Niharika S D'Souza

Niharika is a Research Scientist working at IBM Research, AI in the Silicon Valley (San Jose, CA) since January 2022. Her research interests span various domains such as high-dimensional and statistical representation learning, geometric deep learning , graph signal processing, computer vision, and learning on tabular, heterogenous, and multimodal data.

Currently, she is working on representation learning, optimisation, verification/evaluation, and benchmarking for Agentic DataOps.

In recent news:

• Serving as Area Chair for MIDL 2026
• Invited to give a talk on Maximal Correlation Multigraph Neural Networks for Multimodal Fusion at the inaugural NxtAI summit at San Francisco.
• New work on Automated Data Product Creation and Benchmarking using Large Language Models (LLMs) available. [Paper] [Code] [Data]
• Work on Semantic Alignment in Vision Language Transformer Models (VLMs) accepted into the workshop on Unifying Representations in Neural Models at NeurIPS 2025. SemCLIP was one of six top submissions selected as an oral spotlight.
• Work on Prompt Optimization without Task Cues and Instructions accepted into the IEEE International Conference on Data Mining- ICDM 2025, Demo Track.
• Work on Phrasally Grounded Fact Checking for VLM based Automated Radiology Report Generation accepted into MICCAI 2025
• Served as an Organizing Committee Member for the 6th workshop on GRaphs in biomedicAl Image anaLysis- GRAIL 2025
• Moderated a discussion on Scalable and Translatable Healthcare Solutions at the Conference on Health, Inference, and Learning CHIL 2025 as a Senior Roundtable Leader. [Tehcnical Report]
• Work on Prompt Optimization accepted to VLDB'25 workshop on LLM+KG
• Work on LLMs as Universal Tabular Representation Learners accepted to Transactions of Machine Learning Research ( TMLR ).
• Nominated for Full Membership into Sigma Xi, the Scientific Honor Society.
• Served as an Area Chair for MICCAI 2025 and MIDL 2025
• MMMG toolkit for multi-graph NNs released.
• Work on Encoded Representations and Modern Hopfield Networks was accepted into the workshop on Unifying Representations in Neural Models at NeurIPS 2024. One of six papers selected for an oral spotlight. This work was also selected as an oral presentation (top 13 percent of submission abstracts) at BayLearn 2024.
• Work on Geometrically Constrained U-Nets for segmentation in Radial Imaging modalities was presented at the Machine Learning with Medical Imaging workshop at MICCAI 2024. This was a joint collaboration between the MIT-IBM Watson Lab (with the Medical Vision group, CSAIL, EECS, MIT) and Boston Scientific and a part of her mentee, Yiming Chen's M. Eng. thesis. Achieved 200 percent improvement on state of the art at lower compute cost (GPUs) at +1000x reduction in model parameter cost.
• Recipient Outstanding Technical Achievement Award by IBM Research
• Served as an Area Chair for MICCAI 2024
• Served as an Organising Committee Member for the 6th workshop on GRaphs in biomedicAl Image anaLysis- GRAIL 2024
• Her work on Multiplexed Graph Neural Networks for multimodal fusion appeared as special issue published at MedIA and is now available online.
• Her team was recognised by with an A-Level Technical Accomplishment for fundamental advances to the science of multimodal fusion IBM Research.
• She was recognized as an Outstanding Reviewer (one among the top 12 reviewers) for MICCAI 2023
• She presented her work on Maximal Correlation informed Multi-Layered GNNs for Multimodal Fusion at the ML4MHD workshop at ICML 2023 as an oral.
• Co-wrote a book chapter on Network Comparisons and their applications in Connectomics appeared in Connectome Analysis: Characterization, Methods, and Analysis
• Served as a session chair for the session on Brain Connectomics at IPMI 2023
• Work on Geodesic Mean Estimation for Functional Connectomics manifolds at IPMI 2023 presented as an oral spotlight (top 20 percent of accepted papers)
• Work from 2022 on multiplexed graph neural networks for multimodal fusion was a finalist for the Young Scientist Award for MICCAI 2022 (top 4 percent of papers), and an NIH Travel Award.
• Named top 10 % of reviewers for ICML 2022. Invited as a session chair for session on Optimization.

Between 2016-2021, she obtained her doctoral degree from the Electrical and Computer Engineering at Johns Hopkins University under the supervision of Dr. Archana Venkataraman. In collaboration with researchers from the Malone Center for Engineering in Healthcare and Kennedy Krieger Institute, she developed a suite of mathematical models of brain and behavior spanning network optimization models, deep-generative hybrids, graph neural networks and manifold learning approaches for analyzing functional and structural connectomics data. Her research has been prominently featured in top tier conference venues such as MICCAI, IPMI, MIDL, and journals such as NeuroImage. She has also been the recipient of multiple awards and honours such as the MINDS Data Science Fellowship 2021 (JHU), Computing Research Association (CRA) Richard Tapia Scholarship , Rising Stars in Data Science 2021 (U. Chicago), 2021, Rising Stars in EECS 2020 (UC Berkeley), Best Paper Award (MLCN at MICCAI 2020), IPMI Scholarship For Junior Scientists (2019) and NIH student travel awards for MICCAI (2018, 2020, 2022).

For a complete list of publications, her google scholar profile can be found here.

She also holds a Masters Degree in Applied Mathematics and Statistics (Johns Hopkins University, 2019-2021) with a concentration in Optimisation, Statistics and Statistical Learning (GPA 3.95/4.00), and a Bachelor's Degree (with Honours) in Electrical Engineering (GPA: 9.17/10 Rank 5/120) along with a minor in Electronics and Electrical Communications Engineering (GPA: 9/10) from the Indian Institute of Technology, Kharagpur (2012-2016). During her undergraduate years, she worked with Dr. Debdoot Sheet on developing deep learning frameworks for deblurring and denoising Fluorescence Microscopy images.