Gheorghi Guzun
San Jose State University, CA·Computer Science·Early-career
Deep neural networks, the AI models behind many image and language tools, are highly accurate but need power-hungry hardware to run. Quantization stores numbers with fewer bits to save computation, but fixed low-precision formats can cost accuracy.
- Fields
Computer Science
Engineering
- Topics
deep neural networks
quantization
sparsity
- Grants
- 1 known grant, through Feb 2028
What they work on
- Develop adaptable quantization and compression algorithms that take advantage of sparsity, meaning many zero values, in deep neural networks while keeping accuracy loss and overhead on non-sparse data low.
- Measure the tradeoff between accuracy and performance for the proposed methods, so systems can be tuned automatically to favor accuracy, speed or energy efficiency.
- Build a parallel prototype and an AI vector primitives library for GPUs and CPUs, with an interface that connects it to PyTorch and TensorFlow.
More about the research
Sparsity means much of the data is zero. This project asks how to use sparsity and flexible quantization to make AI faster and more energy-efficient without losing much accuracy.
Why it matters
More efficient AI could run on small, battery-limited devices such as wearables, smart sensors and self-driving vehicles. Running models directly on devices could reduce the load on cloud data centers and keep user data more private. The methods may also lower the power that data centers need for AI.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Closest labs
Miao Pan
University of Houston, TX·Computer Science
Design software-hardware co-designed model pruning and adaptive quantization schemes that exploit the sparsity of model updates to reduce computation and memory access during federated learning on a single 5G mobile device.
Hui Guan
University of Massachusetts, MA·Computer Science·Early-career
Build deep learning systems that adjust in real time which deep neural networks run, responding to task difficulty, changing request loads and resource sharing on edge servers and Internet of Things devices.
Antonio Mastropaolo
William & Mary, VA·Computer Science
Test whether filtering low-quality training examples with task-specific software engineering metrics lowers the computational cost of training large language models for code while preserving their learning ability.
The school
San Jose State University: 1 labs on PI Pages, $88M a year on research, 10 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.