Ing. Jakub Paplhám

This paper addresses the problem of designing reliable prediction models that abstain from predictions when faced with uncertain or out-of-distribution samples - a recently proposed problem known as Selective Classification in the presence of Out-of-Distribution data (SCOD). We make three key contributions to SCOD. Firstly, we demonstrate that the optimal SCOD strategy involves a Bayes classifier for in-distribution (ID) data and a selector represented as a stochastic linear classifier in a 2D space, using i) the conditional risk of the ID classifier, and ii) the likelihood ratio of ID and out-of-distribution (OOD) data as input. This contrasts with suboptimal strategies from current OOD detection methods and the Softmax Information Retaining Combination (SIRC), specifically developed for SCOD. Secondly, we establish that in a distribution-free setting, the SCOD problem is not Probably Approximately Correct learnable when relying solely on an ID data sample. Third, we introduce POSCOD, a simple method for learning a plugin estimate of the optimal SCOD strategy from both an ID data sample and an unlabeled mixture of ID and OOD data. Our empirical results confirm the theoretical findings and demonstrate that our proposed method, POSCOD, outperforms existing OOD methods in effectively addressing the SCOD problem.

Early detection of yeast and filamentous fungi in clinical samples is critical in treating patients predisposed to severe infections caused by these organisms. The patients undergo regular screening, and the gathered samples are manually examined by trained personnel. This work uses deep neural networks to detect filamentous fungi and yeast in the clinical samples to simplify the work of the human operator by filtering out samples that are clearly negative and presenting the operator with only samples suspected of containing the contaminant. We propose data augmentation with Poisson inpainting and compare the model performance against expert and beginner-level humans. The method achieves human-level performance, theoretically reducing the amount of manual labor by 87%, given a true positive rate of 99% and incidence rate of 10%.