When we look at a "red tomato" and a glass of "red wine," our brains instantly recognize the shared concept of redness, despite the fact that the actual RGB pixels, textures, and light reflections are completely different. Traditional computer vision models, however, struggle with this nuance.
In Compositional Zero-Shot Learning (CZSL), the goal is to recognize novel combinations of known primitives—such as identifying an old tiger after only seeing young tigers and old cars. Early approaches assumed that attributes (like "old" or "red") and objects (like "tiger" or "tomato") could be cleanly disentangled into independent, invariant feature spaces.
But as Ishan Misra et al. demonstrated in their foundational work, From Red Wine to Red Tomato: Composition with Context, treating attributes as static properties fails to capture real-world complexity. Visual attributes are fundamentally context-dependent; their appearance shifts dynamically based on the object they modify.
The Failure of Naive Disentanglement
Why does standard feature disentanglement fail? Consider the geometric intuition of mapping concepts to a vector space. If a model learns a single prototype vector for the attribute "old" based on images of old cars, it will look for metallic rust, peeling paint, or vintage chassis designs. When presented with an "old tiger," the model fails because an old animal manifests through entirely different visual cues, such as graying fur or slower movement patterns.
As Muli Yang et al. point out in Learning Unseen Concepts via Hierarchical Decomposition and Composition, forcing a model to decouple an attribute from its object without accounting for this contextuality results in a loss of essential visual nuances. Similarly, Guangyue Xu et al. observe in Zero-Shot Compositional Concept Learning that this variance creates severe feature inconsistency, confusing classifiers when they encounter unseen pairs.
To solve this, a major research direction has emerged that moves away from static attribute representations toward dynamic, context-aware modeling.
Attributes as Mathematical Operators
One of the earliest conceptual shifts came from treating attributes not as points in a semantic space, but as functions. In Attributes as Operators: Factorizing Unseen Attribute-Object Compositions, Tushar Nagarajan et al. proposed modeling attributes as linear operators (matrices) that actively transform an object's base representation.
$$\mathbf{v}_{\text{composition}} = T_{\text{attribute}}(\mathbf{v}_{\text{object}})$$
Under this paradigm, the attribute "sliced" is a matrix that alters the geometric and textural representation of an object vector. Slicing an apple yields a completely different visual transformation than slicing a loaf of bread. By training the model to learn the transformation function rather than a static visual prototype, the network naturally adapts to the underlying object's geometry.
Dynamic Conditioning and Cascaded Architectures
Instead of relying on rigid matrix transformations, modern deep networks use hyper-networks and attention mechanisms to compute attribute representations on the fly.
The Conditional Attribute Network (CANet), introduced by Lingqiao Liu et al. in Learning Conditional Attributes for Compositional Zero-Shot Learning, uses a hyper-learner to generate custom attribute embeddings conditioned directly on the specific object and image context. Rather than looking for a universal "wet" feature, the model generates a "wet-conditioned-on-cat" classifier or a "wet-conditioned-on-pavement" classifier.
[Image Context] + [Object Representation] ──> (Hyper-Learner) ──> [Dynamic Attribute Embedding]
Taking this sequential dependency further, Yanyi Zhang et al. developed CSCNET: Class-Specified Cascaded Network for Compositional Zero-Shot Learning. Their architecture uses a cascaded approach: the model first predicts the object category to establish a contextual anchor, and then uses that prediction to guide and constrain the recognition of the modifying attribute.
Embracing Polysemy via Multi-Representation Spaces
An attribute can be so highly context-dependent that it exhibits primitive polysemy—meaning it possesses multiple distinct visual meanings. To handle this, researchers have begun representing single primitives using multiple distinct vectors or expert networks.
In Decoupling Primitive with Experts: Dynamic Feature Alignment for Compositional Zero-Shot Learning, Xiao Zhang et al. introduced a Mixture-of-Experts (MoE) framework. Instead of forcing the word "old" into a single vector, the model maintains multiple semantic variants (experts) and dynamically aligns the visual features of the image with the most relevant expert at inference time.
This strategy is mirrored by Dongyao Jiang et al. in MRSP: Learn Multi-representations of Single Primitive for Compositional Zero-Shot Learning. The authors model each primitive as a "union" of multiple embeddings. When evaluating an unseen composition, the model selects the representation path that best fits the paired concept, successfully bridging the gap between disparate contexts like a "historical town" and a "senior tiger."
The Universality of Context-Dependency
This context-dependent behavior is not unique to adjectives; it extends across all layers of visual compositionality:
- Relative Modifiers: Adjectives like "small" or "large" possess no absolute physical scale. As noted by Yuval Atzmon et al. in A causal view of compositional zero-shot recognition, a "small bear" is still orders of magnitude larger than a "large bird." Models must evaluate these attributes relative to a category-specific norm.
- Action Verbs: In video understanding, actions are highly dependent on their targets. Dongyao Jiang et al. demonstrated in Beyond Image Classification: A Video Benchmark and Dual-Branch Hybrid Discrimination Framework for Compositional Zero-Shot Learning that the physical motion of "opening" looks entirely different when applied to a book versus a door.
- Object Parts: Even localized components change based on the parent entity. As shown by Vignesh Ramanathan et al. in the PACO dataset (PACO: Parts and Attributes of Common Objects), a generic label like "leg" suffers from extreme visual variance depending on whether it belongs to a chair or a dog.
By shifting from rigid, isolated feature extraction to fluid, context-conditioned representations, this research direction enables computer vision models to move past rote memorization and achieve true, human-like compositional generalization.