Expoliting Confidence-Based Model Fusion for Boosting Image Classification Accuracy

作者：Xinjian Jiang · 年份：2023 · DOI：10.1109/icicml60161.2023.10424922 · 研究领域：Industrial Vision Systems and Defect Detection、Medical Imaging Techniques and Applications、Anomaly Detection Techniques and Applications

In the realm of deep learning, the traditional approach has been to train specialized models for individual tasks, which, although effective, is resource-intensive. The advent of large, universal models has mitigated this issue by offering multitask capabilities, reduced training time, and lower computational costs. However, these generalized models often underperform on specific tasks compared to specialized models. This paper introduces an innovative ensemble approach that integrates specialized and generalized models, specifically focusing on Contrastive Language–Image Pretraining (CLIP) and EfficientNet. This work proposes three fusion strategies: Weighted Voting, Confidence Comparison, and Fully Connected Network Fusion, and evaluate them using the CIFAR-100 dataset. The ensemble model significantly outperforms individual models, achieving an adjusted accuracy of up to 0.848. The paper also introduces a novel evaluation metric, Confidence-Accuracy Correlation, to assess the reliability of model confidence. The findings could revolutionize ensemble learning by making it more adaptive and suited for real-world applications, thereby pushing the boundaries of possibility in artificial intelligence.