2025年1月11日 · Here, we demonstrate in the materials science setting that heuristic evaluations lead to biased conclusions of ML generalizability and benefits of neural scaling, through evaluations of...

The OODDB (OOD Detection Benchmark) is a comprehensive benchmark suite designed to evaluate machine learning models performing Out-Of-Distribution Detection, with a specific focus on its semantic aspect (a.k.a. Semantic Novelty Detection).

In this paper, we propose a simple yet effective framework, Prototypical Outlier Proxy (POP), which introduces virtual OOD prototypes to reshape the decision boundaries between ID and OOD data. Specifically, we transform the learnable classifier into a fixed one and augment it with a set of prototypical weight vectors.

2022年1月11日 · ID指的是in-distribution数据，也就是我们熟悉的训练数据；OOD指的是out-of-distribution，在不同的领域也可能被叫做outlier或者是anomaly data，说的是与ID分布不一致的数据。其实ID和OOD的界定比较模糊，通常我们是将语意信息相差较大的两个数据集构 …

2024年11月27日 · In this work, we analyze OOD detection task from the information theoretical perspective, and show that the close set classification proxy for representation learning may cause over-confidence on known classes and undesired compression of …

2024年9月23日 · We combine the feature vectors from the training, validation, and OOD data into a single dataset. UMAP (Uniform Manifold Approximation and Projection) is used to reduce the dimensionality of the feature vectors from the high-dimensional space to 2D, making it possible to visualize the relationships between different data points.

Motivation: HypStructure with Mahalanobis score leads to improved OOD detection. Main Theorem: Existence of eigenvalue gaps between each level of hierarchy for CPCC-based representations.

UMAP的核心与t-SNE非常相似，两者都使用图形布局（graph layout）算法在低维空间中排列数据。简单来说，UMAP首先构建数据的高维图表示，然后优化低维图以使其在结构上尽可能相似。虽然UMAP用于构建高维图的数学是复杂的，但背后的想法是非常简单的。

2023年6月14日 · 在计算机视觉领域，目标检测任务一直是研究的热点。然而，大多数现有的目标检测方法在面对 开放环境中的未知类别（Out-Of-Distribution, OOD）时性能表现不佳。本文将探讨OOD目标检测的背景、研究现状、挑战和未来。 2. 背景

2024年1月8日 · umap是一种基于概率流线的方法，用于将高维数据映射到低维空间。 它的核心思想是通过将数据点视为流线的节点，然后计算流线之间的欧氏距离来实现数据的可视化。