Assessing Deep Neural Networks as Probability Estimators

By Yu P. et al

Read the original document by opening this link in a new tab.

Table of Contents

Abstract
I. Introduction
II. Related Work
III. Framework
IV. Experiment and Evaluation
A. One Dimensional Case
B. Two Paths
C. Comparison

Summary

This document focuses on the assessment of Deep Neural Networks (DNNs) as probability estimators, particularly in the context of classification uncertainties. It discusses the use of DNNs to estimate conditional probabilities and proposes a framework for uncertainty characterization. The document explores factors impacting the precision of DNNs' estimations, such as probability density and inter-categorical sparsity. The framework involves two paths: Bayesian inference and data sampling/training. The experiment and evaluation section includes a systematic analysis in a one-dimensional case, grid search in parametric space, and consideration of density and sparsity as influencing factors.