Using Stochastic Spiking Neural Networks on SpiNNaker to Solve Constraint Satisfaction Problems

Area of Science:

• Computational neuroscience
• Artificial intelligence
• Neuromorphic computing

Background:

• Constraint satisfaction problems (CSPs) are computationally challenging, belonging to the NP-complete class.
• The human brain effectively solves CSPs using spiking neural networks (SNNs).
• Noise in SNNs can be harnessed as a computational resource for problem-solving.

Purpose of the Study:

• To develop a software framework for implementing noisy neural solvers on SpiNNaker hardware.
• To demonstrate the potential of SNNs with embedded noise for solving CSPs.
• To explore new optimization strategies and understand SNN computational capabilities.

Main Methods:

• Implementation of a software framework for noisy neural solvers on SpiNNaker.
• Utilizing SNNs as stochastic dynamical systems to solve CSPs.
• Demonstrating the framework on Sudoku, map coloring, and spin glass problems.

Main Results:

• The framework successfully implements stochastic search on SpiNNaker for solving CSPs.
• Noise in SNNs facilitates optimal exploration of configuration spaces.
• Discontinuous noise application enables system restarts for enhanced search.

Conclusions:

• Noisy neural solvers on SpiNNaker offer a viable approach for tackling complex CSPs.
• This framework advances the application of SNNs in computational problem-solving.
• The research opens avenues for exploring novel optimization techniques and SNN computational power.