Senior Research Scientist — Efficient & Scalable Vision
Seungwoo Yoo
Current focus: Sub-10ms BEV perception backbone on an Orin-class SoC with INT4 weights / FP8 activations.
Background
PhD UC Berkeley BAIR 2021, advised by Trevor Darrell and Kurt Keutzer; thesis on mixed-precision quantization for vision transformers on automotive SoCs. Member of MLPerf Inference WG (2022–present). Staff Engineer at Tesla Autopilot (2021–2024) where he ported HydraNets to FP8 on HW4 and authored the open-source `torch-quant-vit` library.
Education
- PhD · EECS — UC Berkeley (2021), advisor: Trevor Darrell
- BS · Computer Science — Carnegie Mellon University (2015)
Selected publications
- Yoo et al., 'Q-ViT: Fully Quantized Vision Transformers for Edge Inference', ICCV 2021
- Yoo & Darrell, 'Hardware-Aware Neural Architecture Search for Driving Perception', ECCV 2022
- Yoo et al., 'INT4 Multi-Task Driving Networks on Embedded SoCs', MLSys 2024
Reading list
- Howard et al., 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications', arXiv 2017
- Tan & Le, 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks', ICML 2019
- Touvron et al., 'Training data-efficient image transformers & distillation through attention (DeiT)', ICML 2021
- Dettmers et al., 'GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers', ICLR 2023
- Liu et al., 'EfficientViT: Memory-Efficient Vision Transformer with Cascaded Group Attention', CVPR 2023
Lead experiments
All experiments →No experiments led by this member yet.