Public Lecture #5

"Hardware Design and the Fairness of A Neural Network"

Prof. Yiyu SHI (University of Notre Dame)

Abstract

Ensuring the fairness of neural networks is crucial when applying deep learning techniques to critical applications like medical diagnosis and vital signal monitoring. However, maintaining fairness becomes increasingly challenging when deploying these models on platforms with limited hardware resources. Unfortunately, most existing fairness-aware neural network designs overlook the impact of resource constraints, which can significantly affect their performances. In this talk, I will use designing neural network accelerators on compute-in-memory architecture as an example to analyze the impact of the underlying hardware on the task of pursuing fairness. I will first discuss the relationship between hardware platforms and fairness-aware neural network design. I will then discuss how hardware advancements in emerging computing-in-memory devices, in terms of on-chip memory capacity and device variation management, affect neural network fairness. I will also identify challenges in designing fairness-aware neural networks on such resource-constrained hardware. Finally, I will suggest future research directions in hardware, software, and cross-layer co-design to address these challenges.