Just Leaf It: Accelerating Diffusion Classifiers with Hierarchical Class Pruning

2024-11-18

arXiv.org

AI Summary

Diffusion classifiers require exhaustive evaluation across all classes for each input, incurring prohibitive inference overhead and hindering practical deployment. To address this, we propose the first hierarchical semantic structure-integrated diffusion classification framework, introducing a label-tree-based adaptive pruning method. Our approach performs coarse-grained filtering at high-level categories, focuses subsequent computation on promising subtrees, and employs confidence-driven hierarchical backtracking—all without retraining and enabling plug-and-play integration. It supports flexible speed-accuracy trade-offs and substantially reduces the number of candidate classes evaluated per inference. Experiments on multiple benchmark datasets demonstrate up to 60% inference speedup while maintaining or even improving classification accuracy. The core contribution lies in the principled unification of Bayesian diffusion classification with hierarchical label modeling, establishing a novel paradigm for efficient diffusion-based classification in large-scale scenarios.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Multi-class/Multi-label Learning & Extreme Classification
• Search and Optimization: Learning to Search

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Search and Retrieval-Augmented AI: Efficiency and scalability of Web search engines
• Web Mining and Content Analysis: Content-based information diffusion

Abstract

Diffusion models, celebrated for their generative capabilities, have recently demonstrated surprising effectiveness in image classification tasks by using Bayes' theorem. Yet, current diffusion classifiers must evaluate every label candidate for each input, creating high computational costs that impede their use in large-scale applications. To address this limitation, we propose a Hierarchical Diffusion Classifier (HDC) that exploits hierarchical label structures or well-defined parent-child relationships in the dataset. By pruning irrelevant high-level categories and refining predictions only within relevant subcategories (leaf nodes and sub-trees), HDC reduces the total number of class evaluations. As a result, HDC can speed up inference by as much as 60% while preserving and sometimes even improving classification accuracy. In summary, our work provides a tunable control mechanism between speed and precision, making diffusion-based classification more feasible for large-scale applications.

Problem

• Reduces computational cost in diffusion classifiers
• Exploits hierarchical label structures for efficiency
• Improves speed and accuracy in image classification

Innovation

• Hierarchical label structures for efficient classification
• Pruning irrelevant categories to reduce computations
• Tunable control mechanism balancing speed and accuracy