Racing to learn: statistical inference and learning in a single spiking neuron with adaptive kernels

Area of Science:

• Computational Neuroscience
• Neuromorphic Engineering
• Machine Learning

Background:

• Existing neuron models often involve complex mathematical operations.
• The computational role of dynamic synapto-dendritic kernels at the single neuron level is underexplored.
• Robustness to noise and efficient implementation are key challenges in neuromorphic computing.

Purpose of the Study:

• To introduce and describe the Synapto-dendritic Kernel Adapting Neuron (SKAN) model.
• To investigate the computational power of dynamic synapto-dendritic kernels in a simple spiking neuron.
• To demonstrate the suitability of SKAN for implementation in neuromorphic hardware.

Main Methods:

• Developed a novel spiking neuron model (SKAN) based on simple additive and binary processes.
• Analyzed the neuron's ability to perform statistical inference and unsupervised learning of spatiotemporal spike patterns.
• Implemented SKAN on a Field Programmable Gate Array (FPGA) to demonstrate hardware feasibility.

Main Results:

• SKAN performs statistical inference and unsupervised learning of spatiotemporal spike patterns without complex mathematical operations.
• The model exhibits robustness to signal and parameter noise, utilizing noise in its operations.
• Network-level simulations show competitive dynamics where faster neurons mask learned patterns.

Conclusions:

• SKAN is a computationally powerful, simple, and noise-robust spiking neuron model.
• Its design makes it highly suitable for efficient implementation in digital and analog neuromorphic systems.
• The model opens new avenues for understanding single-neuron computation and developing advanced AI hardware.