🤖 machine learning

Online Continual Learning with Dynamic Label Hierarchies

This paper introduces the DHOCL problem setting for online continual learning with evolving label hierarchies and proposes HALO, a method using adaptive classification heads and organized hierarchical prototypes to overcome partial supervision and granularity-dependent interference, thereby achieving superior performance in hierarchical accuracy and stability.

Original authors: Xinrui Wang, Shao-Yuan Li, Bartłomiej Twardowski, Alexandra Gomez-Villa, Songcan Chen

Published 2026-05-13

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinrui Wang, Shao-Yuan Li, Bartłomiej Twardowski, Alexandra Gomez-Villa, Songcan Chen

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 trying to teach a robot to recognize animals. In the old way of doing this (called "Online Continual Learning"), you would show the robot a stream of pictures one by one. The robot would learn to say "Dog," then "Cat," then "Bird." But there was a catch: the robot treated every animal as a separate, unrelated item. It didn't understand that a "Golden Retriever" is a type of "Dog," and a "Dog" is a type of "Animal."

Furthermore, the old methods assumed you would teach the robot in a strict order: first the big categories (like "Animal"), then the medium ones ("Dog"), and finally the specific ones ("Golden Retriever").

The Real-World Problem
In the real world, life isn't that neat.

The Hierarchy is Fluid: Sometimes you see a specific "Siamese Cat" first, and only later realize it belongs to the "Cat" family, and even later, the "Feline" family. The "family tree" of knowledge is constantly growing and changing.
The Labels are Messy: Sometimes a human expert labels a picture as just "Bird" (coarse), while a novice labels another as "Monarcha melanopsis" (very fine). The robot gets confused because it's getting instructions at different levels of detail all at once.

The authors of this paper call this new, messy reality DHOCL (Dynamic Hierarchical Online Continual Learning). They found that existing robots fail here because:

They get confused when they only get a partial label (like "Bird") but need to understand the whole tree.
They forget old things quickly because learning a new specific detail (like a new bird species) messes up their understanding of the general category ("Bird").

The Solution: HALO
To fix this, the authors built a new system called HALO (Hierarchical Adaptive Learning with Organized Prototypes). Think of HALO as a super-smart librarian who manages a library that is constantly being renovated.

Here is how HALO works, using simple analogies:

1. The "Organized Prototypes" (The Sticky Notes)

Imagine the robot has a wall of sticky notes.

Old way: The robot just memorized the exact shape of every single picture it saw. If a new picture came in, it would overwrite the old ones.
HALO's way: Instead of memorizing every picture, HALO creates "prototypes" (like idealized sticky notes) for different levels of the tree.
- One note says "Bird" (with a generic bird shape).
- One note says "Parrot" (with a parrot shape).
- One note says "Macaw" (with a specific Macaw shape).
Crucially, HALO uses a special rule to glue these notes together. It ensures that the "Macaw" note is physically close to the "Parrot" note, and the "Parrot" note is close to the "Bird" note. Even if the robot learns a new type of bird tomorrow, it knows exactly where to put the new note so it doesn't break the connection between the specific bird and the general bird category. This keeps the robot's memory organized and consistent.

2. The "Two-Headed Brain" (The Flexible and The Stable)

HALO has two different "heads" (classifiers) working together, like a team of two experts:

Head A (The Fast Learner): This head is very flexible. It learns new things instantly. If a new species of bird appears, Head A grabs it immediately. However, it's a bit forgetful; it might lose track of old, rare birds.
Head B (The Stable Keeper): This head is very rigid and stable. It doesn't change its mind easily. It holds onto the old knowledge perfectly but is slow to learn new things.

The Magic Trick: HALO doesn't force these two heads to agree on everything. Instead, it acts as a manager.

When the robot sees a new, tricky bird, it asks Head A for help because Head A is good at spotting new details.
When the robot needs to remember an old, common bird, it asks Head B because Head B won't forget it.
HALO mixes the answers from both heads to make the final decision. This way, the robot learns fast and remembers well.

3. The "Replay Buffer" (The Review Session)

Like any good student, HALO reviews old notes. But instead of just picking random pictures to review, it uses a smart strategy to make sure it reviews the right mix of "Birds," "Parrots," and "Macaws" so it doesn't get confused about which level of the tree it's studying.

The Results
The authors tested HALO on several datasets (like pictures of airplanes, birds, and nature). They found that:

HALO made fewer "severe" mistakes (e.g., confusing a "Dog" with a "Car" is a severe mistake; confusing a "Labrador" with a "Poodle" is a minor one). HALO avoided the severe mistakes.
It learned new categories quickly without forgetting the old ones.
It worked well even when the "family tree" of labels was messy, incomplete, or changing rapidly.

In Summary
This paper introduces a new way for AI to learn from a messy, real-world stream of data where categories are organized in a family tree that keeps growing. They built a system (HALO) that uses "sticky note" prototypes to keep the family tree organized and a "two-headed" strategy to balance learning new things with remembering old things. The result is a robot that is much better at handling the chaotic, evolving nature of real-world knowledge.

Technical Summary: Online Continual Learning with Dynamic Label Hierarchies

Problem Definition: DHOCL
Existing Online Continual Learning (OCL) methods predominantly assume a "flat" label space where categories are independent and semantically equidistant. This assumption fails to capture the hierarchical organization of real-world concepts (e.g., biodiversity monitoring), which evolves both horizontally (new sibling classes) and vertically (refinement or abstraction of existing branches). Furthermore, real-world supervision is often partial; an observer may label an image as "Bird" (coarse) or "Monarcha melanopsis" (fine-grained) depending on expertise, without providing the full taxonomic lineage.

To address this gap, the authors introduce Dynamic Hierarchical Online Continual Learning (DHOCL). Unlike prior hierarchical continual learning works that enforce a rigid, predefined "coarse-to-fine" curriculum, DHOCL allows data instances to arrive with annotations at arbitrary hierarchical levels at any time. The taxonomy is not fixed; it grows organically as new classes emerge, potentially creating unbalanced and incomplete tree structures. The setting presents two fundamental challenges:

Partial Supervision and Structural Decoupling: Each sample provides only a "point-wise" signal on a "path-wise" hierarchy. This constrains plasticity and prevents the propagation of semantic constraints across granularities, leading to a fragmented latent space where taxonomic branches become structurally decoupled.
Granularity-Induced Interference: Different taxonomic depths exhibit heterogeneous learning and forgetting rates. Coarse-grained learning can destabilize fine-grained nuances and vice versa. Existing replay and regularization mechanisms, designed for flat spaces, fail to account for this cross-talk, leading to biased regularization and catastrophic forgetting.

Methodology: HALO
The authors propose HALO (Hierarchical Adaptive Learning with Organized Prototypes), a framework designed to balance rapid adaptation with long-term stability while maintaining hierarchical consistency. HALO consists of two core components:

Hierarchical Prototypes Regularization (HPR):
- HALO introduces a lightweight adapter that maps backbone features into a prototype space. For each class at every hierarchy level, a set of learnable prototypes is maintained.
- Structural Alignment: A margin-based ranking loss enforces that ancestor-descendant prototype pairs are more similar than unrelated pairs, explicitly baking the evolving tree structure into the feature space.
- Temporal Saliency Consistency: The method penalizes response changes at the most informative locations (Top-K patches) between the current model and cached prototypes from the previous step. This ensures stability for coarse semantics while allowing controlled drift for new fine-grained updates.
- These prototypes act as "topological glue," preventing the decoupling of taxonomic branches as the hierarchy evolves.
Hierarchical Adaptive Aggregation (PredLA):
- To resolve the conflict between the need for rapid adaptation (for new nodes) and stability (for old knowledge), PredLA employs a hierarchical adaptive ensemble of two complementary classifiers at each level:
  - A Linear Head trained on trainable features using cross-entropy loss, optimized for adaptability.
  - An Analytic Head (ACIL) trained on frozen features using Recursive Least Squares (RLS), optimized for stability and resistance to forgetting.
- Aggregation: Instead of merging at the parameter level, PredLA aggregates predictions in the probability space. It uses temperature scaling and learns aggregation weights ( $\alpha$ ) exclusively on replayed samples via a bilevel objective. This allows the system to balance the "plastic" linear predictions with the "stable" analytic predictions without interfering with their internal optimization dynamics.

Key Contributions

Formalization of DHOCL: The paper establishes a more realistic and general OCL setting that captures the dynamic evolution of hierarchical knowledge in streaming data, moving beyond rigid coarse-to-fine assumptions.
The HALO Framework: A novel architecture combining hierarchical adaptive aggregation and prototype-based regularization to address the specific challenges of dynamic label hierarchies.
Comprehensive Evaluation: The authors establish a rigorous evaluation protocol using five metrics (AAUC, FAUC, MS, FFAcc, FAAcc) across diverse benchmarks (CIFAR-100, FGVC-Aircraft, CUB-200, iNaturalist, and ImageNet-H) covering various hierarchical structures and domains.

Experimental Results
Extensive experiments demonstrate that HALO consistently outperforms existing OCL baselines (including replay-based, regularization-based, and hierarchical-aware methods) across all datasets and metrics.

Performance: HALO achieves superior Area Under the Curve (AAUC) and Fine-grained Accuracy (FAUC) while significantly reducing Mistake Severity (MS).
Stability-Plasticity Trade-off: Learning curves show that HALO learns new classes and hierarchy levels quickly while effectively limiting catastrophic forgetting, outperforming both pure replay methods (which forget fine-grained details) and pure analytic methods (which struggle with coarse-level adaptation).
Robustness: The method demonstrates robustness to different backbone architectures (ResNet-50, ViT-B) and pretraining strategies (Supervised, CLIP, DINO, MAE). It also exhibits graceful degradation under hierarchical noise (e.g., edge deletions or random perturbations in the taxonomy).
Ablation Studies: Removing either PredLA or HPR leads to performance drops, confirming that both the adaptive aggregation of classifiers and the prototype-based structural regularization are essential for success.

Significance and Claims
The authors position this work as a necessary step toward bridging the gap between hierarchical classification and continual learning. They claim that DHOCL offers a uniquely flexible and realistic setting that reflects the natural evolution of taxonomies in open-world environments. While acknowledging that their setting does not yet cover the full complexity of real-world scenarios (such as highly skewed or corrupted tree structures), they view HALO as a foundational framework that provides a practical benchmark and observations to inform future research in both fields. The work emphasizes that successful DHOCL requires strategies that explicitly align feature evolution with hierarchical topology, rather than relying solely on output-level supervision or generic forgetting mitigation.

1. The "Organized Prototypes" (The Sticky Notes)

2. The "Two-Headed Brain" (The Flexible and The Stable)

3. The "Replay Buffer" (The Review Session)

Technical Summary: Online Continual Learning with Dynamic Label Hierarchies

More like this