The Quest for Universal Master Key Filters in DS-CNNs

2025-09-15

AI Summary

This work investigates whether universal convergence patterns exist in filter learning for depthwise separable convolutional neural networks (DS-CNNs). To address this, we conduct a systematic unsupervised search across diverse architectures and datasets, combined with linear transformation analysis (ax + b) of learned filters. We discover that DS-CNNs consistently converge to a compact set of only eight frozen “principal key filters,” whose shapes closely approximate Gaussian functions, the Difference of Gaussians (DoG), and their derivatives—structures highly consistent with biological visual receptive fields. This reveals that DS-CNNs intrinsically rely on a minimal set of fundamental operators. Initializing DS-CNNs with these eight universal filters achieves over 80% top-1 accuracy on ImageNet without further filter optimization. Moreover, in few-shot learning settings, such initialization significantly outperforms conventional models with thousands of trainable parameters. These findings provide a novel mechanistic explanation for transferability and generalization in DS-CNNs.

Technology Category

• Computer Vision: Learning & Optimization for CV
• Machine Learning: Deep Learning Theory
• Search and Optimization: Learning to Search

Application Category

• Graph Algorithms and Modeling for the Web: Graph neural networks and deep learning approaches for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web

Abstract

A recent study has proposed the "Master Key Filters Hypothesis" for convolutional neural network filters. This paper extends this hypothesis by radically constraining its scope to a single set of just 8 universal filters that depthwise separable convolutional networks inherently converge to. While conventional DS-CNNs employ thousands of distinct trained filters, our analysis reveals these filters are predominantly linear shifts (ax+b) of our discovered universal set. Through systematic unsupervised search, we extracted these fundamental patterns across different architectures and datasets. Remarkably, networks initialized with these 8 unique frozen filters achieve over 80% ImageNet accuracy, and even outperform models with thousands of trainable parameters when applied to smaller datasets. The identified master key filters closely match Difference of Gaussians (DoGs), Gaussians, and their derivatives, structures that are not only fundamental to classical image processing but also strikingly similar to receptive fields in mammalian visual systems. Our findings provide compelling evidence that depthwise convolutional layers naturally gravitate toward this fundamental set of spatial operators regardless of task or architecture. This work offers new insights for understanding generalization and transfer learning through the universal language of these master key filters.

Problem

• Identifies universal master key filters in DS-CNNs
• Shows networks converge to 8 fundamental filter patterns
• Links these filters to biological vision principles

Innovation

• Single set of 8 universal filters
• Unsupervised search across architectures
• Frozen filters achieving ImageNet accuracy