Disentangling Before Composing: Learning Invariant Disentangled Features for Compositional Zero-Shot Learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Compositional Zero-Shot Learning (CZSL) aims to recognize novel combinations of attributes and objects.
• Existing methods often struggle with spurious correlations and bias towards seen compositions due to shared visual representations.

Purpose of the Study:

• To address the limitations of current CZSL methods by reframing it as an out-of-distribution generalization problem.
• To develop a method that learns object-invariant and attribute-invariant features for reliable recognition of novel compositions.

Main Methods:

• Proposes an invariant feature learning framework to align different domains (objects and attributes) at representation and gradient levels.
• Introduces an "encoding-reshuffling-decoding" process to disentangle attribute and object features, preventing spurious correlations by creating synthetic features.

Main Results:

• The proposed method effectively disentangles features representing independent factors of attributes and objects.
• Achieves state-of-the-art or competitive performance in both closed-world and open-world CZSL scenarios.

Conclusions:

• The invariant feature learning framework and feature disentanglement strategy significantly improve generalization in CZSL.
• This approach offers a robust solution for recognizing unseen attribute-object compositions by mitigating biases from seen data.