DOODLER:根据编码器重建确定分布外可能性
DOODLER: Determining Out-Of-Distribution Likelihood from Encoder Reconstructions
摘要:
深度学习模型具有两个关键特征,这两个特征结合在一起,使它们在现实世界中的使用具有风险。第一,他们通常不会在他们接受训练的分布之外很好地泛化,第二,他们倾向于表现出自信的行为,无论他们是否产生有意义的输出。虽然深度学习拥有解决现实、高维问题的巨大能力,但这些特点使人们很难对它们在现实世界中的应用充满信心。为了克服这个困难,已经定义了分布外(OOD)检测的任务,以确定模型何时从其被训练以操作的分布外部接收输入。本文介绍和研究了一种新的方法,DOODLER,面向对象的检测,它直接利用的特点,导致其必要性。通过在与另一个深度学习模型相同的数据上训练变分自动编码器(VAE),VAE学习准确地重建分布内(ID)输入,但不重建OOD输入,这意味着其故障状态可用于执行OOD检测。与该领域的其他工作不同,DOODLER只需要对OOD数据集的存在进行非常弱的假设,从而允许更现实的应用。DOODLER还可以通过OOD可能性对输入图像进行逐像素分割,实验结果表明,它匹配或优于在相同约束条件下操作的方法。
英文摘要:
Deep Learning models possess two key traits that, in combination, make their use in the real world a risky prospect. One, they do not typically generalize well outside of the distribution for which they were trained, and two, they tend to exhibit confident behavior regardless of whether or not they are producing meaningful outputs. While Deep Learning possesses immense power to solve realistic, high-dimensional problems, these traits in concert make it difficult to have confidence in their real-world applications. To overcome this difficulty, the task of Out-Of-Distribution (OOD) Detection has been defined, to determine when a model has received an input from outside of the distribution for which it is trained to operate. This paper introduces and examines a novel methodology, DOODLER, for OOD Detection, which directly leverages the traits which result in its necessity. By training a Variational Auto-Encoder (VAE) on the same data as another Deep Learning model, the VAE learns to accurately reconstruct In-Distribution (ID) inputs, but not to reconstruct OOD inputs, meaning that its failure state can be used to perform OOD Detection. Unlike other work in the area, DOODLER requires only very weak assumptions about the existence of an OOD dataset, allowing for more realistic application. DOODLER also enables pixel-wise segmentations of input images by OOD likelihood, and experimental results show that it matches or outperforms methodologies that operate under the same constraints.
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。