics (rate distribution, Fano factors, weak correlations).
• Optimized mean firing rate and correlation structure enhance task accuracy and reduce prediction uncertainty.
• Achieved enhanced inference speed through joint manipulation of firing rate and correlation.

Conclusions:

• SNC framework provides insights into SNN uncertainty processing.
• Enables the construction of biologically plausible neural circuit models with correlated variability.
• Demonstrated practical application on Intel's Loihi neuromorphic hardware.