Out-of-Distributed Semantic Pruning for Robust Semi-Supervised Learning

TL;DR

This paper introduces OOD Semantic Pruning (OSP), a novel framework that enhances robust semi-supervised learning by effectively removing out-of-distribution semantic information, leading to significant improvements in ID classification and OOD detection.

Contribution

The paper proposes a unified OOD Semantic Pruning framework with a matching module and a regularization technique to improve robustness in semi-supervised learning.

Findings

OSP outperforms previous methods in ID classification accuracy by 13.7%.

OSP improves OOD detection AUROC by 5.9%.

The method is simple and effective on challenging benchmarks.

Abstract

Recent advances in robust semi-supervised learning (SSL) typically filter out-of-distribution (OOD) information at the sample level. We argue that an overlooked problem of robust SSL is its corrupted information on semantic level, practically limiting the development of the field. In this paper, we take an initial step to explore and propose a unified framework termed OOD Semantic Pruning (OSP), which aims at pruning OOD semantics out from in-distribution (ID) features. Specifically, (i) we propose an aliasing OOD matching module to pair each ID sample with an OOD sample with semantic overlap. (ii) We design a soft orthogonality regularization, which first transforms each ID feature by suppressing its semantic component that is collinear with paired OOD sample. It then forces the predictions before and after soft orthogonality decomposition to be consistent. Being practically simple, our method shows a strong performance in OOD detection and ID classification on challenging benchmarks. In particular, OSP surpasses the previous state-of-the-art by 13.7% on accuracy for ID classification and 5.9% on AUROC for OOD detection on TinyImageNet dataset. The source codes are publicly available at https://github.com/rain305f/OSP.