Unsupervised Discovery, Control, and Disentanglement of Semantic Attributes With Applications to Anomaly Detection

摘要

Our work focuses on unsupervised and generative methods that addressthe following goals: (1) learning unsupervised generative representationsthat discover latent factors controlling image semantic attributes,(2) studying how this ability to control attributes formally relates to theissue of latent factor disentanglement, clarifying related but dissimilarconcepts that had been confounded in the past, and (3) developinganomaly detection methods that leverage representations learned in thefirst goal. For goal 1, we propose a network architecture that exploits thecombination of multiscale generative models with mutual information(MI) maximization. For goal 2, we derive an analytical result, lemma 1,that brings clarity to two related but distinct concepts: the ability of generativenetworks to control semantic attributes of images they generate,resulting from MI maximization, and the ability to disentangle latentspace representations, obtained via total correlation minimization. Morespecifically, we demonstrate that maximizing semantic attribute controlencourages disentanglement of latent factors. Using lemma 1 and adoptingMIin our loss function, we then show empirically that for image generationtasks, the proposed approach exhibits superior performance asmeasured in the quality and disentanglement of the generated imageswhen compared to other state-of-the-art methods, with quality assessedvia the Fréchet inception distance (FID) and disentanglement via mutualinformation gap. For goal 3, we design several systems for anomaly detection exploiting representations learned in goal 1 and demonstratetheir performance benefits when compared to state-of-the-art generativeand discriminative algorithms. Our contributions in representationlearning have potential applications in addressing other important problemsin computer vision, such as bias and privacy in AI.