In this paper, error estimates of classification Random Forests are quantitatively assessed. Based on the initial theoretical framework built by Bates et al. (2023), the true error rate and expected error rate are theoretically and empirically investigated in the context of a variety of error estimation methods common to Random Forests. We show that in the classification case, Random Forests' estimates of prediction error is closer on average to the true error rate instead of the average prediction error. This is opposite the findings of Bates et al. (2023) which were given for logistic regression. We further show that this result holds across different error estimation strategies such as cross-validation, bagging, and data splitting.

In the task of comparing two classification algorithms, the widely-used McNemar's test aims to infer the presence of a significant difference between the error rates of the two classification algorithms. However, the power of the conventional McNemar's test is usually unpromising because the hold-out (HO) method in the test merely uses a single train-validation split that usually produces a highly varied estimation of the error rates. In contrast, a cross-validation (CV) method repeats the HO method in multiple times and produces a stable estimation. Therefore, a CV method has a great advantage to improve the power of McNemar's test. Among all types of CV methods, a block-regularized 5$\times$2 CV (BCV) has been shown in many previous studies to be superior to the other CV methods in the comparison task of algorithms because the 5$\times$2 BCV can produce a high-quality estimator of the error rate by regularizing the numbers of overlapping records between all training sets. In this study, we compress the 10 correlated contingency tables in the 5$\times$2 BCV to form an effective contingency table. Then, we define a 5$\times$2 BCV McNemar's test on the basis of the effective contingency table. We demonstrate the reasonable type I error and the promising power of the proposed 5$\times$2 BCV McNemar's test on multiple simulated and real-world data sets.

We present a new approach to bounding the true error rate of a continuous valued classifier based upon PAC-Bayes bounds. The method first con- structs a distribution over classifiers by determining how sensitive each parameter in the model is to noise. The true error rate of the stochastic classifier found with the sensitivity analysis can then be tightly bounded using a PAC-Bayes bound. In this paper we demonstrate the method on artificial neural networks with results of a order of magnitude im- provement vs. the best deterministic neural net bounds.

This technical report describes a new feature of the CleverHans library called "attack bundling". Many papers about adversarial examples present lists of error rates corresponding to different attack algorithms. A common approach is to take the maximum across this list and compare defenses against that error rate. We argue that a better approach is to use attack bundling: the max should be taken across many examples at the level of individual examples, then the error rate should be calculated by averaging after this maximization operation. Reporting the bundled attacker error rate provides a lower bound on the true worst-case error rate. The traditional approach of reporting the maximum error rate across attacks can underestimate the true worst-case error rate by an amount approaching 100\% as the number of attacks approaches infinity. Attack bundling can be used with different prioritization schemes to optimize quantities such as error rate on adversarial examples, perturbation size needed to cause misclassification, or failure rate when using a specific confidence threshold.

We present a new approach to bounding the true error rate of a continuous valued classifier based upon PAC-Bayes bounds. The method first constructs a distribution over classifiers by determining how sensitive each parameter in the model is to noise. The true error rate of the stochastic classifier found with the sensitivity analysis can then be tightly bounded using a PAC-Bayes bound.

We present a new approach to bounding the true error rate of a continuous valued classifier based upon PAC-Bayes bounds. The method first constructs adistribution over classifiers by determining how sensitive each parameter in the model is to noise. The true error rate of the stochastic classifier found with the sensitivity analysis can then be tightly bounded using a PAC-Bayes bound.