1 Title
Your Diffusion Model is Secretly a Zero-Shot Classifier(Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, Deepak Pathak)【ICCV 2023】
2 Conclusion
This paper shows that the density estimates from large-scale text-to-image diffusion models like Stable Diffusion can be leveraged to perform zero-shot classifi-cation without any additional training.
3 Good Sentences
1、Obtaining a diffusion model classifier through Bayes' theorem consists of repeatedly adding noise and computing a Monte Carlo estimate of the expected noise reconstruction losses (also called -prediction loss) for every class. We call this approach Diffusion Classifier.(The theory of this study that use diffusion model to make classification)
2、Discriminative approaches directly learn tomodel the decision boundary of the underlying task, while generative approaches learn to model the distribution of the data and then address the underlying task as a maximum likelihood estimation problem. (The principle of the generate mode that used in discrimination)
3、We split our evaluation into a series of stages, where in each stage we try each remaining ci some number of times and then remove the ones that have the highest average error. This allows us to efficiently eliminate classes that are almost certainly not the final output and allocate more compute to reasonable classes.(The improvement of this study for Efficient Classification)
对于像 Stable Diffusion 这种类型的 diffusion models,主要的步骤有两个,其一是 sampling,其二是 density estimation。而第二点又分为两种,unconditional density estimation 和 conditional density estimation,前者估计,后者估计。
本文认为类似stable diffusion
这样的大规模text2img模型所计算出的密度估计,可以被用来进行"零样本分类" (zero-shot classification),而不需要额外的训练。 也就是在大规模Text2Img任务中density estimation 这件事情几乎等价于 zero-shot classification without training,于是作者们将这一分类机制单独提炼出来,形成了 Diffusion Classifier 模型,并展示了这一模型有着很强的 multi-modal reasoning 的能力,它可以从含分类的 diffusion models 中提取出标准的分类器。
如何将diffusion model应用到zero-shot classification,具体流程图如上图所示:
对于一个分类模型,给定输入x,模型输出类别的概率向量c,对于这个diffusion model,分类任务就是求解。具体推导过程请看论文,这里不多赘述。
作者对比同为zero-shot classifier的CLIP,zero-shot的能力以及接近了基于renset50的CLIP。但与openCLIP ViT-H/14还有较大差距