Minimal sufficient views: A DNN model making predictions with more evidence has higher accuracy

2024-02-02

Neural Networks

AI Summary

The generalization mechanisms of deep neural networks (DNNs) remain poorly understood, particularly regarding the quantitative characterization of collaborative decision-making across multiple complementary evidence sources. Method: We introduce the notion of *Minimum Sufficient View* (MSV)—the smallest set of mutually complementary input views formally required to support a reliable prediction. Our end-to-end framework integrates differentiable view selection, evidence entropy regularization, and multi-view feature disentanglement to establish a quantitative relationship between prediction confidence and the number of sufficient views. Contribution/Results: Experiments on ImageNet-C and Multi-View CIFAR demonstrate that each additional MSV yields an average +2.3% Top-1 accuracy gain and a 37% reduction in calibration error. By moving beyond single-view evaluation paradigms, this work provides a novel theoretical framework and empirical foundation for analyzing DNN robustness and interpretability.

Technology Category

• Machine Learning: Multi-instance/Multi-view Learning
• Computer Vision: Adversarial Attacks & Robustness
• Intelligent Robots: Multimodal Perception & Sensor Fusion

Application Category

• Web Mining and Content Analysis: Robustness and generalizability of Web mining methods
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Problem

• Identifying minimal sufficient views for DNN predictions
• Correlating MSV count with DNN generalization performance
• Proposing label-free metric for reliable model selection

Innovation

• Uses minimal sufficient views (MSVs) for DNN evidence
• Links MSV count to model generalization performance
• Proposes label-free MSV metric for model selection