通过基于置信度的过滤在扩散模型中探索潜在结构的涌现
From Seeds to Semantics: Measuring Semantic Accessibility in Deterministic Diffusion Models
AI总结：
本文通过基于置信度的过滤揭示扩散模型潜在空间中与类别相关的结构，展示了高置信度样本在类别分离性上的显著表现，并提出条件生成作为替代指导方法的应用。
摘要：
扩散模型依赖于高维潜在空间中的初始噪声种子，但尚不清楚该空间是否包含足够的结构来预测生成样本的属性，如其类别。在本工作中，我们通过预训练分类器对生成样本分配的置信度分数来研究潜在结构的涌现。我们发现，当考虑所有噪声现实化时，潜在空间似乎总体上是无结构的，但限制关注于产生高置信度样本的初始噪声种子时，会表现出明显的类别分离性。通过比较不同置信度噪声子集的类别可预测性，并检查潜在空间的类别分离性，我们发现了仅在基于置信度的过滤下才可观察到的与类别相关的潜在结构。作为实际意义，我们讨论了基于置信度的过滤如何使条件生成成为指导方法的一种替代方案。
英文摘要：
Diffusion models generate samples through a sequence of learned denoising steps, and recent work has studied how semantic structure appears along this sampling process. We study this question in deterministic samplers by measuring semantic accessibility: how much information about a final semantic property, such as an image class label or attribute, can be extracted from the seed and intermediate states along the trajectory that produces the sample. Using DDIM sampling, for which each initial noise seed determines a unique trajectory and final image, we train separate classifiers (probes) at several points along the trajectory to predict a semantic property of the final image. We measure how well such a property can be predicted from the state at a given point using top-1 accuracy and normalized mutual information. Across MNIST, Fashion-MNIST, CIFAR-10, and CelebA, class labels and image attributes can be predicted above chance from the initial noise seed, and along DDIM trajectories, this accessibility exceeds matched-noise forward baselines. We find that semantic accessibility is substantially higher along the trajectories whose final images are classified with high confidence than along those classified with low confidence. These measurements provide a quantitative view of when semantic properties can be recovered during deterministic generation.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
