RISK（Recovering Intended-Feature Subspace with Knowledge-Free）

开发了具有无知识 (风险) 的预期特征子空间的恢复。假设由各种偏差引起的快捷特征是无意预测的，风险将其视为冗余特征。当深入研究较低的流形以消除冗余时，RISK揭示了具有预期特征的极低维度子空间可以稳健地表示高度偏向的数据集。实证结果表明，我们的模型可以始终如一地改善模型对分布外集的泛化，并实现了新的最先进的性能1

We develop a method for recovering the expected feature subspace based on the knowledge-free (Risk) framework. We assume that shortcut features induced by various biases are unintended for prediction, and Risk regards them as redundant features. When delving into lower-dimensional manifolds to eliminate such redundancy, RISK reveals that the extremely low-dimensional subspace containing expected features can robustly represent highly biased datasets. Empirical results show that our model consistently improves the generalization performance on out-of-distribution (OOD) sets and achieves new state-of-the-art (SOTA) performance¹.

搜集汇总

背景与挑战

背景概述

RISK是一种无知识方法，旨在恢复预期特征子空间，通过将偏差特征视为冗余并消除它们，以揭示低维子空间来提升模型在分布外数据上的泛化性能。该数据集由复旦大学于2022年发布，相关资源包括GitHub代码和学术论文。