ECE Professor and Research Team’s Paper Chosen for IEEE Top Picks in Test and Reliability 2023

Case where a Self-Driving Car mistakes a transporting truck (left) for a bird (right) due to soft errors in the hardware

Congratulations to Dr. Karthik Pattabiraman, Professor of Electrical and Computer Engineering at UBC, whose paper “Understanding Error Propagation in Deep-Learning Neural Networks (DNN) Accelerators and Applications” was chosen for IEEE Top Picks in Test and Reliability (TPTR) 2023!

The paper was initially published in 2017 with former ECE PhD student Guanpeng Li, now a faculty member at the University of Iowa, and with researchers from Nvidia and MIT.

Submitted publications for IEEE Top Picks in Test and Reliability were reviewed by a committee of renowned experts in the field, and papers were then shortlisted. The paper was chosen by a competitive process, including a 2-page submission highlighting the impact of the paper to be considered. 

The paper was presented at the IEEE Top Picks in Test and Reliability workshop on October 12-13, 2023, to present the most impactful publications in the past six years in the areas of VLSI (Very Large Scale Integration) test and reliability.  

Dr. Pattabiraman and his team’s paper examined Deep Neural Networks (DNNs) usage in safety-critical contexts such as Autonomous Vehicles (AVs) and robots and their reliability to understand the effect of soft errors on DNNs. Soft errors are typically caused as a result of high-energy particles striking electronic devices, causing them to malfunction. It was demonstrated through these researchers’ work that DNNs are not resilient to soft errors and that a single fault can cause safety violations in the operation of DNNs.

Dr. Pattabiraman and his team have continued working on the problem of ML resilience to soft errors and have recently published a paper on the effect of soft errors on Large Language Models (LLMs), “Resilience Assessment of Large Language Models under Transient Hardware Faults”.

Through Dr. Pattabiraman and his team’s work and observations, software-based techniques for DNN resilience have changed to account for soft errors!

For more information on Dr. Pattabiraman’s work and the Dependable Systems Lab at UBC, please visit HERE