Source-linked AI summary
Online Versatile Incremental Learning: Towards Class and Domain-Agnostic Adaptation at Any Time
Jae-Ho Lee, Min-Yeong Park, Jun-Yeong Moon, Jung Uk Kim, Gyeong-Moon Park
TL;DR
Online adaptation must distinguish class knowledge from domain-specific features amid chaotic joint shifts and without prior inputs. The paper introduces Online VIL and TopFlow, which combines domain-agnostic flow matching with global topology preservation. TopFlow is reported to consistently outperform existing state-of-the-art methods across multiple benchmarks.
Problem
Online adaptation must distinguish class-discriminative knowledge from domain-specific features amid chaotic class and domain shifts, without access to prior inputs.
Source: §1 · p. 3Method
TopFlow combines geodesic-flow-based Domain-agnostic Flow Matching with Global Topology Preservation to align features and retain feature-space structure.
Source: §1 · p. 3Results
TopFlow consistently outperforms existing state-of-the-art methods across multiple Online VIL benchmarks.
Source: §1 · p. 3Takeaways & Limitations
Online VIL may serve as a more realistic benchmark for advancing incremental-learning research under ever-changing conditions and without clear task boundaries.
Source: §5 · p. 14The stated architectural setting concerns feature dimensionality across layers, as in ViTs.
Source: §3.2 · p. 7
Abstract
from arXiv · showhide
Continual learning enables vision systems to adapt to ever-changing data distributions. Despite significant advances, existing approaches fail to capture continuous and concurrent shifts in classes and domains, a critical capability for real-world deployment. This work introduces Online VIL (Online Versatile Incremental Learning), a novel scenario where class concepts and visual domains evolve simultaneously online without explicit boundaries. To better adapt to the challenges of such dynamic environments that more closely resemble real-world conditions, we propose a novel framework TopFlow, Topology preservation with Flow matching representation that contains two complementary mechanisms: Domain-agnostic Flow Matching (DFM) and Global Topology Preservation (GTP). DFM guides the model to have domain-agnostic representations by integrating the geodesic flow kernel into contrastive learning. In contrast, GTP maintains the global structure of the feature space without explicitly storing past examples. Our extensive experiments demonstrate that TopFlow effectively addresses the limitations of existing methods within the Online VIL scenario, achieving state-of-the-art performance in challenging Online VIL. The proposed methods suggest potential directions for building continual learning systems in realistic dynamic environments. Our implementation code is available at https://github.com/KU-VGI/Online-VIL.
1 Introduction
Online VIL targets continual learning under unpredictable, continuous shifts in classes and domains, where existing paradigms do not capture their joint evolution without boundaries. TopFlow combines domain-agnostic feature alignment with global topology preservation, and is reported to outperform existing state-of-the-art methods across multiple benchmarks.
1 Introduction
- Online VIL models unpredictable, gradual, heterogeneous shifts across class and domain dimensions without explicit task boundaries.Its sequences allow flexible transitions across both dimensions rather than heuristic patterns.Source: §1 · p. 2
- Online adaptation must distinguish class-discriminative knowledge from domain-specific features without access to prior inputs.Otherwise, models may rely on spurious current-distribution features and exhibit rapid forgetting or poor generalization.Source: §1 · pp. 2–3
- DFM uses a geodesic flow kernel in contrastive learning to align features with pre-trained knowledge while mitigating domain-specific shifts.The motivating feature analysis finds early ViT layers predominantly encode domain-specific patterns and deeper layers class-discriminative features.Source: §1 · p. 3
- GTP preserves invariant geometric configurations in feature space for knowledge retention without complete class coverage or explicit memory of previous inputs.TopFlow combines GTP with DFM to support adaptation to unpredictable shifts.Source: §1 · p. 3
- TopFlow consistently outperforms existing state-of-the-art methods across multiple Online VIL benchmarks.The framework combines Domain-agnostic Flow Matching (DFM) and Global Topology Preservation (GTP).Source: §1 · p. 3
2 Related Work
Prior online continual-learning methods address streaming data and memory constraints, but often focus on class-only or domain-only shifts. Related work on flow alignment and feature topology also leaves gaps for sequential data and partial class information.
Online Continual Learning
- Online continual learning handles streaming data under single-pass, small-batch, memory, and computational constraints, but existing approaches still miss joint class-and-domain shifts.Replay buffers, prototypes, replay-free methods, and pretrained-model prompts are among the approaches discussed.Source: §2 · pp. 3–4
Geodesic Flow Kernel
- Conventional Geodesic Flow Kernel alignment assumes fully available, static source and target domains, unlike sequentially arriving data with evolving domains.DFM reformulates the approach for online feature-geometry alignment without holistic data access or costly higher-order manifold computations.Source: §2 · p. 4
Feature Topology
- Prior feature-topology methods require past samples and complete semantic information, limiting topology construction when online steps reveal only partial classes.GTP is presented as maintaining structural knowledge without storing previous data.Source: §2 · p. 4
3 Method
Online VIL represents continuously shifting domain-class data as a locally observed stream; DFM aligns intermediate features with final-layer semantics, while GTP maintains global feature topology through prototype relationships.
3.1 Problem Setup: Online Versatile Incremental Learning
- Online VIL constructs a continuous stream with stochastic domain-, class-, and time-varying distributions, limited local visibility, smooth shifts, and gradual category appearance and disappearance.Source: §3.1 · p. 5
3.2 Domain-agnostic Flow Matching
- DFM pulls adaptable intermediate features toward final-layer class semantics while pushing them away from domain-sensitive intermediate regions.It uses a batch-wise comparison geometry between intermediate and final representations to normalize contrastive similarities.Source: §3.2 · pp. 7–8
3.3 Global Topology Preservation
- GTP addresses absent-class rank deficiency with batch-derived, EMA-updated global prototypes, then aligns pairwise prototype relationships to preserve coarse global topology without explicit spectral computation.Source: §3.3 · pp. 9–10
4 Experiments
Experiments compare TopFlow with continual-learning methods across Online VIL benchmarks and examine its components in other settings. The reported analyses cover component ablations, layer selection, cross-scenario performance, and feature topology.
4.1 Experimental Setup
- Experiments use iDigits, CORe50, and CLEAR100, comparing TopFlow with continual-learning baselines and assessing online performance with AAUC and ALast.CORe50 and CLEAR100 use 10 tasks, while iDigits uses 5.Source: §4 · p. 10, §4.1 · pp. 10–11
4.2 Experimental Results
- TopFlow outperforms existing methods on AAUC and ALast without replay and consistently outperforms baselines with replay, while maintaining a steadier accuracy trajectory.The trajectory has fewer drastic drops between tasks.Source: Table 1 · p. 11, §4.2 · pp. 11–12, Table 2 · p. 12
4.3 Ablation Studies and Analysis
- DFM and GTP each improve AAUC and ALast, with complementary gains when combined: DFM improves representations while GTP maintains their structure.GTP also preserves performance under varying batch compositions by consolidating transient batch cues into a persistent feature topology.Source: Table 3 · p. 12, §4.3 · pp. 12–13
- DFM and GTP improve performance in both standard CIL and DIL, with more pronounced improvements in DIL.The evaluations use CIFAR-100 10-split CIL and CORe50 8-split DIL.Source: Table 4 · p. 12, §4.3 · p. 13
- 54.82 ALast is achieved with DFM layers (6,11), the best reported configuration; early layers (0,5) degrade performance.Combining (5,10) and (6,11) reaches 53.49.Source: Table 5 · p. 12, §4.3 · p. 13
- TopFlow also improves performance across Offline VIL benchmarks, while the Offline VIL method ICON is not competitive in Online VIL.The authors describe TopFlow as transferring across Offline and Online VIL settings.Source: Table 7 · p. 13, §4.3 · pp. 13–14
- DFM and GTP improve baseline performance independently and together when integrated with other continual-learning algorithms on CORe50.Source: Table 6 · p. 13, §4.3 · p. 13
- With GTP, t-SNE visualizations show better preservation of relative cluster arrangements across tasks than without GTP.The analysis tracks the first 10 classes available across CORe50 tasks.Source: Figure 4 · p. 14, §4.3 · p. 14
5 Conclusion
The paper introduces Online VIL, a setting with stochastic, evolving data and no clear task boundaries, and proposes TopFlow for it. The authors report state-of-the-art performance in Online VIL and note variability from stochastic task construction.
5 Conclusion
- Online VIL models ever-changing states without task concepts or clear boundaries.Its stochastic task construction can introduce variability in results.Source: §5 · p. 14
- TopFlow achieves state-of-the-art performance in the challenging Online VIL scenario.The authors report its effectiveness across various experiments.Source: §5 · p. 14
- The authors hope Online VIL will serve as a more realistic and challenging benchmark for future incremental-learning research.Source: §5 · p. 14