[Paper Review] Not quite unreasonable effectiveness of machine learning algorithms

Egor Illarionov, Roman Khudorozhkov|arXiv (Cornell University)|Apr 7, 2018

Anomaly Detection Techniques and Applications3 references3 citations

TL;DR

This paper argues that state-of-the-art machine learning performance on benchmarks like MNIST and MIT-BIH Atrial Fibrillation is driven by low data variability and high pattern similarity, not model capacity. It shows that only 5–10% of training data is needed for near-optimal accuracy, and that learned features can identify patients, revealing hidden structure that standard metrics fail to capture.

ABSTRACT

State-of-the-art machine learning algorithms demonstrate close to absolute performance in selected challenges. We provide arguments that the reason can be in low variability of the samples and high effectiveness in learning typical patterns. Due to this fact, standard performance metrics do not reveal model capacity and new metrics are required for the better understanding of state-of-the-art.

Motivation & Objective

• To investigate why modern machine learning models achieve near-perfect performance on standard benchmarks like MNIST and MIT-BIH ECG.
• To challenge the assumption that high accuracy reflects model capacity, arguing instead that data structure plays a dominant role.
• To demonstrate that low sample variability and high pattern similarity across data points enable rapid generalization.
• To show that standard evaluation metrics fail to reveal model capacity, necessitating new evaluation frameworks.
• To propose that model performance on highly structured datasets may be misleading without deeper analysis of data complexity and representational learning.

Proposed method

• Trained ResNet18, Inception v1, and fully connected neural networks on MNIST with varying train set sizes to measure accuracy saturation.
• Used logarithmic scaling of training set ratios to detect early saturation points in model convergence.
• Applied global max-pooling on trained models to extract feature vectors from ECG segments for patient identification.
• Trained a random forest classifier on ECG feature vectors to predict patient identity, using test set performance as a proxy for pattern consistency.
• Evaluated model performance on MIT-BIH Atrial Fibrillation dataset using macro F1-score across varying training set proportions.
• Analyzed model variance and generalization by testing on 100 random subsets of training data and measuring accuracy distribution.

Experimental results

Research questions

• RQ1 Why do state-of-the-art models achieve near-100% accuracy on datasets like MNIST and MIT-BIH Atrial Fibrillation?
• RQ2 To what extent is high performance due to low variability in data patterns rather than model capacity?
• RQ3 Can model generalization be explained by the existence of a limited number of dominant patterns in the data?
• RQ4 Can learned representations from deep models be used to reconstruct higher-level data structure, such as patient identity in ECG signals?
• RQ5 Why do standard evaluation metrics fail to reveal the true capacity of modern machine learning models?

Key findings

• On the MNIST dataset, ResNet18 achieves over 97% test accuracy using only 7% of the training data (4.2K samples), with 99% accuracy at 10%.
• For the Inception v1 model, over 97% accuracy is achieved on all 100 random 10% training subsets, indicating robustness to data sampling.
• On the MIT-BIH Atrial Fibrillation dataset, F1-score saturates at 25% of the training data, reaching 0.988 F1-score on test set.
• A random forest classifier trained on global max-pooling features from the ECG model achieves 95% accuracy in identifying the patient of origin.
• The ability to identify patients from ECG feature vectors confirms that data patterns are highly consistent and patient-specific, supporting low variability hypothesis.
• The results suggest that standard metrics like accuracy and F1-score do not adequately reflect model capacity, especially in low-variability datasets.