ON-OFF neuromorphic ISING machines using Fowler-Nordheim annealers

Area of Science:

• Neuromorphic computing
• Quantum mechanics
• Computational neuroscience

Background:

• Ising problems are fundamental in statistical mechanics and computational optimization.
• Classical simulated annealing (SA) is a common heuristic for solving these problems, but can be slow and inefficient.
• Neuromorphic architectures offer potential for faster and more energy-efficient computation.

Purpose of the Study:

• To introduce NeuroSA, a neuromorphic architecture for efficient Ising problem solving.
• To leverage Fowler-Nordheim quantum mechanical tunneling for threshold-annealing.
• To demonstrate NeuroSA's ability to achieve asymptotic convergence to ground states.

Main Methods:

• Designed NeuroSA with asynchronous ON-OFF neurons to mimic simulated annealing dynamics.
• Implemented a Fowler-Nordheim (FN) annealer to adaptively adjust neuron thresholds.
• Mapped classical simulated annealing (SA) dynamics onto integrate-and-fire neurons.
• Validated NeuroSA on combinatorial optimization benchmarks (MAX-CUT, Max Independent Set).

Main Results:

• NeuroSA consistently produced solutions near or surpassing state-of-the-art for benchmark problems.
• Achieved superior performance on Max Independent Set benchmarks without graph-specific tuning.
• Demonstrated effective replication of SA's optimal escape and convergence, especially at low temperatures.
• Successfully mapped NeuroSA onto the SpiNNaker2 neuromorphic platform.

Conclusions:

• NeuroSA offers a powerful neuromorphic approach to solving Ising problems.
• The Fowler-Nordheim tunneling-based annealing process enables efficient convergence.
• NeuroSA's performance and platform compatibility highlight its practical potential for complex optimization tasks.