Grouped Discrete Representation for Object-Centric Learning

2024-11-04

arXiv.org

AI Summary

Existing VAE-based object-centric learning (OCL) approaches treat features as atomic units and employ scalar discrete codes, leading to inadequate attribute-level similarity modeling and limited representational interpretability. To address this, we propose Grouped Discrete Representation (GDR): a novel discrete encoding paradigm that semantically disentangles latent features into attribute-specific groups and replaces scalar codes with tuple-based indices—enabling structurally interpretable and attribute-composable representations. GDR integrates channel-wise grouped decomposition with tuple-indexed discretization, and is architecture-agnostic, supporting VAEs, Transformers, and diffusion models. On multiple benchmarks, GDR significantly improves unsupervised object discovery performance. Visualizations confirm its enhanced object separation and reconstruction fidelity. This work establishes the first attribute-decoupled discretization framework for OCL, advancing both interpretability and compositional reasoning in unsupervised representation learning.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Representation Learning
• Search and Optimization: Mixed Discrete/Continuous Search

Application Category

• Graph Algorithms and Modeling for the Web: Representation, reconstruction, and subgraph or motif discovery in Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking

Abstract

Object-Centric Learning (OCL) can discover objects in images or videos by simply reconstructing the input. For better object discovery, representative OCL methods reconstruct the input as its Variational Autoencoder (VAE) intermediate representation, which suppresses pixel noises and promotes object separability by discretizing continuous super-pixels with template features. However, treating features as units overlooks their composing attributes, thus impeding model generalization; indexing features with scalar numbers loses attribute-level similarities and differences, thus hindering model convergence. We propose extit{Grouped Discrete Representation} (GDR) for OCL. We decompose features into combinatorial attributes via organized channel grouping, and compose these attributes into discrete representation via tuple indexes. Experiments show that our GDR improves both Transformer- and Diffusion-based OCL methods consistently on various datasets. Visualizations show that our GDR captures better object separability.

Problem

• Enhance object separability in Object-Centric Learning
• Address loss of attribute-level similarities in discrete representations
• Improve generalization and convergence in feature decomposition

Innovation

• Decompose features into combinatorial attributes
• Quantize features via tuple code indexes
• Organize channel grouping for better generalization