A Closer Look at In-Distribution vs. Out-of-Distribution Accuracy for Open-Set Test-time Adaptation

Technical Summary: A Closer Look at In-Distribution vs. Out-of-Distribution Accuracy for Open-Set Test-time Adaptation

Problem Statement
Deep neural networks often suffer accuracy degradation when test data undergoes distribution shifts. While Test-Time Adaptation (TTA) methods mitigate this by updating models on test inputs, standard approaches assume all test data belongs to known (In-Distribution or InD) classes. In realistic open-set scenarios, test data may also contain unknown (Out-of-Distribution or OOD) classes. Existing open-set TTA methods aim to adapt to InD data while filtering or rejecting OOD data. However, while these methods have demonstrated improvements in InD accuracy, their ability to accurately detect and reject OOD inputs remains underexplored. There is a critical gap in understanding the trade-off between maintaining high InD accuracy and ensuring reliable OOD rejection, particularly under varying proportions of OOD data and different normalization architectures.

Methodology and Benchmark Setup
The authors conduct a comprehensive benchmark of robust and open-set TTA methods, including SAR, OSTTA, UniEnt, UniEnt+, and SoTTA, alongside a simple baseline (Tent). The evaluation utilizes standard corruption benchmarks:

• Small Scale: CIFAR-10-C (InD) paired with corrupted SVHN-C or CIFAR-100-C (OOD).
• Large Scale: ImageNet-C (InD) paired with corrupted ImageNet-O-C (near OOD, visually similar but different classes) or Textures-C (far OOD, non-object patterns).

The study introduces several experimental variations to reflect realistic deployment conditions:

1. Mixed Batches: Evaluating performance where batches contain varying proportions of InD and OOD data (25%, 50%, 75%).
2. Batch Intervals: Testing scenarios where OOD data appears in consecutive bursts (intervals) between InD batches, rather than being uniformly mixed.
3. Output Distribution: The authors propose a new baseline replacing the standard softmax output with a sigmoid/multi-label output. This allows the model to output low probabilities for all classes simultaneously, enabling explicit rejection of unknown inputs without forcing a class prediction.
4. Normalization Layers: Investigating the impact of Batch Normalization (BN), Group Normalization (GN), and Layer Normalization (LN) on open-set adaptation stability.

Key Contributions and Results

• InD vs. OOD Trade-off: The analysis reveals that current open-set TTA methods struggle to balance InD recognition and OOD rejection. Optimizing for high InD accuracy often results in near-zero OOD accuracy (high false positive rates for OOD data), and vice versa. While methods like UniEnt achieve the best separation, they still face significant trade-offs.
• Limitations of Current Filtering: The paper demonstrates that the internal filtering mechanisms of existing methods (e.g., confidence thresholding in SoTTA, entropy filtering in SAR, prototype clustering in UniEnt) are imperfect. In some cases, their ability to distinguish InD from OOD is inferior to simple thresholding based on confidence scores derived from the model's raw outputs.
• Sigmoid Output Baseline: Replacing softmax with sigmoid output (Tent-Sigmoid and SAR-Sigmoid) yields a distinct trade-off profile. While it may slightly reduce InD accuracy compared to softmax-based counterparts, it significantly improves OOD detection metrics (AUROC, FPR@TPR95, OSCR). The sigmoid entropy distribution shows better separation between InD and OOD inputs, often exhibiting a bimodal distribution that facilitates rejection without complex filtering heuristics.
• Sensitivity to OOD Proportions: Existing methods are highly sensitive to the proportion of OOD data within a batch. As the OOD ratio increases, performance degrades consistently across all metrics. However, the methods are relatively insensitive to the interval of OOD batches (consecutive OOD batches), suggesting that the issue lies in the interaction of mixed data within a single batch update rather than the presence of OOD data over time.
• Role of Normalization Layers: The stability of adaptation is heavily dependent on the normalization layer. Models using Batch Normalization (BN) exhibit large fluctuations in accuracy and OOD metrics when faced with varying OOD proportions. In contrast, Group Normalization (GN) and Layer Normalization (LN) demonstrate significantly more stable performance, with LN showing the most robust results overall.

Significance and Claims
The paper claims to provide the first standardized, multi-metric evaluation of open-set TTA that explicitly measures the trade-off between InD and OOD accuracy. The authors argue that the field has overlooked the reliability of OOD rejection in favor of InD accuracy improvements. By introducing the sigmoid baseline and testing under realistic, variable OOD proportions, the work highlights that current methods do not effectively filter OOD data for adaptation updates.

The authors conclude that while their sigmoid baseline offers a different and often more favorable trade-off for OOD rejection, it is not a panacea; it sacrifices some InD accuracy. They suggest that future work must focus on developing methods that can better balance these competing objectives, potentially through improved filtering techniques or alternative output distributions, to ensure safe and reliable deployment in open-set environments. The paper does not claim to solve the problem entirely but rather identifies the specific gaps in current methodologies and provides a more complete empirical understanding of the challenges involved.