Christopher Kanan
University of Rochester, NY·Computer Science·Early-career
Deep neural networks, the computer systems behind modern face and object recognition, can match people on some perception tasks. They still have major weaknesses, though.
- Fields
Computer Science
Neuroscience
Engineering
- Topics
continual learning
lifelong learning
catastrophic forgetting
- Grants
- 1 known grant, through Dec 2026
What they work on
- Develop brain-inspired continual learning algorithms that learn online and then run sleep-like consolidation periods, based on complementary learning systems theory, so deep neural networks gain new knowledge without forgetting old knowledge.
- Create new measures to test deep neural networks for dataset bias, check whether they acquire robust concepts, and quantify forward transfer when trained over time.
- Evaluate the new algorithms on large-scale incremental image classification, multimodal vision-and-language scene understanding, and abstract reasoning tasks, and release all code and datasets.
More about the research
When they learn new information over time, they tend to forget what they learned before. They learn slowly, needing to loop over the same data many times. They can also amplify biases in their training data, which leads to wrong predictions for groups that are poorly represented. This project takes inspiration from the mammalian brain, which stores memories during waking and strengthens them during sleep, to build networks that learn continually, efficiently and more fairly.
Why it matters
Machine learning systems that keep learning without forgetting and without retraining from scratch could use less power and run on low-powered phones and robots. Methods that detect and reduce dataset bias could make AI predictions more reliable for groups with less training data. Public code and datasets will give other researchers building blocks for their own algorithms and applications.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Rui Yu
University of Louisville, KY·Computer Science
Extract a labeled structural skeleton from single noisy images by fitting geometric primitives such as planes and cylinders to monocular depth estimates, using large vision-language models to assign meaning.
Paul Bogdan
University of Southern California, CA·Computer Science
Develop computational techniques that use multimodal neuroimaging and sensing to identify living neurons, non-neuronal cells such as glia, and the connections between them.
Tai Sing Lee
Carnegie Mellon University, PA·Neuroscience
Develop a machine learning framework that formalizes the rapid recurrent plasticity, or fast rewiring, of local circuits in the early visual cortex.
The school
University of Rochester: 38 labs on PI Pages, $492M a year on research, 308 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.