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

Area of Science:

• Computer Vision
• Machine Learning
• Artificial Intelligence

Background:

• Unsupervised and generative methods are crucial for learning meaningful representations from data.
• Disentangling latent factors in generative models is key to understanding and controlling image attributes.
• Anomaly detection remains a challenging problem, especially in complex, high-dimensional datasets.

Purpose of the Study:

• To develop unsupervised generative methods for learning representations that discover and control image semantic attributes.
• To formally clarify the relationship between attribute control and latent factor disentanglement.
• To create effective anomaly detection systems leveraging learned representations.

Main Methods:

• Proposed a network architecture combining multiscale generative models with mutual information (MI) maximization for attribute control.
• Derived an analytical result (lemma 1) to distinguish attribute control (MI maximization) from latent space disentanglement (total correlation minimization).
• Developed anomaly detection systems utilizing the learned representations.

Main Results:

• Demonstrated that maximizing semantic attribute control encourages latent factor disentanglement.
• Empirically showed superior performance in image generation quality (Fréchet Inception Distance) and disentanglement (mutual information gap) compared to state-of-the-art methods.
• Achieved performance benefits in anomaly detection compared to existing generative and discriminative algorithms.

Conclusions:

• The proposed unsupervised generative approach effectively learns representations for attribute control and disentanglement.
• These learned representations significantly enhance anomaly detection capabilities.
• The work has potential applications in addressing bias and privacy in AI within computer vision.