Conceitos essenciais
Optimal SCOD strategy involves Bayes classifier for ID data and a linear selector in 2D space.
Resumo
This research paper addresses the Selective Classification in the presence of Out-of-Distribution (SCOD) problem. It introduces the optimal SCOD strategy involving a Bayes classifier for In-Distribution (ID) data and a linear selector in a 2D space. The study demonstrates that existing OOD detection methods and Softmax Information Retaining Combination (SIRC) provide suboptimal strategies compared to the proposed optimal solution. Additionally, it establishes the non-learnability of SCOD when relying solely on an ID data sample. The introduction of POSCOD, a method for learning the plugin estimate of the optimal SCOD strategy from both ID data and an unlabeled mixture of ID and OOD data, is shown to outperform existing methods.
Introduction
Addresses Selective Classification in SCOD.
Introduces optimal SCOD strategy with Bayes classifier and linear selector.
Existing methods like SIRC offer suboptimal solutions.
Non-learnability of SCOD with only ID data.
Introduction of POSCOD method for learning optimal SCOD strategy.
Data Extraction
"Linear: 5.84/13.88"
"SIRC: 6.53/15.52"
"θr(x): 17.43/45.3"
"θg(x): 14.11/15.11"
Estatísticas
Linear: 5.84/13.88
SIRC: 6.53/15.52
θr(x): 17.43/45.3
θg(x): 14.11/15.11
Citações
"The devil is in the wrongly-classified samples."
"Optimal prediction strategy comprises Bayes classifier for ID data."