The award-winning paper titled “Lightweight Model Uncertainty Estimation for Deep Neural Object Detection” was authored by Minah Lee, a Ph.D. candidate, Dr. Burhan Mudassar, a recent Ph.D. graduate from GREEN lab who is currently at Samsung, and Dr. Saibal Mukhopadhyay, a Joseph M. Pettit Professor in Tech’s School of Electrical and Computer Engineering.

The purpose of the paper is to quantify model uncertainty for real time reliable object detection. Model uncertainty helps to understand the reliability of a deep neural network (DNN), but its estimation usually requires a huge computational cost from stochastic DNN. This work introduced a deterministic DNN that distills the knowledge of stochastic DNN for probabilistic object detection and quantifies model uncertainty with 38-179 times reduced computational cost.

According to Mukhopadhyay, the paper is a great recognition of Lee’s PhD research on making machine learning more reliable and trustworthy while ensuring fast real-time inference. This paper on lightweight uncertainty estimation for deep neural network models is an important step towards that direction. The paper was based on the work supported by the Defense Advanced Research Projects Agency (DARPA) Reconfigurable Imaging (ReImagine) program.

More than 1,000 papers were presented at the IJCNN 2022, a premier — and one of the oldest — IEEE conferences in neural network research. This year, IJCNN was part of the IEEE World Congress in Computational Intelligence (WCCI 2022) held in Padova, Italy from July 18-23. The WCCI 2022 is the world’s largest technical event on computational intelligence, featuring the three flagship conferences of the IEEE Computational Intelligence Society (CIS) under one roof: IJCNN 2022, the 2022 IEEE International Conference on Fuzzy Systems, and the 2022 IEEE Congress on Evolutionary Computation.