Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

LayerMix: Enhanced Data Augmentation through Fractal Integration for Robust Deep Learning

Hafiz Mughees Ahmad, Dario Morle, Afshin Rahimi

TL;DR

LayerMix tackles the vulnerability of vision models to distribution shifts by introducing a structured, fractal-based data augmentation pipeline with covariance between augmentation stages. The method blends fractals with images and reweights blending operations to balance semantic preservation, diversity, and robustness, achieving state-of-the-art improvements in corruption robustness, prediction consistency, calibration, and adversarial resilience across CIFAR and ImageNet benchmarks. The study provides a theoretical framing, practical pipeline, extensive experiments, and open-source code, demonstrating that carefully designed, label-preserving fractal blending can yield Pareto improvements over baseline augmentations without sacrificing efficiency. This work advances robust AI by delivering a scalable augmentation strategy that enhances generalization and safety metrics, with clear directions for adaptive fractal generation and broader application.

Abstract

Deep learning models have demonstrated remarkable performance across various computer vision tasks, yet their vulnerability to distribution shifts remains a critical challenge. Despite sophisticated neural network architectures, existing models often struggle to maintain consistent performance when confronted with Out-of-Distribution (OOD) samples, including natural corruptions, adversarial perturbations, and anomalous patterns. We introduce LayerMix, an innovative data augmentation approach that systematically enhances model robustness through structured fractal-based image synthesis. By meticulously integrating structural complexity into training datasets, our method generates semantically consistent synthetic samples that significantly improve neural network generalization capabilities. Unlike traditional augmentation techniques that rely on random transformations, LayerMix employs a structured mixing pipeline that preserves original image semantics while introducing controlled variability. Extensive experiments across multiple benchmark datasets, including CIFAR-10, CIFAR-100, ImageNet-200, and ImageNet-1K demonstrate LayerMixs superior performance in classification accuracy and substantially enhances critical Machine Learning (ML) safety metrics, including resilience to natural image corruptions, robustness against adversarial attacks, improved model calibration and enhanced prediction consistency. LayerMix represents a significant advancement toward developing more reliable and adaptable artificial intelligence systems by addressing the fundamental challenges of deep learning generalization. The code is available at https://github.com/ahmadmughees/layermix.

LayerMix: Enhanced Data Augmentation through Fractal Integration for Robust Deep Learning

TL;DR

LayerMix tackles the vulnerability of vision models to distribution shifts by introducing a structured, fractal-based data augmentation pipeline with covariance between augmentation stages. The method blends fractals with images and reweights blending operations to balance semantic preservation, diversity, and robustness, achieving state-of-the-art improvements in corruption robustness, prediction consistency, calibration, and adversarial resilience across CIFAR and ImageNet benchmarks. The study provides a theoretical framing, practical pipeline, extensive experiments, and open-source code, demonstrating that carefully designed, label-preserving fractal blending can yield Pareto improvements over baseline augmentations without sacrificing efficiency. This work advances robust AI by delivering a scalable augmentation strategy that enhances generalization and safety metrics, with clear directions for adaptive fractal generation and broader application.

Abstract

Deep learning models have demonstrated remarkable performance across various computer vision tasks, yet their vulnerability to distribution shifts remains a critical challenge. Despite sophisticated neural network architectures, existing models often struggle to maintain consistent performance when confronted with Out-of-Distribution (OOD) samples, including natural corruptions, adversarial perturbations, and anomalous patterns. We introduce LayerMix, an innovative data augmentation approach that systematically enhances model robustness through structured fractal-based image synthesis. By meticulously integrating structural complexity into training datasets, our method generates semantically consistent synthetic samples that significantly improve neural network generalization capabilities. Unlike traditional augmentation techniques that rely on random transformations, LayerMix employs a structured mixing pipeline that preserves original image semantics while introducing controlled variability. Extensive experiments across multiple benchmark datasets, including CIFAR-10, CIFAR-100, ImageNet-200, and ImageNet-1K demonstrate LayerMixs superior performance in classification accuracy and substantially enhances critical Machine Learning (ML) safety metrics, including resilience to natural image corruptions, robustness against adversarial attacks, improved model calibration and enhanced prediction consistency. LayerMix represents a significant advancement toward developing more reliable and adaptable artificial intelligence systems by addressing the fundamental challenges of deep learning generalization. The code is available at https://github.com/ahmadmughees/layermix.
Paper Structure (31 sections, 10 equations, 8 figures, 10 tables)

This paper contains 31 sections, 10 equations, 8 figures, 10 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Affinity and Diversity
  4. Augmentation Pipelines
  5. Robust Deep Learning
  6. Methodology
  7. Pipeline Covariance
  8. Fractals Augmentation
  9. Reweighted Blending
  10. Augmentation Pipeline
  11. Experiments
  12. Evaluation Benchmarks and Metrics
  13. Corruption Robustness:
  14. Prediction Consistency:
  15. Adversarial Robustness:
  16. ...and 16 more sections

Figures (8)

  • Figure 1: Comparison of different augmentation pipelines.
  • Figure 2: Complete Pipeline of LayerMix. The resulting image produced by the LayerMix pipeline is uniformly selected from samples 1, 2, and 3 produced by the pipeline. All Aug blocks are correlated by the covariance structure described in \ref{['sec:meth.cov']}. All Blend blocks are independent and sample from the re-weighted blending mixture distribution described in \ref{['sec:meth.blend']}.
  • Figure 3: Corruption Error values for various methods on CIFAR-100-C.
  • Figure 4: Individual Corruption Error (CE) values for various methods on ImageNet-200-C
  • Figure 6: Samples of grayscale fractals used in LayerMix.
  • ...and 3 more figures