Active Data Fusion in Deep Neural Networks via Separation Index
Movahed Jamshidi, Ahmad Kalhor, Abdol‐Hossein Vahabie · 2024
In this paper, we aim to enhance Convolutional Neural Networks (CNNs) by merging features from different pretrained models to improve information representation. Amidst advancements in network depth and architectures, we highlight feature fusion as a method for boosting CNN performance. Despite challenges in managing large feature sets and potential information loss during fusion, our paper introduces the Separation Index (SI) for feature selection. Unlike passive fusion, where all feature maps are simply concatenated, we employ the SI metric for active fusion. This method selects informative and discriminative features for the classification task, reducing computational complexity and model sizes. We evaluate our approach using medical and public datasets, specifically Skin Cancer data and CIFAR-10. The results demonstrate significant improvements in classification performance, confirming the effectiveness of our method. Notably, achieving a 95.63% classification accuracy on the CIFAR-10 dataset showcases the superiority of our approach over baselines and other competing methods.