Vijay Sadashivaiah

I am a Senior Scientist in Machine Learning in the Data AI and Genomics Group at Merck Research Laboratories. My research focuses on pretraining large foundation models, optimizing training through efficient kernels, and evaluating model performance. My broader research interests include foundation models, deep learning, machine learning systems, training optimization, and explainable AI.

Email  /  CV  /  Resume  /  Google Scholar  /  Twitter  /  LinkedIn  /  GitHub

• 03/2025: Started as Senior Scientist in Machine Learning at Merck Research Laboratories
• 02/2025: Awarded the Robert McNaughton Prize for Best Computer Science Graduate Student at Rensselaer Polytechnic Institute
• 02/2025: Successfully defended my Ph.D. in Computer Science at Rensselaer Polytechnic Institute
• 05/2024: I returned as a Research Intern at Bosch Center for AI
• 04/2024: I passed my candidacy exam to become a Ph.D. candidate
• 04/2024: Our work on studing bias in vision language models was accepted to NAACL TrustNLP workshop
• 01/2024: Our work on suppressing user determined semantic concepts in representation transfer was accepted to TMLR
• 05/2023: Started as summer intern at Bosch Center for AI
• 01/2023: Our paper on spectral unmixing was published at BMC Neuroscience
• 11/2022: I was a finalist at Quad Fellowship
• 10/2022: Our work on Improving langugage model predictions using knowledge graphs was accepted to DL4KG Workshop at ISWC-22
• 07/2022: I was awarded a second best poster award at ISWS summer school in Bertinaro, Italy
• 04/2022: I was accepted to ISWS summer school in Bertinaro, Italy
• 02/2022: Our work on improving transfer learning using multi-armed bandits was accepted to ICLR-22
• 01/2022: Started my research co-advised by Professor James A. Hendler and Professor Chris R. Sims
• 05/2021: Started a summer project with Trustworthy AI and RL groups at IBM Research
• 01/2021: Started MS/PhD in Computer Science at Rensselaer Polytechnic Institute

We demonstrate that replacing softmax with sigmoid attention in biological foundation models yields 25% improved cell-type separation, up to 10% faster training, and improved stability. We develop TritonSigmoid, an efficient GPU kernel achieving 515 TFLOPS on H100s.

We present TEDDY, a family of transformer-based foundation models for single-cell biology, scaled to 116 million cells. The models (70M, 160M, and 400M parameters) can identify disease states and distinguish diseased from healthy cells, with performance improving predictably with data volume and parameter count.

We propose to explain chest X-ray pathology models using textual concepts. This is achieved by leveraging the joint latent space of image and text in vision-language models.

We evaluated the influence of bias on multimodal text generation models. In particular, we studied the impact of visual augmentation using state-of-the-art diffusion models when generating text.

We propose to suppress user-determined semantically meaningful concepts (viz. eyeglasses, smiling) from intermediate representations in computer vision tasks.

We introduce multi-armed bandit based representation routing to improve transfer learning in computer vision tasks.

We propose to imporve language model predictions by enriching the prompts from knowledge graphs.

An automated approach to spectral unmixing of fluorescent images

A single-nucleus RNA-sequencing resource of 70,615 high-quality nuclei to generate a molecular taxonomy of cell types across five human brain regions.

We constructed a mechanistic, stochastic and functional models of nerve fiber to quantify interactions.

We studied the sensory motor interactions in mice brain using voltage sensitive dye imaging.

© 2025 Vijay Sadashivaiah | Source code credit to Dr. Jon Barron