DS-UI

DS-UI enhances Bayesian estimation-based uncertainty inference for deep neural network (DNN)-based image recognition by integrating a dual-supervised mechanism with a mixture of Gaussian mixture models (MoGMM) to enable probabilistic interpretation of classifier features and improved detection of misclassification and out-of-distribution samples.

Key Features:

• MoGMM-FC Layer Integration: The MoGMM-FC layer combines the classifier's last fully-connected (FC) layer with a mixture of Gaussian mixture models (MoGMM) to provide direct probabilistic interpretation of input features.
• Dual-Supervised Stochastic Gradient-Based Variational Bayes (DS-SGVB): DS-SGVB is an optimization algorithm for the MoGMM-FC layer that models positive samples within classes and negative samples from other classes to reduce intra-class distances and increase inter-class margins.
• Enhanced Misclassification Detection: Experimental evaluations report improved misclassification detection compared with existing uncertainty inference methods through MoGMM-based probabilistic modeling and dual supervision.
• Open-Set Out-of-Domain/Distribution Detection: The framework improves identification of out-of-domain or distribution-shifted samples in open-set scenarios using MoGMM-driven uncertainty estimates.
• Visualizations of Feature Spaces: Visual analyses demonstrate more compact intra-class representations and clearer inter-class separation in feature space under DS-UI.

Scientific Applications:

• Medical Imaging: Provides reliable uncertainty estimates and misclassification detection for DNN-based medical image interpretation tasks.
• Autonomous Systems: Supports detection of out-of-distribution inputs and misclassifications in perception modules of autonomous systems.
• DNN-based Image Recognition: Improves uncertainty quantification, misclassification detection, and open-set robustness across image-recognition applications.

Methodology:

The methodology combines probabilistic modeling via mixture of Gaussian mixture models (MoGMM) in a MoGMM-FC layer with dual-supervised learning and a DS-SGVB optimization algorithm that explicitly models positive and negative samples to reduce intra-class distances and increase inter-class margins.