Improving Extrapolation Behavior in Deep Neural Networks

Executive Summary

We will develop capabilities to detect, control, and improve extrapolation and strong generalization behavior in deep neural networks. By enabling neural networks to extrapolate better, we will make deep neural networks better suited for a broad range of scientific explorations, transforming mission-critical research in areas such as biosecurity, physical data modeling, and cognitive simulation.

Publications, Presentations, and Patents

Narayanaswamy, V., Mubarka, Y., Anirudh, R., Rajan, D., Spanias, A. and Thiagarajan, J.J., “Know Your Space: Inlier and Outlier Construction for Calibrating Medical OOD Detectors.” Accepted for publication at Medical Imaging and Deep Learning (MIDL) 2023.

M.L.Olson,S. Liu, R. Anirudh, J.J. Thiagarajan, P.T. Bremer, and W.K. Wong, “Cross-GAN Auditing: Unsupervised Identification of Attribute Level Similarities and Differences between Pretrained Generative Models.” (Presentation and accepted for publication, IEEE/CVF Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, 2023).

S. Mitra, A. Shukla, R. Anirudh, J.J. Thiagarajan, and P. Turaga, “Adapting Blackbox Generative Models via Inversion” (Presentation, ICML Workshops on Challenges of Deploying Generative AI, Honolulu, HI, 2023). 

Thopalli, K., Devi, S., Thiagarajan, J. J., InterAug: A Tuning-Free Augmentation Policy for Data-Efficient and Robust Object Detection, IEEE Access (In press).

Anirudh, R., & Thiagarajan, J. J. (2023). "Single Model UQ estimation with DeltaUQ." (Version 0.1) [Computer software]. https://github.com/LLNL/DeltaUQ

Narayanaswamy, V. (2023), "Know Your Space: Inlier and Outlier Construction for Calibrating Medical OOD Detectors." (Version 0.1) [Computer Software] https://github.com/LLNL/OODmedic