Task Identification for Class Incremental Learning

This direction is split into other directions: task-agnostic inference (class-incremental learning challenge). Its last version stays readable and citable here.

Version 1 of 195 papers101 conceptsfrozen 2026-10-04

In the world of continual learning, there is a massive performance gap between two ways of testing a model. In Task-Incremental Learning, the model is told which task it is currently performing (e.g., "you are now looking at birds"). In Class-Incremental Learning (CIL), the model is given an image and must decide what it is across all categories it has ever seen, without any hints.

As Andrea Cossu and Antonio Carta have noted, when task labels are provided, forgetting almost disappears. The "catastrophic" part of catastrophic forgetting isn't necessarily that the model loses the ability to recognize a specific object; it's that the model loses the ability to distinguish which context or task the object belongs to.

Recent research has converged on a shared insight: to solve Class-Incremental Learning, we must turn it into a Task Identification problem. If a model can accurately infer the task ID at inference time, it can route the input to the correct specialized parameters, effectively bridging the gap between the two settings.

The Task ID Gap

The primary challenge in CIL is the lack of a task identifier at test time. Without this metadata, a model must resolve a "global" decision boundary between classes that were never seen together during training. Bing Liu and Gido M. van de Ven argue that this ambiguity is the root of performance degradation. If the system cannot correctly infer the underlying context, it cannot apply the appropriate knowledge.

To address this, researchers are developing mechanisms that allow models to autonomously determine which subset of their learned parameters is relevant for a given input.

Task Identification as OOD Detection

One of the most powerful conceptual shifts in this field is the realization that identifying a task is essentially an Out-of-Distribution (OOD) detection problem.

In Open-World Continual Learning, Gyuhak Kim and Bing Liu provide a theoretical bridge between these fields. They prove that if a model can detect when an input does not belong to the current task (OOD), it can use that signal to identify which previous task the input does belong to. Their algorithm, MORE, uses this distance-based task identification to achieve state-of-the-art results. Similarly, Xuchen Xie uses OOD detection paired with task-specific Batch Normalization to turn the difficult CIL problem into a manageable task-prediction challenge.

Dynamic Routing and Gating

If the model can't be told the task, it must "route" the input itself. This has led to the development of modular architectures where different "experts" or "adapters" handle different tasks.

  • Conditional Gating: Davide Abati and Rita Cucchiara introduced Conditional Channel Gated Networks, which use task-specific gating modules to select relevant filters.
  • Learned Routing: Vladimir Araujo and Tinne Tuytelaars developed L2R (Learning to Route), which trains a specialized router to dynamically compose task-specific adapters at inference time.
  • Task Descriptors: Tenghao Cai proposed a multi-centroid task descriptor that acts as a "gate network," predicting the task ID so the model can focus on the most relevant parameters.

Subspace Isolation and Orthogonal Projections

Another way to handle the lack of task labels is to ensure that tasks never interfere with each other in the first place. By projecting new knowledge into mathematical subspaces that are orthogonal to previous tasks, researchers can prevent "cross-talk" between classes.

Huancheng Chen and Seohyeon Cha have advanced this through methods like DualLoRA and FedProTIP. These frameworks use Task Identity Prediction (TIP) mechanisms to navigate these subspaces during inference. By keeping the features of different tasks geometrically separated, the model can more easily identify which "subspace" an input belongs to, even without an explicit label.

Unified Embedding Spaces

Finally, some researchers argue that instead of explicitly predicting a task ID, we should focus on building a unified embedding space where classes naturally stay separated.

Jary Pomponi and Simone Scardapane introduced Centroids Matching (CM), which represents classes as fixed points (centroids) in an embedding space. By ensuring that new learning doesn't distort the geometric relationships of old features, the model maintains a unified decision boundary. This task-agnostic approach allows the model to recognize all learned categories simultaneously, fulfilling the core requirement of Class-Incremental Learning without needing to "guess" the task ID.

Papers in this version

Showing 60 of 95 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_task_identification_for_class_incrementa_e7a07661,
  title = {Task Identification for Class Incremental Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/task-identification-for-class-incremental-learning/rdr_e7a07661b54249b7bef996bede242f71},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_task_identification_for_class_incrementa_e7a07661_v1,
  title = {Task Identification for Class Incremental Learning},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/task-identification-for-class-incremental-learning/rdr_e7a07661b54249b7bef996bede242f71/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}