Bharath Hariharan
Cornell University, NY·Computer Science·Early-career
Modern computer vision systems can recognize thousands of objects, but only after training on millions of images that people have labeled by hand. That approach breaks down in many settings.
- Fields
Computer Science
Engineering
Earth & Environmental Science
- Topics
computer vision
machine learning
artificial intelligence
- Grants
- 3 known grants, through Feb 2027
What they work on
- Build recognition systems that learn difficult visual concepts from as few as 1–2 labeled images and about 1,000 unlabeled ones, by drawing on a memory of past visual tasks across domains such as microscopy.
- Develop vision systems that learn by asking expert teachers detailed questions, rather than relying only on passively labeled images.
- Adapt perception systems in self-driving cars to each user's conditions by recording sensor data during use, generating pseudo-labels from physical constraints and cross-sensor agreement, and fine-tuning offline.
- Personalize perception through repeated routes by aligning recordings over time and propagating labels, and detect and remove noisy labels using learning dynamics and user verification, tested on driving datasets and home-robot navigation.
- Extend foundation models for satellite imagery to select among sparsely labeled satellite, drone and internet images, balancing cost against accuracy, and validate them on detecting submerged aquatic vegetation.
- Couple image models with self-improving large language models (LLMs) that generate and refine interpretable hypotheses about ecosystem change, such as environmental causes of invasive species spread.
More about the research
Labels may be expensive or require experts, the world a system faces may differ from the one it was trained on, or the data may be mostly unlabeled. This research builds visual recognition systems that learn from very little labeled data and adapt to new conditions. It applies them to areas ranging from microscopy and self-driving cars to monitoring the environment from space.
Why it matters
This work could make computer vision usable where labeled data are scarce. Examples include scientists analyzing microscopy images or tracking invasive species, and groups that cannot afford large annotation efforts. More adaptable perception could make self-driving cars and home robots more reliable in snow, ice or unfamiliar places. Satellite-based discovery of the causes of ecosystem change may help environmental scientists, and the under-resourced communities most exposed to environmental change and food insecurity, respond faster and at lower cost.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
CAREER: Bootstrapping Recognition from Little Data in New Domains
NSF · Mar 2022 – Feb 2027
Collaborative Research: ACED: Planet-scale AI for accelerating environmental science - Invasive species and beyond
NSF · Jun 2025 – Nov 2026
Collaborative Research: RI: Medium: Robust Perception through End-User Adaptation
NSF · Oct 2021 – Sep 2026
Closest labs
Carl Vondrick
Columbia University, NY·Computer Science·Early-career
Build open-world recognition models for satellite, aerial and ground imagery from vision-language models without language annotations, using cross-modal contrastive learning and reasoning about change over time.
Somil Bansal
Stanford University, CA·Engineering·Early-career
Develop training techniques that make machine-learning perception policies robust by construction, so camera-based controllers stay safe in new situations.
Noah Snavely
Cornell University, NY·Computer Science
Develop compositional implicit neural scene representations that pair a network's ability to capture fine detail with physical models of scene structure, such as light fields.
The school
Cornell University: 37 labs on PI Pages, $1.6B a year on research, 537 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.