Presented at IJCAI-2025, this paper introduces Homogeneous Group Representation Learning (HGRL), a novel framework designed to solve the "conditional dependency" problem in Compositional Zero-Shot Learning (CZSL). By mimicking human hierarchical reasoning, the authors move away from the binary choice of learning one-size-fits-all prototypes versus instance-specific features, instead opting for a "middle ground" of group-based representations.
Abstract
Compositional Zero-Shot Learning (CZSL) aims to recognize novel state-object combinations (e.g., "green truck") after training on different pairs (e.g., "red truck", "green car"). A major hurdle is "conditional dependency," where a state's visual appearance depends heavily on the object it modifies. This paper proposes Homogeneous Group Representation Learning (HGRL), which adaptively clusters categories into groups to balance semantic transferability and discriminability. HGRL employs group-specific experts and prompts, achieving state-of-the-art results on MIT-States, UT-Zappos, and C-GQA benchmarks.
The Challenge of Conditional Dependency
In CZSL, the primary hurdle is domain bias. A state like "old" looks very different when applied to a "dog" (pale skin, bloated body) versus a "bus" (corroded metal, dirty surfaces). Existing models usually take one of two extreme paths:
- All-to-one: Learning a single, "pure" representation for a state. This aids generalization (transferability) but loses the specific details needed to distinguish objects (discriminability).
- One-to-one: Modeling the specific relationship between every state and object. This is accurate for seen pairs but fails to generalize to unseen ones due to overfitting.
The authors argue that humans use hierarchical clustering. We recognize that "old dogs" and "old cats" share visual properties as animals, while "old cars" belong to a different mechanical group. HGRL exploits this latent homogeneity to create a more robust learning framework.
Methodology: The HGRL Framework
HGRL integrates three core components on top of a frozen CLIP backbone to manage group-aware learning:
- Group-Aware Visual Representation (GAVR): Utilizing a Mixture-of-Experts (MoE) approach, the model routes features through specific "expert" sub-networks based on group membership. This preserves group-specific details without over-fitting to individual instances.
- Decoupled Group Prompt (DGP): DGP adds an exclusive "context prompt" for each group. This helps the model learn distributed prototypes that capture the nuances of different clusters (e.g., a prompt for "animal states" vs. "mechanical states").
- Group-Aware Pair Enhancement (GAPE): To increase robustness, the model augments data by combining features from "compatible" group members. For example, an "old dog" feature might be enhanced by an "old cat" but not an "old car," ensuring semantically meaningful augmentation.
To guide this clustering without manual labels, the authors use a Text Co-occurrence Probability Graph (TCPG) based on GloVe word embeddings to calculate the likelihood that categories belong to the same super-class.
Experimental Results
The authors evaluated HGRL on three standard CZSL benchmarks: MIT-States, UT-Zappos, and C-GQA. HGRL achieved State-of-the-Art (SOTA) results across all datasets:
- UT-Zappos: Outperformed the second-best method by 2.3% in AUC and 2.5% in Harmonic Mean (HM) in closed-world settings. In open-world settings, the lead increased to 4.2% (AUC).
- C-GQA: On this large-scale dataset, HGRL outperformed previous methods by 1.9% (AUC) and 2.6% (HM).
- Ablation Studies: Removing GAVR or DGP led to significant performance drops, confirming that both visual and textual group-awareness are critical for success.
Conclusion and Future Work
HGRL successfully balances the "seesaw" of transferability and discriminability in CZSL. However, the model remains sensitive to the predefined number of groups and relies on external word embeddings like GloVe. Future research may focus on dynamic methods for determining optimal group numbers to further enhance the model's flexibility and performance.
Summary updated