Professor Brian Sierwerski presented Dr. Rachel Brewer’s thesis presentation at the Vanderbilt Seminar Series.
Neuromorphic computing seeks to emulate the brain’s information‑processing strategies, offering promising approaches for image‑processing, audio analysis, and optimization tasks. Assessing the robustness of such systems under radiation exposure is critical for deploying neuromorphic hardware in space, high‑energy, or defense environments. This resource investigates the impact of proton‑induced single‑event effects on image‑classification performance within a neuromorphic computing architecture.
The study centers on the IBM True North chip, a CMOS‑based spiking neural network that implements a 14‑layer convolutional architecture for MNIST digit classification. Each of the more than 4 000 cores contains local SRAM for weight storage, dedicated neuron logic, and a suite of controllers, schedulers, and routers that enable event‑driven compute‑in‑memory operations without a global clock. Input images are encoded as pulse trains, propagated through weighted summations and thresholding, and the highest‑magnitude output node determines the classified digit. Pre‑radiation testing yields a deterministic 98.82 % classification accuracy across 10 000 images, while controlled proton irradiation introduces single‑event upsets in memory and routing elements, allowing quantitative measurement of accuracy degradation.
By examining these results, learners can explain how radiation alters neuromorphic chip classification accuracy, compare pre‑ and post‑radiation performance metrics of the spiking neural network, and describe the key components of the IBM True North architecture—including distributed memory, event‑driven cores, and on‑chip communication infrastructure. This insight supports the design of radiation‑hardened neuromorphic systems and guides future research on resilient, low‑power image‑processing hardware for hostile environments.