Collaborative Research: SHF: Medium: Verifying Deep Neural Networks with Spintronic Probabilistic Computers

Northwestern University, Evanston IL

Investigators

Abstract

This project addresses the critical need for verifying the safety and reliability of deep neural networks (DNNs) used in various applications, such as autonomous driving, aircraft control, and consumer products like smartphones. As the demand for computational power in artificial intelligence continues to grow, this project explores innovative domain-specific architectures (DSAs) such as quantum and probabilistic computing platforms as potential solutions to the verification problem. The research's significance lies in ensuring the correct functioning of DNNs when exposed to perturbations or attacks, with the goal of benefiting society by enhancing the safety and trustworthiness of Artificial Intelligence (AI)-driven technologies. This interdisciplinary project will not only advance the field of DNN verification and energy-efficient domain-specific computers but also support education and workforce development, increasing collaboration between academia and industry through targeted activities. The project aims to solve the exact DNN verification problem using quantum annealing and probabilistic computers, contrasting conventional classical computing approaches. The research contributions span device, circuit, system, and algorithm levels. At the device level, the project will improve the energy efficiency of existing probabilistic-bit (p-bit) designs by leveraging the voltage-controlled-magnetic anisotropy (VCMA) phenomenon. At the circuit and architecture level, the team will design array-level spintronic p-computers (i.e., computers powered by p-bits) and investigate the dynamics between the feedback circuitry and p-bits, complemented by scaled complementary metal-oxide semiconductor (CMOS) emulators (Field Programmable Gate Arrays - FPGAs) with more than 1000 p-bits. At the algorithm level, the project will focus on formulating the exact verification of a neural network as an Ising model problem, which will be solved using the developed hybrid classical/probabilistic computers. The project will create pathways for large-scale spintronic probabilistic computers and explore new research directions, such as applying the simulated quantum annealing algorithm for DNN verification. This work will lay the foundations for p-computers with more than one million p-bits, enabled by Magnetic Random Access Memory (MRAM) technology with far-reaching applications beyond verification, including machine learning and quantum simulation. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.