Interpretable Compositional Representations for Robust Few-Shot Generalization.

Area of Science:

• Computer Vision
• Cognitive Science
• Machine Learning

Background:

• Current deep learning models struggle with low-shot generalization and adversarial robustness.
• Human object recognition relies on decomposing objects into parts and a compact concept vocabulary.

Purpose of the Study:

• To introduce Recognition as Part Composition (RPC), a novel image encoding method inspired by human cognition.
• To demonstrate RPC's effectiveness in improving generalization and robustness in machine learning tasks.
• To leverage RPC's interpretability for synthetic data generation.

Main Methods:

• RPC decomposes images into salient parts.
• Each part is encoded as a mixture of learned prototypes representing concepts.
• A classifier is trained using the RPC image encoder.

Main Results:

• RPC enhances performance in zero-shot learning, few-shot learning, and unsupervised domain adaptation.
• Classifiers utilizing RPC exhibit significant robustness against adversarial attacks.
• Crowdsourcing experiments confirm the interpretability of RPC encodings by humans.

Conclusions:

• Cognition-inspired image encoding offers a promising alternative to traditional deep learning methods.
• RPC provides a robust and interpretable framework for image representation learning.
• RPC enables novel applications, such as generating synthetic annotations for evaluating zero-shot learning models.