Optimal decision rule with class-selective rejection and performance constraints

Revisiting Confidence Estimation: Towards Reliable Failure Prediction

Reliable confidence estimation is a challenging yet fundamental requirement in many risk-sensitive applications. However, modern deep neural networks are often overconfident for their incorrect predictions, i.e., misclassified samples from known classes, and out-of-distribution (OOD) samples from unknown classes. In recent years, many confidence calibration and OOD detection methods have been developed. In this paper, we find a general, widely existing but actually-neglected phenomenon that most confidence estimation methods are harmful for detecting misclassification errors. We investigate this problem and reveal that popular calibration and OOD detection methods often lead to worse confidence separation between correctly classified and misclassified examples, making it difficult to decide whether to trust a prediction or not. Finally, we propose to enlarge the confidence gap by finding flat minima, which yields state-of-the-art failure prediction performance under various settings including balanced, long-tailed, and covariate-shift classification scenarios. Our study not only provides a strong baseline for reliable confidence estimation but also acts as a bridge between understanding calibration, OOD detection, and failure prediction.

What are the differences between Bayesian classifiers and mutual-information classifiers?

In this paper, both Bayesian and mutual-information classifiers are examined for binary classifications with or without a reject option. The general decision rules are derived for Bayesian classifiers with distinctions on error types and reject types. A formal analysis is conducted to reveal the parameter redundancy of cost terms when abstaining classifications are enforced. The redundancy implies an intrinsic problem of nonconsistency for interpreting cost terms. If no data are given to the cost terms, we demonstrate the weakness of Bayesian classifiers in class-imbalanced classifications. On the contrary, mutual-information classifiers are able to provide an objective solution from the given data, which shows a reasonable balance among error types and reject types. Numerical examples of using two types of classifiers are given for confirming the differences, including the extremely class-imbalanced cases. Finally, we briefly summarize the Bayesian and mutual-information classifiers in terms of their application advantages and disadvantages, respectively.

Gene-based multiclass cancer diagnosis with class-selective rejections

Supervised learning of microarray data is receiving much attention in recent years. Multiclass cancer diagnosis, based on selected gene profiles, are used as adjunct of clinical diagnosis. However, supervised diagnosis may hinder patient care, add expense or confound a result. To avoid this misleading, a multiclass cancer diagnosis with class-selective rejection is proposed. It rejects some patients from one, some, or all classes in order to ensure a higher reliability while reducing time and expense costs. Moreover, this classifier takes into account asymmetric penalties dependant on each class and on each wrong or partially correct decision. It is based on ν-1-SVM coupled with its regularization path and minimizes a general loss function defined in the class-selective rejection scheme. The state of art multiclass algorithms can be considered as a particular case of the proposed algorithm where the number of decisions is given by the classes and the loss function is defined by the Bayesian risk. Two experiments are carried out in the Bayesian and the class selective rejection frameworks. Five genes selected datasets are used to assess the performance of the proposed method. Results are discussed and accuracies are compared with those computed by the Naive Bayes, Nearest Neighbor, Linear Perceptron, Multilayer Perceptron, and Support Vector Machines classifiers.