Reliable Deep Learning
Confidence calibration, refinement, uncertainty estimation, and OOD detection.
Senior Scientist, TCS Research · Doctoral Researcher, IIT Delhi
Ramya Hebbalaguppe
Reliable machine learning for deep neural networks and vision-language models — spanning confidence calibration, out-of-distribution detection, refinement, distillation, and test-time adaptation.
I am a Senior Scientist at TCS Research, working in the Deep Learning and Artificial Intelligence Group at the Research and Innovation Park, IIT Delhi. My doctoral research, supervised by Prof. Chetan Arora at IIT Delhi, focuses on improving the reliability of deep neural network classifiers.
A central theme of my research is that accuracy alone is not sufficient for deployment in safety-critical settings. I develop methods that make models not only accurate, but also aware of their uncertainty, robust under distribution shift, and better calibrated for real-world decision making.
Before my doctoral research, I worked with Prof. Ramakrishna Kakarala at Nanyang Technological University on HDR imaging for smartphone camera pipelines. This work received the Best Student Paper Award at SPIE 2012. I completed my Master’s degree at Dublin City University, advised by Prof. Noel O'Connor and Prof. Alan Smeaton, where I worked on reducing false alarms in surveillance camera networks.
Research Themes
Reliability as a first-class design principle rather than a post-hoc correction.
Lightweight Reliability
Knowledge distillation, compression, and calibrated student models.
Vision-Language Models
Prompt calibration, VQA reliability, and test-time adaptation.
Creative & Immersive AI
Generative image editing, motion transfer, and 3D/4D vision.
Selected News
Recent publications and research updates.
- 2026ICLR Trustworthy AI: Enhancing DNN Reliability with Refinement and Calibration.
- 2026ECML PKDD : Prompting without Panic.
- 2025WACV: Curve Skeletonization in Continuous Domain.
- 2025ECML PKDD: Better Features, Better Calibration.
- 2024ACCV Oral: Calibration Transfer via Knowledge Distillation.
- 2022CVPR Oral: A Stitch in Time Saves Nine.
Selected Publications
Representative work across reliability, calibration, VLMs, and visual computing.
Refine and Align: Confidence Calibration Through Multi-Agent Interaction in VQA
A debate-based multi-agent framework where specialized VLMs answer and generalist agents critique and aggregate responses to produce better-calibrated confidence estimates for VQA.
Enhancing Deep Neural Network Reliability with Refinement and Calibration
A unified reliability view connecting refinement and calibration, showing how representation learning can improve both discriminative capability in classifiers and predictive confidence quality.
Prompting without Panic: Attribute-aware, Zero-shot, Test-Time Calibration
An attribute-aware zero-shot test-time calibration approach for vision-language models, improving confidence reliability without retraining on task-specific labeled data.
Curve Skeletonization in Continuous Domain for Meshes and Point Clouds
A continuous-domain approach to curve skeletonization for meshes and point clouds, designed to handle geometric structure across diverse 3D inputs.
Better Features, Better Calibration: A Simple Fix for Overconfident Networks
A representation-focused calibration method that addresses overconfidence by improving feature quality rather than only reshaping logits after training.
Calibration Transfer via Knowledge Distillation
Knowledge Distillation for Calibration (KD(C)) extends distillation beyond accuracy transfer to produce lightweight, accurate, and well-calibrated models.
Lifelong Learning in StyleGAN through Latent Subspaces
StyleCL enables lifelong learning in StyleGAN through task-specific latent subspace dictionaries and lightweight feature adaptors while reducing catastrophic forgetting.
LoMOE: Localized Multi-Object Editing via Multi-Diffusion
A zero-shot multi-diffusion framework for localized multi-object editing, supporting additions, replacements, and edits in a single pass.
ReMOVE: A Reference-free Metric for Object Erasure
A reference-free metric for assessing object erasure in diffusion-based image editing, especially where no clean reference image exists.
Transfer4D: A Framework for Frugal Motion Capture and Deformation Transfer
A frugal alternative to expensive motion-capture pipelines, transferring motion from incomplete single-view depth video to semantically similar target meshes.
Calibrating Deep Neural Networks Using Explicit Regularisation and Dynamic Data Pruning
A differentiable calibration loss and dynamic data pruning strategy to improve calibration while reducing training effort.
A Novel Data Augmentation Technique for Out-of-Distribution Sample Detection using Compounded Corruptions
Compounded Corruptions (CnC) synthesizes OOD-like samples without hold-out OOD data and improves OOD detection accuracy and inference-time behavior.
A Stitch in Time Saves Nine: A Train-Time Regularizing Loss for Improved Neural Network Calibration
MDCA is a train-time multi-class calibration objective that can be combined with task losses across image, natural language classification, speech, and segmentation settings.
Current Research Team
Research and Innovation Park, IIT Delhi.
- Ayush Pandey — Calibration in Agentic AI and LLMs
Interns (Summer 2025): Arman Fatima (IIT Dharwad), Ishita Jain (GaTech), Khuushi Maheshwari (MIT).
Alumni
Researchers including full-time members, pre-doctoral fellows, and research interns.
- Srinidhi Hegde → Doctoral Student at Purdue University (Prof. Sotiris Nousias)
- Jai Bardhan → Researcher at the Czech Technical University in Prague (with Dr. Josef Sivic)
- Adarsh Kappiyath → Doctoral student at the University of Surrey, UK (with Dr. Lu Yin and Dr. Xiatian Zhu)
- Pranay Gupta → Doctoral student at CMU (with Prof. Henny Admoni)
- Meghal Dani → Doctoral student at IMPRS-IS, Max Planck School (with Prof. Zeynep Akata and Prof. Bernhard Schölkopf)
- Surabhi Nath → Doctoral student at the Max Planck School of Cognition and the MPI for Biological Cybernetics (with Prof. Peter Dayan)
- Goirik Chakrabarty → Doctoral student, University Göttingen (with Prof. Fabian Sinz)
- Jatin Prakash → Doctoral student at New York University (with Rajesh Ranganath)
- Neelabh Madan → Doctoral student at New York University
- Gaurav Gupta → Doctoral student at Rice University
- Apoorv Khattar → Doctoral student at University of Manchester, UK
- Neel Rakholia → Masters Student at Stanford
- Sharan Yalburgi → Visiting researcher at MIT probabilistic ML project
- Shubh Maheshwari → Graduate student at UCSD
- Jitender Maurya → Researcher, Toshiba
- Archie Gupta → SDE, Microsoft
- Varun Jain → Masters student at CMU → Microsoft Fellow
- Additya Popli → SDE at Google
- Kshitijz Jain → Grad student at IIT Delhi
- Aravind Udupa → Grad student at IIT Delhi
- Soumya Suvra Ghosal → Masters Student at University of Wisconsin
- Gaurav Garg → Accenture
- Ramakrishna Perla → TTEC Digital
Academia: Thesis Supervision
Student supervision and thesis mentoring.
- Aditya C (IISc Bangalore) — co-supervisor for M.Tech thesis on metrics for image editing; Best Thesis Award.
- Shreyash Mohatta (BITS Goa) — M.Tech thesis on egocentric real-time gesture recognition with Dr. Ashwin Srinivasan → Masters student at NCSU.
- Rishabh Patra (BITS Goa) — B.Tech thesis on uncertainty calibration with Dr. Tirtharaj Dash → SDE, Amazon.
- Ashwin Vaswani (BITS Goa) — BTP on data-free iterative knowledge distillation with Prof. Ashwin Srinivasan → Google Research → Masters student at CMU.
- Het Shah (BITS Goa) — BTP on knowledge distillation, pruning, and quantization with Prof. Ashwin Srinivasan → Research Associate at Google Research.