Natural gradient enables fast sampling in spiking neural networks

Area of Science:

• Computational Neuroscience
• Machine Learning

Background:

• Animals require uncertainty estimation for navigating dynamic environments.
• Implementing fast sampling algorithms in biologically plausible spiking neural networks remains a challenge.

Purpose of the Study:

• To propose a method for implementing fast sampling algorithms in spiking neural networks.
• To unify different classes of spiking samplers within a common framework.
• To demonstrate rapid parameter inference in high-dimensional distributions.

Main Methods:

• Leveraging population geometry, neural code, and neural dynamics.
• Simulating Langevin sampling with efficient balanced spiking networks.
• Implementing Metropolis-Hastings sampling with probabilistic spike rules.

Main Results:

• Unified two classes of spiking samplers (Langevin and Metropolis-Hastings).
• Demonstrated rapid inference of parameters from strongly-correlated high-dimensional distributions.
• Showcased the impact of population geometry on inference speed.

Conclusions:

• Population geometry is key for fast probabilistic inference in spiking neural networks.
• Provides design principles for sampling-based inference algorithms.
• Offers potential inspiration for neuromorphic computing and testable neurobiological predictions.