Building Correlations Between Filters in Convolutional Neural Networks.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Conventional Convolutional Neural Networks (CNNs) train filters independently, neglecting inter-filter relationships.
• This independent training limits the learning capacity and generalization of CNN models.
• Exploring filter correlations offers a potential avenue for improving CNN performance.

Purpose of the Study:

• To introduce a novel optimization approach for CNNs using explicitly related filters.
• To develop a method for joint initiation and training of filters based on predefined correlations.
• To enhance the generalization capabilities of CNNs through cooperative filter learning.

Main Methods:

• Designed a new optimization approach for CNNs incorporating correlative filters (CFs).
• Initiated and trained CFs jointly, leveraging predefined correlations for cooperative learning.
• Validated the CF approach on five benchmark image classification datasets: CIFAR-10, CIFAR-100, MNIST, STL-10, and street view house number.

Main Results:

• The proposed CF approach demonstrated improved CNN performance across all tested benchmark datasets.
• Comparative experiments showed the CF method outperforming several state-of-the-art CNN approaches.
• Joint training of correlated filters led to more generalized and efficient optical systems.

Conclusions:

• The correlative filters (CFs) approach offers a significant advancement in CNN optimization.
• Jointly trained filters improve cooperative learning, leading to enhanced model generalization.
• This method represents a promising direction for developing more effective CNN architectures.