Core Concepts
HYPO proposes a novel hyperspherical learning algorithm for out-of-distribution generalization, focusing on low intra-class variation and high inter-class separation.
Abstract
Abstract:
Out-of-distribution (OOD) generalization is crucial for real-world machine learning models.
HYPO framework learns domain-invariant representations in a hyperspherical space.
Introduction:
Challenges of generalizing under distributional shifts are highlighted.
Importance of OOD generalization in various scenarios is emphasized.
Problem Setup:
Multi-class classification task with random variables X, Y over instances x ∈ X ⊂ Rd and labels y ∈ Y := {1, 2, · · · , C}.
Motivation of Algorithm Design:
Theoretical findings by Ye et al. [70] provide the basis for designing practical learning algorithms.
Method:
Introduction of the HYPO learning algorithm focusing on hyperspherical embeddings.
Experiments:
Strong OOD generalization performance demonstrated through extensive evaluations on various benchmarks.
Why HYPO Improves Out-of-Distribution Generalization?:
Formal justification provided on how the loss function reduces OOD generalization error theoretically.
Stats
この論文は、実験においてCIFAR-10(ID)対CIFAR-10-Corruption(OOD)タスクにおいて、Gaussian noiseのOOD一般化精度を78.09%から85.21%に向上させたことを示しています。
また、PACSデータセットで88.0%の精度を達成し、他のドメイン汎化ベンチマークでも優れたパフォーマンスを発揮しています。