Efficient AI Algorithms on-Device
Efficient model design for always-on and interactive AR/VR: efficient multimodal foundation models, contextualized AI, neural architecture search, visual memory, and foveated perception.
Ziyun Li
Efficient AI for smart glasses and edge systems
Passionate AI System Researcher, Meta Reality Labs
I work on efficient on-device AI for AR/VR, smart glasses and robots, spanning vision foundation models, contextualized AI applications, foveated and behavior-driven perception, and efficient models and deployment on edge systems. My research connects AI algorithms, hardware, and full-stack systems, from model design and deployment to custom sensing and computing silicon.
Research Areas
Efficient AI HW and System
Full-stack AI acceleration through advanced technologies across sensors, processors, and memory systems: low-power vision SoCs, eMRAM/RRAM/NVM, compute-in-memory, and ASIC-aware algorithm co-design.
Selected Papers
UTOPI: Efficient Egocentric Long-Video Understanding in AR via User-Guided Token Pre-Compression
NeurIPS, 2026
Motion- and gaze-guided token pre-compression for efficient egocentric long-video understanding on AR devices.
AdaDino: Adaptive Vision Foundation Models for Edge Intelligence via LLM-Guided Execution
arXiv, 2026
LLM-guided routing for vision foundation models under edge latency and power constraints.
EgoEverything: A Benchmark for Human Behavior Inspired Long Context Egocentric Video Understanding in AR Environment
ECCV, 2026
A benchmark for long-context egocentric video understanding in AR, using human attention signals to build behavior-inspired questions.
Foveated Instance Segmentation
CVPR, 2025
A gaze-aware segmentation strategy that focuses computation on the instance of interest.
H4H: Hybrid Convolution-Transformer Architecture Search for NPU-CIM Heterogeneous Systems for AR/VR Applications
ASP-DAC, 2025
A search workflow for hybrid CNN/ViT models targeting heterogeneous NPU and compute-in-memory systems.
RoboVisio: A Micro-Robot Vision Domain-Specific SoC for Autonomous Navigation Enabling Fully-on-Chip Intelligence via 2-MB eMRAM
IEEE Journal of Solid-State Circuits, 2024
A micro-robot vision SoC with on-chip intelligence and eMRAM for autonomous navigation workloads.
EyeCoD: Eye Tracking System Acceleration via FlatCam-based Algorithm and Accelerator Co-design
ISCA, 2022; IEEE Micro Top Picks
First Place, University Demo Best Demonstration at DAC 2022A FlatCam-based eye tracking algorithm and accelerator co-design for compact, high-throughput AR/VR systems.
SplitNets: Designing Neural Architectures for Efficient Distributed Computing on Head-Mounted Systems
CVPR, 2022
Neural architecture design for distributing computation across sensors and aggregators on head-mounted systems.
RRAM-DNN: An RRAM and Model-Compression Empowered All-Weights-on-Chip DNN Accelerator
IEEE Journal of Solid-State Circuits, 2020
A DNN accelerator using RRAM and model compression to keep weights on chip and reduce memory energy.
An 879 GOPS 243 mW 80 fps VGA Fully Visual CNN-SLAM Processor for Wide-Range Autonomous Exploration
ISSCC, 2019
A low-power CNN-SLAM processor for real-time visual navigation and wide-range exploration.
A Low-Power 6D Vision Processor for Unified Optical Flow and Stereo Depth
IEEE JSSC, 2019; VLSI/ISSCC-era processor work; SiPS algorithm thread
First Place, Bob Owens Best Student Paper Award at SiPS 2016A unified optical-flow and stereo-depth vision pipeline and processor lineage for low-power autonomous navigation.
Stereo-Depth Processor for Energy-Efficient Autonomous Navigation of Micro Aerial Vehicles
ISSCC, 2017; IEEE JSSC invited version
A stereo-depth processor delivering real-time high-resolution depth for energy-efficient micro aerial vehicle navigation.
Education
University of Michigan, Ann Arbor
PhD in Electrical and Computer Engineering, 2019
Thesis: Energy-Efficient, Mobile Computer Vision and Machine Learning Processors
Doctoral research mentors: Hun-Seok Kim, Dennis Sylvester, and David Blaauw.
BS in Electrical Engineering, 2014
Work Experience
Aug 2019-present
Senior Staff Research Scientist, Meta Reality Labs
Leading research and technology investigations for efficient, low-power, low-latency vision processing and contextualized AI applications for smart glasses and AR/VR systems.
May 2018-Aug 2018
Research Intern, Facebook Reality Labs
Worked on efficient on-device machine learning for VR/AR platforms.
Oct 2016-Jan 2017
Research Intern, NVIDIA
Worked on agile design methodology and sparse machine-learning processor architectures.
Service
Workshop and Session Chairing
- Chairing ISCAS workshop/session activity on efficient contextualized AI for smart glasses.
- Chairing DAC Efficient ML sessions.
- Chairing ICCAD Efficient Training & Inference sessions.
Program Committee and Workshop Service
- MLsys On-device AI Workshop Committee.
- ICCAD Efficient Training & Inference Technical Program Committee.
- DAC Efficient ML Technical Program Committee.