Neural networks forget. Train a model to classify birds, then train it to classify flowers, and watch its bird accuracy collapse. This phenomenon — catastrophic forgetting — is the central obstacle in continual learning, the problem of building systems that accumulate knowledge over time without destroying what they already know. A growing body of work suggests that concept bottleneck architectures may offer a structural solution: models that learn through interpretable concepts appear to forget less, and recent advances at CVPR 2025 are making this connection precise.
Why Concepts Resist Forgetting
The intuition is straightforward. Standard neural networks encode task knowledge as distributed patterns across millions of parameters. When new task data arrives, gradient updates overwrite these patterns indiscriminately — the network has no way to identify which parameters encode which knowledge. Concept bottleneck models, by contrast, organize knowledge through an explicit intermediate layer of human-interpretable concepts. If "wing shape" and "beak curvature" are the concepts used to classify birds, these concept representations can be preserved or updated selectively when flower classification is introduced, because the model knows which parameters correspond to which concepts.
This architectural separation between concept encoding and task prediction creates a natural mechanism for knowledge preservation. The concept layer acts as a stable anchor: concepts learned for one task can transfer to another (both birds and flowers have color, texture, and shape), while task-specific prediction heads can be added without disturbing shared concept representations.
Language as the Concept Anchor
Yu, Han, and Tao (2025), in work presented at CVPR 2025, formalize this intuition through Language Guided Concept Bottleneck Models for continual learning. Their key insight is that natural language provides a stable, task-independent vocabulary for concept representation. Rather than learning concept vectors from scratch for each task — which risks the same catastrophic overwriting that plagues standard networks — they ground concepts in the embeddings of a frozen language model.
When a new task arrives, the system generates concept descriptions in natural language (such as "has a pointed beak" or "has serrated leaf edges"), maps them to the language model's embedding space, and uses these embeddings as the concept bottleneck layer. Because the language model is frozen, the concept representations remain stable across tasks. The model learns to map visual inputs to these linguistically grounded concepts, and then maps concepts to task predictions. New tasks add new concept descriptions and new prediction heads but do not modify existing concept representations.
The approach achieves two goals simultaneously. It mitigates catastrophic forgetting by anchoring concepts in a fixed representational space. And it maintains interpretability across the entire learning sequence — at any point, the model's predictions can be traced through human-readable concept descriptions, making it possible to audit what the system knows and how it reasons.
Concept-Driven Task Separation
Yang, Oikarinen, and Weng (2024) approach the same problem from a different angle in their work on concept-driven continual learning. Their method uses concepts not merely as stable anchors but as active organizers of the learning process itself. By identifying which concepts are shared across tasks and which are task-specific, the system can partition its representational space — allocating dedicated capacity for task-specific concepts while sharing parameters for universal ones.
This concept-driven partitioning addresses a limitation of replay-based continual learning methods, which preserve old knowledge by storing and periodically retraining on exemplars from previous tasks. Replay is expensive in memory and computation, and it raises privacy concerns when training data cannot be retained. A concept-driven approach reduces the need for replay by structurally preventing the interference that replay is designed to mitigate. If the model knows that "wing shape" belongs to the bird task and "petal arrangement" belongs to the flower task, it can protect the relevant parameters without needing to store bird images.
Scaling Concepts to Unseen Classes
Zhang, Luo, Yang et al. (2025), also at CVPR 2025, address a complementary challenge: how to organize the concept bottleneck so that it scales to classes the model has never seen. Their Attribute-formed Language Bottleneck Model organizes concepts not as a flat list but as an attribute-structured space where concepts describe specific attributes of specific classes.
The distinction matters for two reasons. A flat concept list invites spurious inference — the model might classify an image as a dog because it matches the concept "four legs," ignoring that the concept also applies to cats, horses, and tables. By organizing concepts as class-specific attribute descriptions ("dog: floppy ears" versus "cat: pointed ears"), each class is identified by the conjunction of its specific attribute values rather than by individual concepts that may be shared across many classes.
The second benefit is compositional generalization. Because the attribute set is shared across classes — every class has entries for shape, texture, color, size — the concept classifier trained on known classes transfers to unseen classes that share the same attribute vocabulary. A model that has learned to evaluate "wing shape" and "body color" for familiar bird species can apply those same attribute evaluators to an unfamiliar species, constructing its concept profile without additional training.
Experiments on nine few-shot benchmarks demonstrate that ALBM matches or exceeds the accuracy of standard approaches while providing concept-level explanations and generalizing to new classes with minimal examples. The combination of interpretability, scalability, and compositional transfer suggests that structuring the concept space is as important as having a concept space at all.
The Deeper Pattern
These three lines of work share a common insight: concepts are not just a tool for interpretability but a structural principle for organizing knowledge. A model that reasons through concepts can remember better, transfer more efficiently, and explain itself — not because concepts were added as an afterthought but because they determine how knowledge is represented, partitioned, and preserved.
The analogy to human cognition is suggestive. We do not store knowledge as undifferentiated neural patterns overwritten by each new experience. We organize knowledge through concepts, categories, and relationships — and this organization is precisely what allows us to learn new things without forgetting old ones. Whether neural networks can replicate this trick through engineered concept bottlenecks remains an open empirical question, but the evidence from CVPR 2025 suggests the approach is productive.