Encoding integers and rationals on neuromorphic computers using virtual neuron.

Area of Science:

• Neuroscience
• Computer Science
• Energy-Efficient Computing

Background:

• Neuromorphic computers mimic the human brain for power-efficient computation, primarily in spiking neural networks.
• Despite theoretical Turing completeness, efficient data encoding remains a major challenge for general-purpose computation on neuromorphic hardware.
• Existing encoding methods like binning and rate-based encoding are insufficient for broad computational applications.

Purpose of the Study:

• To introduce a novel virtual neuron abstraction for efficient encoding and addition of integers and rational numbers.
• To enable general-purpose computation on neuromorphic platforms by overcoming current data encoding limitations.
• To evaluate the performance and energy efficiency of this new encoding mechanism.

Main Methods:

• Developed a virtual neuron abstraction using spiking neural network primitives.
• Implemented and evaluated the virtual neuron for encoding and adding numbers.
• Tested performance on both physical and simulated neuromorphic hardware, including memristor-based processors.
• Demonstrated utility within μ-recursive functions, fundamental to general-purpose computation.

Main Results:

• The virtual neuron abstraction enables efficient encoding and addition of integers and rational numbers.
• Addition operations are estimated to consume as little as 23 nJ on average using a memristor-based neuromorphic processor.
• The method was successfully integrated into μ-recursive functions, showcasing its potential for general-purpose computation.

Conclusions:

• The virtual neuron abstraction is a viable solution for efficient number encoding on neuromorphic computers.
• This approach significantly advances the potential of neuromorphic systems for energy-efficient general-purpose computing.
• Further research can leverage this mechanism to unlock the full capabilities of brain-inspired computing architectures.