Conformal Prediction and Its Extensions

Please enable javascript in your browser.

Moradi, Mohammad Mahdi | 2024

0 Viewed

Type of Document: M.Sc. Thesis
Language: Farsi
Document No: 57669 (02)
University: Sharif University of Technology
Department: Mathematical Sciences
Advisor(s): Alishahi, Kasra
Abstract:
Nowadays, machine learning is used to solve various problems, and different machine learning models have increasingly become a part of our daily lives. This raises the importance of a more detailed examination of the models' outputs. In sensitive areas such as medical uses, this issue becomes even more critical, and using these models without having a framework to quantify the uncertainty of the outputs can lead to significant challenges and hinder their broader application. Conformal prediction is one of the frameworks that can help quantify the uncertainty of outputs. This framework is model-agnostic, meaning that we can use it for any model, including black-box models. The outputs of this method are prediction sets that we can have statistical inference for these sets. In this thesis, we introduce and examine conformal prediction and its extensions, analyze their advantages, disadvantages, and applications, run simulations on local data, and finally discuss the current issues and challenges in this field
Keywords:
Machine Learning ; Model Agnostic ; Conformal Prediction ; Uncertainty