Department of Computational and Data Sciences
Department Seminar
==================================================================
Speaker : Dr. Vivek Sivaraman Narayanaswamy
Title : “Anchoring Deep Models for Reliable and Robust Machine Learning”
Date & Time: June 23rd, 2025 (Monday), 04:00 PM
Venue : # 102, CDS Seminar Hall
ABSTRACT
As machine learning and AI systems are increasingly deployed in high-stakes, real-world settings, ensuring their reliability under diverse conditions is more important than ever. Reliability spans multiple dimensions, including robustness to corruptions and distribution shifts, out-of-distribution (OOD) detection, calibration, generalization, extrapolation, uncertainty quantification and failure detection. Despite ongoing progress, achieving consistent performance across these axes remains a fundamental challenge.
This talk introduces anchoring, a simple yet effective training paradigm for deep networks. Anchoring reparameterizes each input as a tuple comprising a reference sample and a residual (the difference between the input and the reference), while preserving the original label. Both reference and input samples are drawn from the training distribution. Though lightweight in design, this modification enables the model to explore a richer hypothesis space, leading to better convergence and improvements across multiple reliability dimensions without requiring significant architectural changes. Importantly, anchoring is model-agnostic and applicable even to state-of-the-art vision transformers. The talk will present key insights about anchoring across tasks and modalities, demonstrate its consistent impact on reliability, and discuss its integration into standard training pipelines along with promising future directions.
BIO: Vivek Sivaraman Narayanaswamy is a Machine Learning Research Scientist with the Machine Intelligence Group at the Center for Applied Scientific Computing (CASC). His work lies at the intersection of deep learning, computer vision, and scientific machine learning with a focus on building trustworthy AI systems. He develops neural network surrogates for complex simulations, embedding uncertainty quantification to enhance reliability and designing protocols for better extrapolation, calibration, and anomaly detection. He earned his Ph.D. in Electrical Engineering from Arizona State University, USA in 2023, where his research centered on improving the fidelity and reliability of deep learning models. His work has been published at top-tier AI conferences such as NeurIPS, ICML, ECCV, AAAI, and ICASSP. He is also passionate about mentorship and has co-supervised student research on foundation models, conformal prediction, and scientific ML.
Host Faculty: Yogesh Simmhan
ALL ARE WELCOME
