Research
Research directions Visual perception of non-cooperative targets is our central task. Open-environment learning and on-device embodied intelligence provide supporting methods. Each direction lists selected publications and related foundational work.
Visual Perception of Non-Cooperative Targets Machine Learning in Open Environments On-Device Embodied Intelligence Visual Perception of Non-Cooperative Targets Target discovery, localization, and tracking in complex real-world environments.
Machine Learning in Open Environments Foundational learning methods that support reliable visual perception in open environments.
2.1 Out-of-Distribution DetectionFairness confusion can cause sensitive attributes and feature shifts to distort out-of-distribution decisions. Figure 1 from Hu et al., Neural Networks 205, 109564 (2027; online 2026); Long Lan is a coauthor. ↗ 2.2 Continual LearningSAIM reduces forgetting through sharpness-aware updates and adaptive isotropic merging across continual tasks. Method panel from the official ICML 2026 poster for Yang et al., Merge to Remember; Long Lan is a coauthor. Cropped. ↗ 2.3 Few-Shot LearningThe SHA-Pipeline captures and uses class hierarchy in large-scale few-shot learning. Figure 3 from Lin et al., Applied Sciences 14(16), 7060 (2024), CC BY 4.0. Resized; Long Lan is a coauthor. ↗ 2.4 Adversarial LearningBiDO and BiDO+ defend against model inversion by controlling dependencies while preserving classification utility. Figure 3 from Peng et al., TPAMI 47(8), 6382–6395 (2025); Long Lan is a coauthor. Cropped. ↗ 2.5 Learning with Noisy LabelsTNLPAD separates parameters associated with clean and mislabeled samples over the course of training. Figure 3 from Wang et al., TPAMI 46(9), 6341–6354 (2024); Long Lan is a coauthor. ↗ 2.6 Reinforcement LearningMulti-head attention filters shared observations for fault-tolerant multi-agent reinforcement learning. Figure 2 from Gu, Geng, and Lan, Entropy 23(9), 1133 (2021), CC BY 4.0. Resized; Long Lan is a coauthor. ↗ On-Device Embodied Intelligence Efficient models and active perception as enabling technologies for visual tasks.
3.1 Model Compression through Knowledge DistillationKnowledge distillation through complementary magnitude and diversity enhancement. Figure 2 from Li et al., Sensors 24(11), 3617 (2024), CC BY 4.0. Resized; Long Lan is a coauthor. ↗ 3.2 On-Device Generative ModelsJuZhou 1.0 combines Chinese prompt refinement, compact denoising, distillation, and a lightweight decoder for on-device generation. Figure 4 from Chen et al., JuZhou 1.0, arXiv:2606.28421 (2026); Long Lan is a coauthor. ↗ 3.3 Agent-Based Active PerceptionGeoEyes learns task-adaptive, coarse-to-fine visual exploration for evidence-grounded remote-sensing reasoning. Figure 2 from Wang et al., GeoEyes, arXiv:2602.14201 (2026); Long Lan is a coauthor. ↗ 3.4 Behavior Prediction and Self-Verification with World ModelsRelated foundations for sequential prediction and goal inference in multi-agent settings.
Mamba-CDSP models sequential observations and debiases state-space representations for counterfactual prediction. Figure 3 from Wang et al., ICLR 2025; Long Lan is a coauthor. Cropped from the author-hosted paper. ↗