Core Concepts
Formulating a new GP-based MIL method using Hyperbolic Secant and Gamma distributions for improved predictive performance and efficiency.
Abstract
The content discusses the application of Hyperbolic Secant representation in probabilistic MIL methods, specifically VGPMIL, for CT intracranial hemorrhage detection. It introduces PG-VGPMIL and G-VGPMIL models, showcasing their equivalence to VGPMIL and their enhanced predictive performance. The study includes experiments on synthetic datasets, benchmarks, and real-world medical problems.
Directory:
Introduction to Multiple Instance Learning (MIL)
MIL as weakly supervised learning with reduced annotation effort.
Gaussian Processes for MIL
Formulation of MIL problem using GPs and logistic classification.
P´olya-Gamma Variables for GP-MIL
Introduction of P´olya-Gamma variables in GP classification.
General Inference Framework for Logistic Observation Model
Development of a general framework using differentiable GSM densities.
Gamma Variables for GP-MIL
Utilization of Gamma variables in GP-MIL models.
Results Analysis on MNIST, MUSK1 & 2 Datasets, RSNA, CQ500 Datasets
Evaluation of G-VGPMIL's performance in various datasets with comparisons to VGPMIL.
Stats
"VGPMIL introduces the following expression of the bag label likelihood given the labels of the instances..."
"The P´olya-Gamma distribution PG (b, c), with b > 0 and c ∈R..."
"This equality ensures that the following joint density is well defined..."