Harish Haresamudram
Portrait of Harish Haresamudram

Harish Haresamudram

Postdoctoral Research Associate · University of Illinois Urbana-Champaign

I develop machine learning methods that learn from everyday wearable sensors, such as accelerometers, PPG, and ECG, with minimal human supervision, and use them to understand human activity, behavior, and health.

About

I am a Postdoctoral Research Associate in the Grainger College of Engineering at the University of Illinois Urbana-Champaign, working with Prof. Jim Rehg. I received my PhD in Electrical and Computer Engineering from the Georgia Institute of Technology, where I was advised by Prof. Thomas Plötz and Prof. Irfan Essa. My dissertation, Learning Representations for Sensor-Based Human Activity Recognition for Challenging Application Scenarios, developed self-supervised and multimodal representation learning methods for wearable sensors.

During my PhD I was a research intern at Microsoft Research and Facebook Reality Labs. My work has been supported by the National Science Foundation (through the AI-CARING Institute), Optum AI, and Google.

News

  1. New preprint: Variable Input Masked Autoencoders for Wrist-Worn Accelerometry, a single 3.8M-parameter foundation model that transfers across sampling rates and input durations.

  2. New preprint: TEX-MEX, a contrastive learning objective for time series that unifies structural and temporal similarity, with state-of-the-art results across five sensing modalities.

  3. Models Got Talent was accepted as a brief at ISWC 2026!

  4. New preprint: How Well Do Multimodal Models Reason on ECG Signals? A reproducible framework for evaluating the reasoning of multimodal LLMs on ECG.

  5. New preprint: Models Got Talent, on predicting HAR model performance without any training.

  6. Honored to be runner-up for the Gaetano Borriello Outstanding Student Award at UbiComp 2025.

  7. Co-organized the GenAI4HS workshop at UbiComp 2025, to a packed room.

  8. New preprint: Leveraging Shared Prototypes for a Multimodal Pulse Motion Foundation Model, a foundation model for PPG and accelerometer data built on shared prototypes.

  9. Joined Prof. Jim Rehg’s lab at UIUC as a Postdoctoral Research Associate.

  10. Successfully defended my PhD dissertation at Georgia Tech!

  11. Our surveying tutorial on sensor-based HAR was published in IMWUT.

  12. Natural-language supervision for HAR accepted to AAAI 2025.

  13. Organized the second SOAR tutorial at UbiComp 2024 in Melbourne.

  14. Discrete representations for HAR published in Sensors.

  15. Organized the SOAR tutorial at UbiComp 2023 in Cancún.

Research

Wearables such as smartwatches and phones continuously record how we move and how our bodies respond, but turning those signals into reliable insight is hard: labels are scarce and noisy, devices and people vary widely, and the data rarely looks like what models were trained on. My research builds representation learning methods that address these challenges. I build foundation models of sensor data and develop approaches to enable longitudinal health monitoring.

  1. Self-supervised learning for wearable sensors

    Labeled sensor data is expensive to collect, while unlabeled data from phones and watches is abundant. I design pretext tasks and training recipes that learn general-purpose representations of human movement from unlabeled data, and study when and why they work.

    CPC (IMWUT ’21) · Assessing SSL for HAR (IMWUT ’22) · Discrete representations (Sensors ’24) · How much unlabeled data? (ISWC ’23)

    All papers in this theme →

  2. Sensors and language: multimodal foundation models

    Multimodal pre-training has transformed vision and language. I study how to bring these advances to sensor data: why natural-language supervision breaks down for wearables and how to fix it, how text descriptions let models transfer across environments, and whether multimodal LLMs actually reason about physiological signals.

    Language supervision for HAR (AAAI ’25) · TDOST (IMWUT ’25) · ECG reasoning (’26) · LLMs memorize sensor data (’24)

    All papers in this theme →

  3. Generalizing across sensing settings

    Models must work outside the lab, with new people, devices, and settings. I work on transfer to new domains with few labels, efficient model selection, and generating training data from video, with applications to workplace stress and eating behavior.

    Cross-domain HAR (TIST ’25) · Models Got Talent (ISWC ’26) · Stress (ACII-W ’23) · IMUTube (IMWUT ’20)

Publications

Google Scholar ↗