竞争性约束如何揭示表征学习的规律
What competing constraints reveal about representation learning
AI 导读
一篇被 Patterns(Cell Press)接收的论文提出六个互补性质来指导从时序、多模态和部分观测数据中学习潜在表征:预测充分性、最小性、时间一致性、观测兼容性、干扰不变性和结构约束。受控合成实验显示,看似有利的诊断指标与表征实际保留的信息之间存在偏差,且同等预测性能下存在潜在坐标歧义。
正文 · 原文
Abstract:Learning latent states from temporal, multimodal, and partially observed data requires decisions about what information representations should retain, discard, and organize. I argue that six complementary properties help clarify these decisions: predictive sufficiency, minimality, temporal coherence, observation compatibility, nuisance invariance, and structural constraints. Drawing on established approaches, I distinguish desired representation properties from the objectives used to encourage them and the diagnostics used to assess them, since neither a training objective nor an evaluation metric guarantees that a property has been achieved. Controlled synthetic experiments examine how training objectives influence predictive performance, nuisance decodability, and latent organization. The results reveal discrepancies between seemingly favorable diagnostics and the information retained in learned representations, as well as latent-coordinate ambiguities under equivalent predictive performance, illustrating how complementary evaluations can improve interpretation and guide design choices. Together, these analyses provide a basis for designing, comparing, and evaluating learned latent states.
| Comments: | Accepted for publication in Patterns (Cell Press). Code repository: this https URL |
| Subjects: | Machine Learning (cs.LG); Artificial Intelligence (cs.AI) |
| Cite as: | arXiv:2605.15995 [cs.LG] |
| (or arXiv:2605.15995v5 [cs.LG] for this version) | |
| https://doi.org/10.48550/arXiv.2605.15995 arXiv-issued DOI via DataCite |
Submission history
From: Gwenole Quellec [view email]
[v1]
Fri, 15 May 2026 14:25:40 UTC (250 KB)
[v2]
Thu, 23 Jul 2026 16:07:36 UTC (547 KB)
[v3]
Wed, 7 Oct 2026 11:34:25 UTC (552 KB)
[v4]
Mon, 16 Nov 2026 15:33:07 UTC (552 KB)
[v5]
Wed, 18 Nov 2026 13:46:20 UTC (969 KB)
来源:arXiv cs.AI · arxiv.org