Sumit Jha
University of Florida, FL·Computer Science
Artificial intelligence uses large amounts of energy because standard computers constantly move data between memory and processor. In-memory computing reduces this by doing calculations directly inside the memory hardware.
- Fields
Computer Science
Engineering
- Topics
formal methods
formal verification
in-memory computing
- Grants
- 1 known grant, through Aug 2027
What they work on
- Develop formal verification methods that prove neural networks run correctly when accelerated on analog in-memory computing hardware, which can be more error-prone than standard CMOS chips.
- Synthesize memory crossbars for neural networks using decision diagrams, neural networks and reinforcement learning, producing hybrid analog-digital in-memory circuit designs.
- Verify bidirectional digital in-memory computing circuits and then demonstrate the resulting systems through fabrication.
More about the research
The newer memory technologies that make this possible can be more error-prone than conventional CMOS chips, the standard silicon technology. That raises a question: how can we be sure a neural network running on such hardware gives correct answers?
Why it matters
As AI models grow larger, the energy needed to train and run them keeps rising. Methods that prove neural networks work correctly on error-prone in-memory hardware could let chip designers use these lower-power technologies safely. This could help deploy reliable AI with lower power consumption.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Nathaniel Cady
University at Albany, NY·Engineering
Evaluate the commercial translation potential of an enhanced fluorescence-based diagnostic platform for Lyme disease, which detects antibodies and other biomarkers in under 40 minutes, through first-hand investigation of the diagnostics industry ecosystem.
Abhijit Chatterjee
Georgia Institute of Technology, GA·Engineering
Develop post-manufacture test and tuning algorithms for deep neural networks running on analog crossbar arrays, recovering accuracy lost to manufacturing variation in seconds rather than hours and detecting shorts and opens.
Deliang Fan
Arizona State University Campus Immersion, AZ·Engineering·Early-career
Design, fabricate and validate a hybrid in-memory computing chip that combines RRAM, a non-volatile resistive memory, with SRAM to run deep neural networks with high energy efficiency.
The school
University of Florida: 65 labs on PI Pages, $1.3B a year on research, 795 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.