CondenseNet with exclusive lasso regularization.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Group convolution is a common technique for improving computational efficiency in deep learning.
• Neural networks can suffer from overfitting, where they perform well on training data but poorly on unseen data.
• CondenseNet is a type of deep learning architecture that utilizes group convolutions.

Purpose of the Study:

• To develop a novel method, CondenseNet-elasso, to address feature correlation and overfitting in deep learning models.
• To improve the efficiency and generalization of CondenseNet architectures.

Main Methods:

• The study introduces CondenseNet-elasso, which applies exclusive lasso regularization to CondenseNet.
• Exclusive lasso regularization encourages distinct convolution groups to utilize different subsets of input channels.
• This encourages the learning of more diverse features.

Main Results:

• CondenseNet-elasso demonstrated superior efficiency compared to standard CondenseNets.
• The proposed method also showed better performance than other DenseNet variants.
• Experiments were conducted on CIFAR10, CIFAR100, and Tiny ImageNet datasets.

Conclusions:

• CondenseNet-elasso effectively reduces feature correlation and mitigates overfitting in deep neural networks.
• The developed technique offers improved computational efficiency and feature diversity.
• CondenseNet-elasso represents a promising advancement for deep learning model optimization.