Learning to Parameterize a Stochastic Process Using Neuromorphic Data Generation

Deep learning is consistently becoming more integrated into scientific computing workflows. These high-performance methods allow for data-driven discoveries enabling, among other tasks, classification, feature extraction, and regression. In this paper, we present a unique approach to solving an inverse problem-determining the initial parameters of a system from observed data-using not only deep learning-powered AI but also simulation data generated using neuromorphic, brain-inspired hardware. We find this approach to be both scalable and energy efficient, capable of leveraging future advancements both in AI algorithms and neuromorphic hardware. Many high performing deep learning approaches require large amounts of training data. And, while great progress is being made in new techniques, current methods suggest that data-heavy approaches are still best-suited for maintaining critical generalization required for an inverse problem. However, that data comes at a cost, often in the form of expensive high-fidelity numerical simulations. Instead, we make use of recent advances in spiking neural networks and neural-inspired computing wherein we can use Intel's Loihi to compute hundreds of thousands of random walk trajectories. Statistics from these random walkers effectively simulate certain classes of physical processes. Moreover, the use of neuromorphic architectures allows these trajectories to be generated quickly and at drastically lower energy cost. This generated data can then be fed into a deep learning regression network, modified to incorporate certain known physical properties. We find the resulting networks can then determine the initial parameters and their uncertainties, and we explore various factors that impact their performance.