Introduction
In the evolving field of Task Incremental Learning (TIL), models are typically evaluated in "closed-world" environments where they only encounter data from previously learned tasks. However, real-world applications, such as autonomous driving or robotics, operate in an "open-world." In these scenarios, a model must not only classify known tasks but also identify Out-of-Distribution (OOD) samples—data that does not belong to any previously seen category.
Existing OOD detection methods struggle with three primary challenges in continual learning: threshold sensitivity, where manually tuning a score-based threshold is impractical in non-stationary streams; model dependence, where poor classification performance degrades OOD detection; and a lack of task identity, as standard methods often fail to specify which learned task an in-distribution (ID) sample belongs to.
The H2ST Framework
To address these hurdles, we introduce H2ST (Hierarchical Two-Sample Tests). This framework shifts OOD detection from simple score-based thresholding to a robust statistical hypothesis testing paradigm.
Hierarchical Architecture and Early-Exit
Instead of a monolithic classifier, H2ST employs a cascade of task-specific binary classifiers. As a new sample enters the system, it traverses this hierarchy. If a classifier identifies the sample as belonging to its specific task, the process triggers an early-exit, and the sample is passed to that task's inference head. If the sample passes through all $T$ layers without being matched, it is definitively labeled as OOD. This hierarchical approach significantly improves efficiency; on average, a sample only requires $(T+1)/2$ tests.
Feature-Level Statistical Testing
H2ST performs tests on intermediate feature maps rather than final logits. High-level semantic representations are often more stable than classification layers, making the system more resilient to "catastrophic forgetting." For each task, a small binary classifier—the Classifier Two-Sample Test (C2ST)—is trained to distinguish between "source" samples (from a memory buffer) and new incoming data. Decisions are made using the Clopper-Pearson (CP) interval, providing a statistically grounded, threshold-free deployment.
Experimental Evaluation
We evaluated H2ST across diverse benchmarks, including MNIST, SVHN, CIFAR-10, CIFAR-100, and real-world datasets like CoRe50 and Stream-51. The experiments utilized a "6k-2k-2k" configuration, simulating an open-world stream where incremental task learning is interspersed with OOD detection rounds.
Performance and Efficiency
H2ST demonstrates a significant performance leap over traditional baselines such as Maximum Softmax Probability (MSP), Energy-based models, and ODIN.
As shown in the results, H2ST achieves an average F1 score of approximately 92% on CIFAR-10, nearly 30 percentage points higher than the next best method (MORE). Crucially, H2ST provides a massive boost in Task Accuracy (TA)—the ability to correctly identify the task-id of an ID sample—reaching ~90% where others hover around 20%.
Beyond accuracy, H2ST is computationally efficient. In a 9-task setup, it processed inputs in 16.7ms, compared to 22.0ms for a non-hierarchical C2ST approach, validating the benefits of the early-exit strategy.
Conclusion
H2ST offers a robust, threshold-free solution for open-world continual learning. By decoupling OOD detection from classification performance and utilizing a hierarchical statistical testing structure, it provides superior detection and task identification. While the method currently relies on a memory buffer for source samples, its compatibility with standard continual learning strategies like Experience Replay (ER) makes it a versatile tool for real-world deployment. Future work will focus on further reducing the overhead of online binary classifier updates and optimizing the task order within the hierarchy.
Summary updated