Deliang Fan
Arizona State University Campus Immersion, AZ·Engineering·Early-career
Artificial intelligence, especially deep neural networks (DNNs), needs huge amounts of computation and memory. Today's chips waste much of their energy shuttling data back and forth between memory and processor.
- Fields
Engineering
Computer Science
Materials Science
- Topics
in-memory computing
processing-in-memory
deep neural networks
- Grants
- 3 known grants, through Oct 2027
What they work on
- 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.
- Develop dynamic neural networks and on-device learning methods for the RRAM-SRAM chip that adapt latency and power at run time and resist hardware and adversarial noise.
- Build an end-to-end design (CAD) tool that trains, optimizes, maps and evaluates neural networks on the in-memory hardware to meet goals for energy, speed, accuracy and robustness.
- Help design and tape out an AI processing-in-memory chip using two-terminal SOT-MRAM, a fast magnetic memory, for digital in-memory sparse multiply-and-accumulate operations in training and inference.
- Develop situation-aware continual learning and network architecture search algorithms that limit costly on-chip memory writes on the SOT-MRAM processing-in-memory system.
- Model adversarial weight-duplication fault attacks in multi-tenant cloud FPGAs, and develop ML algorithms and FPGA tamper-resistant defenses that protect model parameters in transit.
More about the research
This bottleneck is called the memory wall, and it adds to the power limits of silicon devices. In-memory computing addresses it by doing calculations inside the memory itself. The lab works across levels, from devices and circuits to chip architecture and algorithms, to build energy-efficient AI hardware that can keep learning on the device. A related project protects machine learning running on shared cloud hardware from hardware-based attacks.
Why it matters
Faster, far more energy-efficient AI chips could benefit computer vision, autonomous driving, robotics, personalized speech and smart connected health, especially in devices that must learn on the spot. Defenses against hardware fault attacks could help cloud providers and their users run machine learning safely on shared FPGAs.
Known grants
From NIH and NSF research grants. Labs often have other funding we can't see.
Collaborative Research: FuSe: Efficient Situation-Aware AI Processing in Advanced 2-Terminal SOT-MRAM
NSF · Oct 2024 – Oct 2027
Collaborative Research: SaTC: CORE: Small: Secure and Robust Machine Learning in Multi-Tenant Cloud FPGA
NSF · Oct 2024 – Jul 2027
CAREER: Efficient, Dynamic, Robust, and On-Device Continual Deep Learning with Non-Volatile Memory based In-Memory Computing System
NSF · Oct 2024 – May 2027
Closest labs
Shan Wang
Stanford University, CA·Engineering
Improve the thermal robustness, switching speed, endurance and energy efficiency of MRAM, a memory that stores bits in nanoscale magnets, by tackling nanofabrication and design challenges for mobile AI.
Shaahin Angizi
New Jersey Institute of Technology, NJ·Engineering·Early-career
Design reconfigurable circuits and architectures for a multi-modal in-sensor accelerator that moves AI computation onto sensor nodes as efficiently as possible.
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.
The school
Arizona State University Campus Immersion: 36 labs on PI Pages, $1.0B a year on research, 625 research doctorates a year.
Reach out
A start for your email. Fill in the brackets.