Local Intrinsic Dimension

Local intrinsic dimension (LID) estimation aims to determine the effective dimensionality of data points within their local neighborhood, revealing the underlying structure of high-dimensional datasets. Recent research focuses on leveraging deep generative models, particularly diffusion models, to improve LID estimation accuracy and scalability, addressing limitations of older non-parametric methods. These advancements are proving valuable in diverse applications, including improving neural architecture search efficiency, detecting anomalies and adversarial examples, and assessing the complexity of data such as images. The development of more robust and efficient LID estimation techniques is crucial for advancing various machine learning tasks and data analysis methods.