🤖 machine learning

When Classes Evolve: A Benchmark and Framework for Stage-Aware Class-Incremental Learning

This paper introduces Stage-Aware Class-Incremental Learning (Stage-CIL) to address the challenge of learning classes that undergo morphological evolution over time, proposing the Stage-Bench benchmark and the STAGE framework to effectively mitigate both inter-class forgetting and intra-class stage-level degradation.

Original authors: Zheng Zhang, Tao Hu, Xueheng Li, Yang Wang, Rui Li, Jie Zhang, Chengjun Xie

Published 2026-08-05

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Zheng Zhang, Tao Hu, Xueheng Li, Yang Wang, Rui Li, Jie Zhang, Chengjun Xie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are teaching a robot to recognize animals. In the world of artificial intelligence, there is a famous challenge called "Class-Incremental Learning." Think of this like a student taking a long series of exams. First, they learn to identify dogs. Then, they learn cats. Then, birds. The tricky part is that the robot has a very limited memory. Every time it learns a new animal, it tends to "forget" the old ones, a glitch known as "catastrophic forgetting." Usually, scientists assume that once the robot learns what a "dog" looks like, a dog stays a dog forever. They assume the animal doesn't change its shape or appearance over time.

But in the real world, things aren't that static. A caterpillar doesn't stay a caterpillar; it turns into a butterfly. A baby car doesn't look exactly like a rusted, dented old car years later. If a robot is taught that a "caterpillar" and a "butterfly" are two completely different things, it misses the point that they are the same creature at different ages. If it's taught they are the same thing but doesn't know how to handle the drastic change in appearance, it gets confused. This paper tackles a specific corner of this problem: what happens when the things you are learning to recognize actually evolve while you are learning them?

The researchers behind this paper, led by Zheng Zhang and Tao Hu, realized that standard AI training methods are failing at this specific task. They argue that current methods treat a changing object as if it were a static one, which causes the AI to lose track of the object's true identity as it morphs. To fix this, they created a new "game" for AI called Stage-Aware Class-Incremental Learning (Stage-CIL). Instead of just asking, "Is this a dog?", the new game asks, "Is this a dog, and is it a puppy or an adult?" while also making sure the AI doesn't forget the puppies it saw weeks ago.

To prove this is a real problem, the team built a massive test suite called Stage-Bench. Imagine a giant photo album with 10 different chapters (like Animals, Plants, and Objects). Inside each chapter, there are 20 different items, and for every single item, there are two photos: one of it in its "early stage" (like a seedling or a larva) and one of it in its "later stage" (like a tree or a butterfly). In total, they created a dataset with nearly 19,000 images. They used this to test how well different AI models could learn new items without forgetting the old ones, and crucially, without forgetting that the "early version" and the "late version" are actually the same thing.

When they ran the tests, the results were eye-opening. They tried all the top-performing AI methods that are usually considered the "champions" of learning new things. However, when faced with these evolving objects, even the best existing methods struggled. They suffered from what the authors call "intra-class forgetting." This is a fancy way of saying the AI would learn to recognize a butterfly, but then completely forget what the caterpillar looked like, even though they are the same species. The standard methods tried to force the caterpillar and butterfly into a single, static box, and the AI got confused by the huge difference in their looks.

To solve this, the authors introduced a new method called STAGE. Think of STAGE as a smart detective who keeps a "mugshot" of the object in its early form (the identity anchor). When the object shows up later in a new, evolved form, STAGE doesn't just guess; it uses a special "memory pool" of transformation patterns to predict what the new version should look like based on the old one. It's like having a mental map that says, "If I see a caterpillar, I know exactly how it will look when it becomes a butterfly."

The experiments showed that STAGE was significantly better than the competition. On the main test, STAGE achieved an average accuracy of 75.11%, while the next best method only reached 59.74%. More importantly, STAGE barely forgot the early stages of the objects, with an "intra-class forgetting" score of just 7.48%, whereas other methods struggled with scores as high as 58.61%. This suggests that by explicitly teaching the AI to understand that objects can change over time, rather than pretending they stay the same, we can build much smarter, more adaptable systems. The paper concludes that while existing methods are good at learning new categories, they need a major upgrade to handle the messy, evolving reality of the world we live in.

Technical Summary: When Classes Evolve: A Benchmark and Framework for Stage-Aware Class-Incremental Learning

Problem Definition: Stage-Aware Class-Incremental Learning (Stage-CIL)
Conventional Class-Incremental Learning (CIL) operates under the implicit assumption that semantic classes are morphologically static; the primary challenge is mitigating catastrophic forgetting of previously learned categories as new ones are introduced. However, in many real-world scenarios, instances of the same semantic class undergo substantial morphological evolution over time (e.g., a larva transforming into a butterfly, or a vehicle accumulating damage).

The paper identifies a fundamental dilemma when applying standard CIL to such scenarios:

Grouping all stages under a single label flattens the evolutionary structure, inducing severe intra-class feature interference.
Splitting stages into separate pseudo-classes severs the underlying identity link, failing to preserve semantic consistency.

This leads to a new form of intra-class forgetting, where a model fails to maintain recognition consistency across different morphological stages of the same semantic class, even if it successfully discriminates between different classes. Standard domain-shift protocols do not explicitly evaluate this class-conditioned identity preservation within an expanding label space.

Methodology: The STAGE Framework
To address Stage-CIL, the authors propose STAGE, an evolution-aware reference baseline designed to disentangle stable semantic identity from stage-specific dynamics. The framework operates on a "predict-then-classify" pipeline:

Identity Anchor Construction (Phase 0):
- The model freezes the backbone (using a Pre-trained Model, specifically CLIP ViT-B/16) and learns task-specific image and text projection heads.
- For each new class, it constructs a stable identity prototype ( $p^y_0$ ) based on the initial morphological stage (Stage-0). This prototype is stored as a fixed vector and serves as the semantic anchor.
Predictive Evolution via Memory Pool (Phase 1):
- Instead of replacing the anchor with mixed-stage features, STAGE predicts the evolved class prototype ( $\hat{p}^y_1$ ) for the later stage (Stage-1) using the Stage-0 anchor.
- This is achieved through an Evolution-aware Memory Pool, a fixed-size trainable memory containing reusable transformation patterns.
- The model uses an attention mechanism to select the top- $k$ patterns from the pool that best match the current class anchor.
- An evolution network ( $E$ ) applies a residual update to the anchor based on the attended context to predict the Stage-1 prototype: $\hat{p}^y_1 = p^y_0 + E(p^y_0, c^y)$ .
Classification and Training Objectives:
- Stage-Agnostic Classification: At inference, stage labels are unavailable. The model classifies an input by computing logits against the best-matching prototype across all available stages (either the original anchor or the predicted evolved prototype).
- Loss Functions: The training objective combines standard cross-entropy with specific terms:
  - Evolution Loss ( $L_{evo}$ ): Minimizes the distance between the predicted and actual Stage-1 prototypes, aligning the predicted displacement direction with the ground truth.
  - Pattern Rehearsal ( $L_{pattern}$ ): Ensures the evolution predictor remains consistent on previously learned patterns in the memory pool, preventing arbitrary changes to the transformation operator.
  - Prototype Rehearsal ( $L_{proto}$ ): Maintains alignment with the fixed anchor space to preserve inter-class discrimination.

Benchmark: Stage-Bench
The authors introduce Stage-Bench, the first multi-domain benchmark tailored for Stage-CIL.

Scope: It spans 10 diverse semantic domains (Animals, Plants, Artifacts) with 20 classes per domain, totaling 400 morphological stages.
Protocol: It utilizes a two-stage regime (Stage-0: initial form; Stage-1: evolved form) with a controlled progressive learning protocol.
Metrics: In addition to standard average accuracy, the benchmark introduces two diagnostic metrics:
- Inter-class Forgetting (Inter-F): Measures the drop in accuracy for previously learned classes.
- Intra-class Forgetting (Intra-F): Specifically measures the performance drop on the earliest morphological stage of a class after the model has been updated with later-stage data. This isolates the degradation of identity consistency across stages.

Key Results
Extensive experiments on Stage-Bench demonstrate that:

Limitations of Existing Methods: State-of-the-art CIL and Domain-Incremental Learning (DIL) methods (e.g., iCaRL, L2P, DualPrompt, MOS, SimpleCIL) perform significantly worse under Stage-CIL. They exhibit high Intra-F scores, indicating they fail to preserve identity consistency across evolving forms.
STAGE Performance: STAGE consistently outperforms all competitors. On the primary two-stage benchmark, it achieves an average accuracy of 75.11% and a final accuracy of 56.44%, significantly surpassing the next best method, MOS, which achieved 59.74% average accuracy.
Forgetting Reduction: STAGE drastically reduces forgetting, achieving an Intra-F of 7.48% compared to 24.44%–58.61% for other methods.
Generalization: When tested on a restructured three-stage setting (Object domain), STAGE maintains a clear lead with an Intra-F of only 0.31%, suggesting the approach scales to more complex evolutionary regimes.

Significance and Claims
The paper claims three primary contributions:

Formalization of Stage-CIL: It establishes a new continual-learning setting that explicitly addresses the challenge of preserving semantic identity across intra-class morphological evolution, distinguishing it from standard CIL and DIL.
Benchmarking: It provides Stage-Bench and the associated diagnostic metrics (Inter-F and Intra-F) to rigorously evaluate and isolate intra-class forgetting, which previous benchmarks failed to capture.
Proof of Concept: Through STAGE, the authors demonstrate that explicit modeling of stage transformation is a viable and promising direction for solving intra-class forgetting. By separating stable identity anchors from dynamic transformation patterns, the framework enables the prediction of later morphological forms from earlier representations, effectively mitigating the "intra-class forgetting" problem.

The authors conclude that conventional CIL reductions (ignoring stage structure) are insufficient for evolving classes and that future continual learning systems must account for temporal, stage-specific evolution within semantic categories.

Technical Summary: When Classes Evolve: A Benchmark and Framework for Stage-Aware Class-Incremental Learning

More like this