Ph.D. Researcher · NUDT 博士研究生 · 国防科技大学
Continuous Bachelor–Master–Ph.D. Program 本硕博连读项目
College of Systems Engineering · Laboratory for Big Data and Decision 系统工程学院 · 大数据与决策实验室
I am a Ph.D. student in the continuous Bachelor–Master–Ph.D. program at the National University of Defense Technology. My research sits between self-supervised representation learning and efficient visual adaptation: I am interested in what a representation should preserve, how little supervision or adaptation is enough, and what changes when a network is built to respect the structure of its input.
我是国防科技大学本硕博连读项目的博士研究生。我的研究位于自监督表征学习与高效视觉适配之间：关注一个表征应当保留什么、多少监督或适配才算足够，以及当网络被构造成尊重输入结构时会发生什么。
Recent work covers rotation-stable 360° perception (EquiSSL), minimal contrastive objectives (CORE), semantic augmentation from foundation-model priors (SDI), the link between self-supervised representations and latent diffusion (CMAE), and asymmetric multi-modal fusion (HTNet, TriCM-Net). I review for NeurIPS, ICLR, ICML, AAAI, CVPR, IEEE TNNLS, Pattern Recognition and Neural Networks, and maintain the daily Visual SSL Digest.
近期工作包括旋转稳定的 360° 感知（EquiSSL）、极简对比学习目标（CORE）、由基础模型先验驱动的语义增广（SDI）、自监督表征与潜在扩散的衔接（CMAE），以及非对称的多模态融合（HTNet、TriCM-Net）。我担任 NeurIPS、ICLR、ICML、AAAI、CVPR、IEEE TNNLS、Pattern Recognition 与 Neural Networks 的审稿人，并维护每日更新的 Visual SSL Digest。
Our paper “Gauge-Equivariant Attention for Rotation-Stable 360° Scene Understanding” (EquiSSL) is accepted by ACM Transactions on Graphics (SIGGRAPH Asia 2026 Journal Track).
论文《Gauge-Equivariant Attention for Rotation-Stable 360° Scene Understanding》（EquiSSL）被 ACM Transactions on Graphics 接收（SIGGRAPH Asia 2026 期刊轨）。
First-author paper “Less Is More: Minimal Components for Contrastive Learning” (CORE) is published in IEEE TNNLS (CAS Tier 1 Top · CCF-B).
第一作者论文《Less Is More: Minimal Components for Contrastive Learning》（CORE）发表于 IEEE TNNLS （中科院一区 Top · CCF-B）。
First-author paper “DAGA: Dynamic Attention-Guided Adaptation for Self-Supervised Vision Transformers” is accepted to NeurIPS 2026. Project page ↗
第一作者论文《DAGA: Dynamic Attention-Guided Adaptation for Self-Supervised Vision Transformers》被 NeurIPS 2026 接收。项目主页 ↗
First-author paper “Semantic-Guided Data Augmentation for Self-Supervised Visual Pattern Learning” (SDI) is accepted by Pattern Recognition (CAS Tier 1 Top · CCF-B).
第一作者论文《Semantic-Guided Data Augmentation for Self-Supervised Visual Pattern Learning》（SDI）被 Pattern Recognition 接收（中科院一区 Top · CCF-B）。
Co-first author paper “CMAE: Contrastive Manifold Autoencoder for Efficient Latent Diffusion”, with Yishan Li, is accepted for oral presentation at IEEE ICME 2026.
与 Yishan Li 合作的共同第一作者论文《CMAE: Contrastive Manifold Autoencoder for Efficient Latent Diffusion》被 IEEE ICME 2026 接收，作 口头报告。
ACM Transactions on Graphics, Vol. 45, No. 6, Article 242 · SIGGRAPH Asia 2026 Journal Track
TL;DR — Removes gauge dependence from a spherical transformer's attention with a parameter-free Reynolds average, cutting the 90° rotation drop from 53% to 0.8%.
一句话总结——以无参数的 Reynolds 平均消除球面 Transformer 注意力中的规范依赖，将 90° 旋转下的精度下降从 53% 降至 0.8%。
Advances in Neural Information Processing Systems (NeurIPS 2026)
TL;DR — A dynamic, parameter-efficient adapter that steers attention in frozen self-supervised ViTs towards the features a downstream task needs.
一句话总结——一种动态、参数高效的适配器，在冻结的自监督 ViT 中把注意力引导到下游任务真正需要的特征上。
IEEE Transactions on Neural Networks and Learning Systems, 2026
TL;DR — Shows mathematically that momentum encoders and deep projection heads are dispensable, cutting training compute by 42% while improving linear-probe accuracy.
一句话总结——从数学上说明动量编码器与深层投影头可以去掉，训练计算量减少 42%，线性探针精度反而提升。
Pattern Recognition, Vol. 160, 2026
TL;DR — Generates semantically consistent counterfactual views from diffusion foundation-model priors, avoiding the feature collapse that heavy hand-designed augmentation causes.
一句话总结——借助扩散基础模型先验生成语义一致的反事实视图，避免强力人工增广导致特征坍缩。
Ph.D. researcher, NUDT — continuous Bachelor–Master–Ph.D. program, College of Systems Engineering; research at the Laboratory for Big Data and Decision. 国防科技大学博士研究生——本硕博连读项目，就读于系统工程学院，研究工作依托大数据与决策实验室。
Reviewer — NeurIPS, ICLR, ICML, AAAI, CVPR; IEEE TNNLS, Pattern Recognition, Neural Networks. 审稿人——NeurIPS、ICLR、ICML、AAAI、CVPR；IEEE TNNLS、Pattern Recognition、Neural Networks。
Four national invention patents granted by CNIPA — self-supervised feature extraction, multi-scale contrastive enhancement, pedestrian trajectory prediction, and retinal image segmentation. 四项国家发明专利获国家知识产权局授权——自监督特征提取、多尺度对比增强、行人轨迹预测与视网膜图像分割。
Curator — the Visual SSL Digest, a daily digest of new work on self-supervised visual representation learning. 维护者——《Visual SSL Digest》，每日梳理自监督视觉表征学习的新工作。
Contact联系
For questions about a paper, a possible collaboration, or anything in the digest, email is the most reliable way to reach me.
关于论文、合作或日报的任何问题，邮件是最可靠的联系方式。