AI 中文总结
该研究通过测量扩散模型生成过程中的committor概率，发现模型在约两倍于细粒度问题的噪声水平下解决粗粒度问题，并提出一种更便宜的测量方法以正确排序决策顺序。
AI 中文摘要
扩散模型从纯噪声开始，逐步去除噪声。在生成过程中的某个时刻，它不再能够生成“任何东西”，而是致力于生成，比如说，一匹马而不是一辆卡车。我们在CIFAR-10上测量了这一时刻的发生时间，以及训练好的模型对此的误解。最直接的测量方法是冻结一幅半成品图像，从该点多次重新启动生成过程，并统计每个类别出现的频率。我们将这个概率称为committor。通过这种方式测量，模型解决粗粒度问题（车辆还是动物？）的噪声水平大约是解决细粒度问题（哪种动物？）的两倍。一种更便宜的测量方法，即分类器对两个类别的意见分裂成两个不同群体时的噪声水平，能够正确判断这些决策的顺序（秩相关0.73-0.88），但无法准确判断其具体时间。然后，我们将预训练模型与其
英文摘要
A diffusion model starts from pure noise and removes it step by step. Somewhere along the way it stops being able to become "anything" and becomes committed to, say, a horse rather than a truck. We measure when this happens, and what a trained model gets wrong about it, on CIFAR-10. The most direct measurement is to freeze a half-finished image, restart the generation from that point many times, and count how often each class comes out. We call this probability the committor. Measured this way, the model settles coarse questions (vehicle or animal?) at roughly twice the noise level of fine ones (which animal?). A much cheaper measurement, the noise level at which a classifier's opinion about two classes splits into two distinct groups, gets the order of these decisions right (rank correlation 0.73-0.88) but not their exact timing. We then compare pretrained models with their
