Mapping Generative Models onto a Network of Digital Spiking Neurons.

Area of Science:

• Artificial Intelligence
• Computational Neuroscience
• Hardware Engineering

Background:

• Stochastic neural networks like Restricted Boltzmann Machines (RBMs) excel in generative tasks but rely on computationally intensive Markov Chain Monte Carlo methods (e.g., Gibbs sampling).
• Neuromorphic systems offer low-power, parallel cognitive computing but lack robust applications and automation for complex algorithms.
• Bridging these fields is crucial for advancing AI hardware and applications.

Purpose of the Study:

• To develop a systematic method for mapping generative Restricted Boltzmann Machines (RBMs) onto digital neuromorphic systems.
• To demonstrate the feasibility of generative RBM inference on neuromorphic hardware using a pattern completion task.
• To address challenges in network connectivity, weight, and bias quantization for neuromorphic implementation.

Main Methods:

• Proposed a Gibbs sampler implementation using bio-inspired digital noisy integrate-and-fire neurons.
• Described the offline training and mapping process of generative RBMs onto the IBM TrueNorth neurosynaptic processor.
• Developed a design automation procedure for optimal resource utilization on neuromorphic VLSI substrates.

Main Results:

• Successfully implemented generative RBM inference on a neuromorphic VLSI substrate, a first of its kind.
• Analyzed generative performance to validate neuromorphic requirements and optimize neuron parameters.
• Demonstrated a pattern completion task as proof of concept for the proposed method.

Conclusions:

• This work establishes a viable pathway for deploying generative RBMs on low-power neuromorphic hardware.
• The proposed methods and design automation address key challenges in mapping complex AI algorithms to neuromorphic systems.
• This research paves the way for more efficient and powerful cognitive computing applications.