Exploring Open-world Continual Learning with Knowns-Unknowns Knowledge Transfer

TL;DR

This work targets Open-World Continual Learning (OWCL), where models must detect unknowns while incrementally learning new knowledge without forgetting. It formalizes four OWCL scenarios and reveals a strong interaction between open-detection and incremental classification, challenging the notion that these components are orthogonal. The authors propose HoliTrans, a framework that combines nonlinear random projection (NRP) and distribution-aware prototypes (DAPs) to enable simultaneous knowledge transfer for known and unknown samples, along with pseudo-open samples to update open representations. They provide theoretical bounds on open risk and incremental prediction error and demonstrate, through extensive experiments on CIRO and KIRO benchmarks, that HoliTrans outperforms 22 competitive baselines and maintains robustness across tasks. The approach advances open-world learning by bridging theory and practice and offering a scalable, open-sample-aware continual learning paradigm.

Abstract

Open-World Continual Learning (OWCL) is a challenging paradigm where models must incrementally learn new knowledge without forgetting while operating under an open-world assumption. This requires handling incomplete training data and recognizing unknown samples during inference. However, existing OWCL methods often treat open detection and continual learning as separate tasks, limiting their ability to integrate open-set detection and incremental classification in OWCL. Moreover, current approaches primarily focus on transferring knowledge from known samples, neglecting the insights derived from unknown/open samples. To address these limitations, we formalize four distinct OWCL scenarios and conduct comprehensive empirical experiments to explore potential challenges in OWCL. Our findings reveal a significant interplay between the open detection of unknowns and incremental classification of knowns, challenging a widely held assumption that unknown detection and known classification are orthogonal processes. Building on our insights, we propose \textbf{HoliTrans} (Holistic Knowns-Unknowns Knowledge Transfer), a novel OWCL framework that integrates nonlinear random projection (NRP) to create a more linearly separable embedding space and distribution-aware prototypes (DAPs) to construct an adaptive knowledge space. Particularly, our HoliTrans effectively supports knowledge transfer for both known and unknown samples while dynamically updating representations of open samples during OWCL. Extensive experiments across various OWCL scenarios demonstrate that HoliTrans outperforms 22 competitive baselines, bridging the gap between OWCL theory and practice and providing a robust, scalable framework for advancing open-world learning paradigms.

Figures (9)

• Figure 1: A toy example of the KIRO scenario in OWCL. During training Task 1 and Task t, Robots A and B learn Golden Retrievers and Labradors, both labeled as [dog]. In testing Task 1, both robots correctly classify [dog] and [bus] samples while identifying the unknown wolf without errors. However, in testing Task 2, Robot A struggles with Golden Retrievers and misclassifies the wolf as [dog] due to its similarity to Labradors. In contrast, Robot B accurately classifies all known samples and continually detects the wolf as an open sample with knowledge transfer.
• Figure 2: From task $t$ to task $t+1$ in the KIRO scenario, we illustrate the embedding space utilizing different baselines. Red circles and red triangles denote unknown samples encountered in tasks $t$ and $t-1$, respectively. Orange, blue, and green circles represent the labeled training data learned during tasks $t-1$, $t$, and $t+1$, respectively. Dashed lines depict the decision boundaries inferred by the model during OWCL. (a) The original embedding distributions for each task learned by a classical CL baseline, EWC. (b) A current competitive OWCL approach enforces tighter embedding cohesion for each task. (c) Our proposed method concurrently addresses open detection and incrementally reduces known sample classification errors with knowledge transfer. [Best view in color]
• Figure 3: Unscaled Classification Scores on the CINO scenario.
• Figure 4: Unscaled Classification Scores on the CIRO scenario.
• Figure 5: Unscaled Classification Scores on the KINO Scenario.

Theorems & Definitions (8)

• Definition 1
• Definition 2
• Theorem 1
• Theorem 2
• Lemma 1
• Lemma 2
• Theorem 3
• Corollary 1