Quantifying Correlations of Machine Learning Models

Yuanyuan Li, Neeraj Sarna, Yang Lin

Abstract

As machine learning models become increasingly integrated into safety-critical systems, the assumption that using multiple models provides a "safety net" is being challenged. This paper investigates the phenomenon of error correlation—the tendency for different models to fail on the same data points. By analyzing various architectures, datasets, and foundation models, the research demonstrates that shared training data and common architectural lineages create systemic risks, leading to synchronized failures across supposedly diverse AI systems.

1. Introduction

In fields such as healthcare, autonomous driving, and finance, redundancy is a standard safety protocol. The prevailing logic suggests that if one AI model fails, a secondary model might succeed, provided their errors are independent. However, the modern AI landscape is characterized by "homogenization": a trend where the industry converges on a few massive foundation models (e.g., Llama, ResNet) and a handful of public datasets (e.g., ImageNet, CIFAR).

This paper addresses whether this convergence leads to shared blind spots. If two independent applications inherit the same underlying biases or limitations from a shared foundation model, they may fail simultaneously when presented with the same challenging input. This study provides a mathematical framework to quantify exactly how much these models "fail in sync."

2. Methodology

To quantify error correlation, the researchers defined metrics for both regression and classification tasks:

  • Pearson’s Correlation Coefficient: Employed for continuous variables in regression to measure the relationship between prediction errors.
  • $ϕ_K$ Measure: A versatile metric used for classification tasks to capture non-linear relationships across categorical and numerical variables.

2.1 Correlation Scenarios

The study evaluated three primary deployment scenarios to identify the drivers of correlated failures:

  1. Shared Datasets: Different model architectures (e.g., Random Forest vs. Neural Networks) trained on the identical dataset.
  2. Overlapping Features: Models with the same architecture trained on datasets sharing key variables.
  3. Foundation Model Inheritance: Different models fine-tuned for distinct tasks but originating from the same pre-trained foundation model.

3. Experimental Setup

The researchers utilized a diverse suite of benchmarks to ensure the generalizability of their findings:

  • Tabular Data: California Housing price prediction.
  • Image Classification: CIFAR-10, EuroSAT (satellite imagery), MNIST, and Fashion-MNIST.
  • Text Classification: Financial news sentiment analysis, emotion-balanced tweets, and AG News.

4. Key Insights and Results

4.1 The Impact of Algorithmic Families

The research found that models based on similar logic—such as tree-based methods like Random Forest and XGBoost—exhibit very high error correlations. When one model in a specific "family" struggles with a data pattern, its relatives are likely to fail as well. In many tests, correlation coefficients exceeded 0.7, indicating a strong systemic link between failures.

4.2 "Heavy Hitter" Features in Tabular Data

In tabular datasets, the presence of highly predictive features (e.g., "Median Income" in housing data) acts as a primary driver of correlation. Any model that utilizes these dominant features tends to make similar judgments—and similar mistakes—regardless of the underlying architecture.

4.3 Foundation Model Inheritance and LLMs

The most significant finding involves foundation models. Fine-tuned Large Language Models (LLMs) showed more widespread error correlation than convolutional neural networks (CNNs). Because most modern LLMs share the Transformer architecture and similar pre-training regimes, they exhibit consistent error patterns even when fine-tuned for entirely different tasks. The researchers observed that if a specific foundation model (e.g., VGG16 vs. ResNet) was more prone to error on one dataset, that trend remained remarkably consistent across almost all other datasets tested.

5. Discussion and Systemic Risk

The results suggest that achieving true "diversity" in AI ensembles is significantly more difficult than previously assumed. Simply swapping a model architecture is insufficient to guarantee independent failures if the underlying data or the foundation model remains the same. This creates a systemic risk where multiple AI systems, intended to act as checks on one another, may inherit the same vulnerabilities.

6. Conclusion and Future Work

While this study provides robust empirical evidence of error correlation, the authors note that it is primarily an empirical investigation. Future research is required to derive formal mathematical bounds for these correlations under different distributional assumptions. Additionally, the development of risk management strategies to mitigate "cascading failures" in multi-model systems remains a critical priority for the safety-critical deployment of AI.

Summary updated

BibTeX
@inproceedings{li2025quantifying,
  author = {Yuanyuan Li and Neeraj Sarna and Yang Lin},
  title = {{Quantifying Correlations of Machine Learning Models}},
  booktitle = {ICSTW},
  year = {2025},
  doi = {10.1109/icstw64639.2025.10962482},
}
Figure