Spooky Action at a Distance: Normalization Layers Enable Side-Channel Spatial Communication

2025-07-07

AI Summary

Normalization layers (e.g., BatchNorm, LayerNorm) in convolutional networks implicitly enable long-range spatial communication via cross-location statistic computation—bypassing the local receptive field constraint and forming an unrecognized side channel. This effect risks unintended information leakage and structural distortion in tasks requiring strictly bounded receptive fields, such as diffusion-based trajectory generation. Method: Through theoretical analysis and controlled experiments on localization tasks, we systematically characterize how normalization layers progressively strengthen long-range spatial dependency modeling across iterative convolutional layers. Contribution/Results: (1) We identify normalization layers as implicit global communication channels; (2) we quantify their non-negligible expansion of effective receptive field boundaries; and (3) we highlight their underappreciated risks in spatiotemporally sensitive applications, providing principled design constraints for receptive-field-restricted modeling.

Technology Category

• Computer Vision: Diffusion Models for Vision
• Natural Language Processing: Safety and Robustness
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Social Networks and Social Media: Influence propagation, information diffusion, and the prediction on networks
• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Web Mining and Content Analysis: Normalization, clustering, classification, and summarization of Web text

Abstract

This work shows that normalization layers can facilitate a surprising degree of communication across the spatial dimensions of an input tensor. We study a toy localization task with a convolutional architecture and show that normalization layers enable an iterative message passing procedure, allowing information aggregation from well outside the local receptive field. Our results suggest that normalization layers should be employed with caution in applications such as diffusion-based trajectory generation, where maintaining a spatially limited receptive field is crucial.

Problem

• Normalization layers enable cross-spatial communication in tensors
• Normalization facilitates message passing beyond local receptive fields
• Caution needed in diffusion tasks requiring limited receptive fields

Innovation

• Normalization layers enable spatial communication
• Message passing beyond local receptive fields
• Caution needed in diffusion-based applications