FeatInv: Spatially resolved mapping from feature space to input space using conditional diffusion models

2025-05-27

AI Summary

Deep neural network feature spaces lack interpretability, particularly due to the absence of an exact, pixel-level inverse mapping from spatial feature maps back to input images. To address this, we propose Spatially Aligned Conditional Diffusion (SACD), the first method to condition high-fidelity pre-trained diffusion models—pixel-wise—on spatially resolved feature maps extracted from CNNs or Vision Transformers. SACD enables a probabilistic, high-fidelity, and sampleable inverse mapping from features to inputs. Crucially, it supports concept-guided visualization (“concept steering”) and feature disentanglement analysis without fine-tuning or restrictive approximations. We evaluate SACD on multiple ImageNet-pretrained classifiers, demonstrating superior reconstruction quality and robustness over existing feature inversion approaches. Our method establishes a new paradigm for interpretable deep representation learning by bridging generative modeling and feature-space analysis in a principled, scalable manner.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Machine Learning: Deep Generative Models & Autoencoders
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Web Mining and Content Analysis: Large pretrained models with web data
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs

Abstract

Internal representations are crucial for understanding deep neural networks, such as their properties and reasoning patterns, but remain difficult to interpret. While mapping from feature space to input space aids in interpreting the former, existing approaches often rely on crude approximations. We propose using a conditional diffusion model - a pretrained high-fidelity diffusion model conditioned on spatially resolved feature maps - to learn such a mapping in a probabilistic manner. We demonstrate the feasibility of this approach across various pretrained image classifiers from CNNs to ViTs, showing excellent reconstruction capabilities. Through qualitative comparisons and robustness analysis, we validate our method and showcase possible applications, such as the visualization of concept steering in input space or investigations of the composite nature of the feature space. This approach has broad potential for improving feature space understanding in computer vision models.

Problem

• Mapping feature space to input space for interpretability
• Using conditional diffusion models for high-fidelity reconstruction
• Improving understanding of feature space in vision models

Innovation

• Conditional diffusion models for feature mapping
• Probabilistic mapping from feature to input space
• High-fidelity reconstruction across CNN and ViT models