Jianxin Wu (吴建鑫) Professor (中文简介) School of Artificial Intelligence Nanjing University Office: Room 1018, Computer Science Building I am in the LAMDA group. Email: For research related matters (paper, code, review, etc.): wujx2001 {AT} gmail.com Email: For teaching related matters (teaching, admission, hiring, etc.): wujx2001 {AT} nju.edu.cn My email box will treat 126, 163, 189, QQ, Sina, Yeah email addresses as spam. Emails from these addresses will not be read by me. 请申请进组的本校本科生或申请读研（直博）的同学首先阅读这个文档。 怎样从别人那里要求获得代码/论文/模型/数据?

Essentials of Pattern Recognition: An Accessible Approach is now available!

《模式识别》教材中文版现已出版！

Education

Career

Services

Publications

Teaching

Students

Research

• CV & ML with limited computing resources
• Finetuning large deep models
• DTL: How to save 2/3 GPU memory in parameter-efficient finetuning?: Paper [C70]
• Memory- and parameter- efficient finetuning (extension of DTL): Paper[J57]
• How to save communication (transferring data between cloud and edge) and fine-tune on edge?: Paper [C79]
• Model Quantization
• PTQ quantizing all sorts of networks: fast & accurate, simple & general, on GPUs & simulators---QwT (Quantization without Tears): Paper [C78]
• Quantizing large language models (LLM): Paper [C77]
• Quantization for self-supervised learning: Paper [C62]
• To make quantization work for ViT detection etc. by KD from quantized penultimate layer: Paper [C65]
• Deep network compression, acceleration, and generation
• Compressing Diffusion models to 1 bit storage: Paper [C80]
• Teacher-free KD (SSL models like DINO as teacher--no teacher training on downstream task): Paper [C76]
• Zero-/Few-shot dense compression of Vision Transforms: Paper [C71]
• Few-shot network compression---the Practise algorithm: Paper [C67] and an interesting theory to support it: Paper [J54]
• A unified compression scheme for Vision Transformers and variants: Paper [J49]
• Few-shot, unlabeled compression of (all) linear layers (in particular in Transformers) with compression and acceleration but small accuracy drop: Paper [C66]
• Compressing recommender systems (embeddings+model): faster, smaller, and better: Paper [C73]
• Versatile, Full-spectrum and Swift network generation: Paper [J48]
• Feature mimicking with few-samples: Paper [C60]
• Feature mimicking: a new knowledge distillation paradigm: Paper [J46]
• CURL: Network compression with only small dataset and/or residual connections: Paper [C51]
• AutoPruner: Variable ratio channel pruning: Paper [J42]
• Channel pruning based on activation approximation: Papers [C41], [J37]
• A simple acceleration trick for detection using deep networks: Paper [C42]
• Deep learning beyond CNN/RNN/Attention
• NRS: Neural random subspace: Paper [J45]
• RSS: NRS for learning with missing data: Paper [J51]
• CV & ML with limited data resources
• Few-shot deep learning
• A new & practical few-shot learning problem (pFSL): Paper: [C72]
• Worst-case accuracy more important and difficult than average accuracy in few-shot learning: Paper [C63]
• Deep learning with only very few samples (training from scratch)
• Training vision Transformers with only 2040 images (IDMM self-supervised learning): Paper [C61]
• Weakly supervised localization & detection
• Weakly supervised object localization (WSOL): Paper [C52] (proposing a paradigm shift for WSOL)
• Object co-localization using deep models: Papers [C39], [J38]
• Weakly supervised object detection (WSOD): Papers [C59] (Salvage of Supervision, or SoS)
• T-PAMI extension of SoS (see the above item) for WSOL, WSOD, WSIS (weakly supervised instance segmentation) and WSSS (weakly supervised semantic segmentation): Papers [J52]
• Weakly supervised foreground learning for both WSOL and WSOD: Paper [J50]
• Learning with zero, partial, incomplete, noisy, and weak labels
• Multi-Label SSL: coarse cropping & turn it into object-centric SSL. Papers [J56]
• Multi-Label SSL: self-supervised learning on scene images. Papers [C69]
• ActionFormer: one-shot, anchor-free action localization using Transformers. Papers [C64]
• Tobias: A random network (without any training) can localize objects! Papers [C58], the theoretical underpinning of Tobias [J53]
• Theoretical results and practical algorithm for semi-supervised deep learning: Paper [C50]
• End-to-end deep learning in the presence of noisy labels: Paper [C48]
• Using weak labels for recognition: Papers [C38], [J32], [C46], CSRA [C57] (simple but powerful multi-label recognition)
• Fine-grained classification and retrieval without using bounding box annotations: Papers [J22], [J31], New datasets [C55]
• Dealing with imbalanced & long-tailed data distribution
• Reviving undersampling for imbalanced learning. PR paper
• Long-tailed recognition with the help of synthetic (diffusion-generated) data: NeurIPS'24 paper
• Improving long-tailed detection by rectifying the regression bias: ECCV'24 paper
• Improving the worst instead of average accuracy in long-tailed recognition: CVPR'23 paper
• Long-tailed recognition: AAAI'21 paper (bag of tricks for long-tailed recognition), ICCV'21 paper (DiVE: balanced virtual example distribution)
• Learning with imbalanced datasets: Papers [C5], [C12], [J6]
• Imbalance in face detection: Papers [C4], [J4], [J5]
• Multi-instance learning
• Scalable MIL: Papers [C33], [J21]
• MIL with multi-view: Paper [C34]
• Earlier work & other work
• CV & ML with limited computing resources
• Kernel approximation (in SVM and beyond): Papers [C10], [C17], [J10], [J19], [C25]
• Cascade structured classifier and detector: Papers [C3], [C4], [J4], [J5]
• Creating visual codebooks using additive kernels: Papers [C8], [J8], [J17]
• High-dimensional visual features and their compact representations: Papers [C28], [C37], [J24]
• Real time object detection based on HIK: Papers [C10], [C13], [J12]
• Detection and recognition using sensors beyond camera (RFID, mobile sensor, etc.), and beyond the computer (robot, mobile phone etc.): Papers [C6], [C11],[C15]
• Visual representation based on the Census Transform: Papers [C7], [C8], [J7], [J14]
• Actions: Papers [J11], [J13], [C30], [C32] (physics based modeling), [J20] (good practices), [J27] (from single image)
• Visual Place Categorization, mapping, and navigation: Papers [C9], [C29], my Ph.D. dissertation, [J16]
• Find the appropriate level of sparsity: [C14]
• Ensemble learning: [C1], [C2], [J2] (many could be better than all)
• (Very) early work on faces: Papers [J1], [J3], [J18]