Yawei Li
Democratizing AI for pervasive applications.
School of EEE, NTU Singapore
50 Nanyang Avenue, Singapore 639798
– Think big, aim high;
– Start small, stay patient;
– Never stop, never rush
I am a Nanyang Assistant Professor at the School of Electrical and Electronic Engineering, Nanyang Technological University (NTU), Singapore, where I lead the Deep Efficiency Lab. Previously, I was a Lecturer and Established Researcher at Integrated Systems Laboratory, ETH Zürich, where I worked with Prof. Luca Benini on AI accelerators and foundation models. I got my PhD degree from the Computer Vision Lab at ETH Zürich, supervised by Prof. Luc Van Gool.
I also collaborated with Prof. Radu Timofte, Dr. Michele Magno, and Prof. Ming-Hsuan Yang. During my PhD, I was mentored by Babak Ehteshami Bejnordi, Tijmen Blankevoort, and Amir Habibian at Qualcomm and Rakesh Ranjan at Meta.
Research
My research focuses on AI efficiency: accelerating computation, reducing memory footprint, and mitigating power consumption in AI systems, across both cloud and edge environments. Efficiency is rarely won at a single layer; the largest gains come from evolving the algorithm and the hardware that runs it together.
What is the right computational mechanism for a given architecture? I have studied this for CNNs, graph neural networks, vision Transformers (LocalViT, GRL, SemanIR, VRT), and state-space models (MambaIR, FEMBA).
How much of a trained model is actually needed? My work spans weight pruning (group sparsity, random pruning, DHP), token pruning (FastVAR), quantization (SliM-LLM, OBR), tensor decomposition (learning filter basis), and low-rank adaptation (IntLoRA), applied to LLMs and diffusion models.
Once the model is fixed, the bottleneck moves to how it is mapped onto silicon. I work on serving modern models efficiently both on large clusters (FlatAttention) and on severely constrained edge devices (smart glasses, Retina, TinyTracker).
Open-source processor and AI accelerator architectures based on RISC-V, as well as EDA methodologies for hardware PPA (Power, Performance, Area) optimization and 3D integration. This is the newest direction in the lab, growing out of my work at ETH Zürich with Prof. Luca Benini's group.
Biosignals record physiological activity and underpin healthcare, disease monitoring, BCI, HCI and AR/VR. They demand accurate modelling under a tight on-device compute budget — here efficiency is not an optimization but a precondition. This is my most active recent direction (PhysioWave, LUNA, WaveFormer, FEMBA, CEReBrO).
News
| Nov 10, 2026 | PhysioWave-T is released on arxiv! PhysioWave-T is a token-efficient wavelet foundation model for biosignals. Congrats Yanlong! |
|---|---|
| Oct 15, 2026 | I will serve as an Area Chair for ICML 2027. |
| Sep 28, 2026 | PrismQuant is released on arxiv! In this work, we derive optimal null-space rotations for grouped quantizers. Congrats Yanlong! |
| Sep 25, 2026 | One work on floating-point microformat quantization and pruning for efficient MU-MIMO neural receivers is released on arxiv. |
| Sep 21, 2026 | Our paper “TinyMyo: A Tiny Foundation Model for Flexible EMG Signal Processing at the Edge” is accepted to IEEE Transactions on Biomedical Engineering (TBME)! Congrats Matteo! |
| Sep 21, 2026 | Our paper “SPEAR: A System for Post-Quantization Error-Adaptive Recovery Enabling Efficient Low-Bit LLM Serving” is accepted to EuroSys 2027! |
| Aug 15, 2026 | I am invited to serve as an Area Chair for ICLR 2027. |