Researcher · ML & Bionics

Hello, I am Abhi

Portrait

Building Human-Centered AI with Multimodal Learning and Generative Models across Vision, Medical Imaging, Speech, Wearables, and Assistive Systems.

Across these areas, I’m motivated by a common question: how can machine learning systems understand people well enough to interact with them naturally and assist them effectively in the real world?

My recent work explores how machine learning can make sense of complex real-world data under three recurring constraints—noise, scale, and uncertainty—in ways that better support human interaction:
Speech and interaction. As an Applied ML Scientist at Ethosphere, I work on speech privacy for an AI-powered training assistant for retail employees.
Efficient learning from large-scale visual data. As a Postdoctoral Scholar with the Computational and Integrative Pathology Group at Northwestern University, I work on active learning and efficient adaptation of foundation models for classification and segmentation in histopathology, including gigapixel whole-slide images.
Generative modeling for medical imaging. At the University of Washington, I worked with KurtLab on generative methods for improving tumor segmentation in 3D brain MRI.

Multimodal sensing for assistive intelligence. A central theme of my PhD was understanding how machine learning and wearable sensing can be used to analyze, predict, and generate human movement in real-world environments. I completed my PhD at the bionics research lab ROMBOLABS at the University of Washington, where my research focused on making assistive devices more adaptive, expressive, and personalized. Here is my PhD work in a nutshell. Here is Everything (Almost) You Always Wanted to Know About Lower Limbs* (*But Were Afraid To Ask).

Embodied AI and robotics. My interest in this broader area began during my bachelor’s degree in Mechanical Engineering at IIT Bombay. Toward the end of my undergraduate studies, I became deeply interested in robotic hands and soft robotics, which led to my first research project: Design of a Soft Anthropomorphic Hand with Active Stiffness Control. Around that time, I was influenced by ideas from embodied intelligence, especially the work of Rodney Brooks and Rolf Pfeifer on how bodies shape behavior, cognition, and interaction with the world, including ideas related to morphological computation.

Understanding and predicting human movement through complex environments

Our environment shapes how we move through space. We have to alter our gait to avoid obstacles and move towards the intended goal. Humans use vision to plan their movement through space. So, can egocentric vision help predict their gait forward in time?

Complexity of human activities

Human bodies are capable of generating a diverse set of movements. We have some intuition about which activites are more complex than the others, for example walking forward in a straight line might be considered less complex than navigative obstacles. Then, there are mathematical notions of complexity for example dimensionality, variability etc. In this work, we explore whether different measures of complexity agree with our intuition.

Personalizing Assistive Devices

Everyone has a unique style with which they move. Then why should the assistive devices be generic?

Active Learning and High Performance Computing for Computational Pathology

Digital pathology has allowed improved data collection and labeling processes, as well as better data sharing. Consequently, there is a greater scope of using machine learning to build better tools for pathological diagnoses, and analyse digital pathology datasets. The challenges however include limited labeled data, fat-tailed distributions compared to natural domains, higher computational requirements due to large whole slide images, requirement of domain expertise for data labeling, as well as interobserver variability in data labeling. There is also a greater emphasis on developing human-in-loop solutions as compared to fully automated systems.

Generative Models for Medical Imaging

Medical imaging is crucial for early detection and prevention of fatal diseases. The datasets involved are hard to collect, and much smaller in quantity compared to most computer vision tasks. Can generative modeling techniques alleviate these challenges?

Skills

Machine Learning

Representation Learning, Recurrent Neural Networks, Convolutional Neural Networks, Autoencoders, Generative Adversarial Networks, Denoising Diffusion Probabilistic Models, Decision Trees, Style Transfer, Spectral Clustering, Semi-Supervised Learning

Bionics

Wearable Motion Capture, Gait Analysis, Biomechanics, Prosthetic Control.