On the Robustness of Open-World Test-Time Training: Self-Training with Dynamic Prototype Expansion

Generalizing deep learning models to unknown target domain distribution with low latency has motivated research into test-time training/adaptation (TTT/TTA). Existing approaches often focus on improving test-time training performance under well-curated target domain data. As figured out in this work, many state-of-the-art methods fail to maintain the performance when the target domain is contaminated with strong out-of-distribution (OOD) data, a.k.a. open-world test-time training (OWTTT). The failure is mainly due to the inability to distinguish strong OOD samples from regular weak OOD samples. To improve the robustness of OWTTT we first develop an adaptive strong OOD pruning which improves the efficacy of the self-training TTT method. We further propose a way to dynamically expand the prototypes to represent strong OOD samples for an improved weak/strong OOD data separation. Finally, we regularize self-training with distribution alignment and the combination yields the state-of-the-art performance on 5 OWTTT benchmarks. The code is available at https://github.com/Yushu-Li/OWTTT.

1. Introduction

The distribution gap between training and testing data poses great challenges to the generalization of modern deep learning methods [34], [4]. To improve model’s generalization to testing data which may feature a different data distribution from the training data, domain adaptation has been extensively studied [44] to learn domain invariant features. Nevertheless, the existing unsupervised domain adaptation paradigm requires simultaneous access to both source and target domain data with an off-line training stage [12], [41]. In a realistic scenario, access to target domain data may not become available until the inference stage, and instant prediction on testing data is required without further ado. Therefore, these requirements give rise to the emergence of a new paradigm of adaptation at test time, a.k.a. test-time training/adaptation (TTT/TTA) [40], [43].