A Comparison of the Delta Method and the Bootstrap in Deep Learning\n Classification

2021/07/04 by Geir Kjetil Nilsen, Nilsen, Geir K., Antonella Z. Munthe-Kaas +5
Computer Science · Materials Science · #Adversarial Robustness in Machine Learning #Anomaly Detection Techniques and Applications #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning (stat.ML) #Machine Learning and Data Classification #Machine Learning in Materials Science

paper · pdf · doi:10.48550/arxiv.2107.01606

openalex publication_date 2021/07/04 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28

Abstract

We validate the recently introduced deep learning classification adapted\nDelta method by a comparison with the classical Bootstrap. We show that there\nis a strong linear relationship between the quantified predictive epistemic\nuncertainty levels obtained from the two methods when applied on two\nLeNet-based neural network classifiers using the MNIST and CIFAR-10 datasets.\nFurthermore, we demonstrate that the Delta method offers a five times\ncomputation time reduction compared to the Bootstrap.\n

Citations

• A Practical Bayesian Framework for Backpropagation Networks
• Bootstrap Methods: Another Look at the Jackknife
• Comprehensive Review of Neural Network-Based Prediction Intervals and New Advances
• Deep Exploration via Bootstrapped DQN

Related

• Epistemic Uncertainty Quantification in Deep Learning Classification by the Delta Method
• Transgressing the boundaries: towards a rigorous understanding of deep learning and its (non-)robustness
• Providing theoretical learning guarantees to Deep Learning Networks
• A Bootstrap Algorithm for Fast Supervised Learning
• A Survey of Techniques All Classifiers Can Learn from Deep Networks: Models, Optimizations, and Regularization