Avesta Sasan
University of California-Davis, CA·Engineering
Machine learning models called transformers now drive language tools and much of computer vision, but they burn a lot of energy. The models themselves are complex, and today's chips are not built to execute them well.
- Fields
Engineering
Computer Science
- Topics
vision transformers
transformer models
attention mechanism
- Grants
- 1 known grant, through Sep 2027
What they work on
- Build vision transformer models that sample an image incrementally across encoder layers, then use that accumulating context to drop unattended tokens and mask unimportant image patches in later samples.
- Test learned, context-aware coarse-grain pruning of transformers through attention-head dropping, encoder-layer skipping and early termination, measuring efficiency gains against accuracy loss.
- Architect a stochastic pre-processing unit that approximates matrix-matrix multiplication to drive the lightweight classifiers that decide which patches, tokens, attention heads or encoders to eliminate.
- Design multiply-and-accumulate (MAC) units using temporal carry-bit deferment, which avoids carry propagation, and exploit their shallow logic to build diffusible MACs with dynamic precision trade-offs.
- Develop a scheduler for sparse attention graphs that balances load and minimizes memory accesses, supporting out-of-order token processing, processing-element clustering and precision-aware scheduling.
More about the research
This project pursues efficiency from two directions at once: redesigning the model so it skips work it does not need, and designing custom hardware that performs the remaining work faster and with less power.
Why it matters
Transformer models are accurate but expensive to run, which keeps them out of power- and cost-constrained settings. Pruning that is guided by what the model actually attends to, paired with accelerator hardware built for it, could cut the energy and time per inference. That would make strong vision and language models more widely usable, and the tunable-precision circuit ideas could inform future accelerator design more broadly.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Yuxiong Wang
University of Illinois Urbana-Champaign, IL·Computer Science
Develop a multi-resolution video encoding strategy that adjusts token granularity to balance visual detail against efficiency in transformer-based vision-language models.
Zhuoran Yang
Yale University, CT·Mathematics & Statistics
Test whether large language models carry out in-context learning and chain-of-thought prompting as implicit Bayesian inference, and analyze how transformer attention, the mechanism that weighs relevant words, could approximate these estimators.
Omiya Hassan
Boise State University, ID·Engineering
Co-design an electro-optical application-specific integrated circuit that pairs photonic (light-based) and electronic processing units, assigning AI training and inference to the best-suited unit.
The school
University of California-Davis: 49 labs on PI Pages, $1.0B a year on research, 647 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.