Coupled convolution layer for convolutional neural network.

Area of Science:

• Computer Vision
• Deep Learning
• Computational Neuroscience

Background:

• Convolutional Neural Networks (CNNs) are fundamental to modern computer vision.
• Existing CNN architectures often have a large number of parameters, leading to computational and memory inefficiencies.
• Biological vision systems exhibit efficient feature extraction mechanisms that can inspire artificial systems.

Purpose of the Study:

• To propose a novel coupled convolution layer for CNNs.
• To enhance feature representation and reduce model complexity.
• To investigate the effectiveness of biologically inspired constraints on convolution filters.

Main Methods:

• A coupled convolution layer is designed with multiple parallel convolutions.
• Filters within the layer are mutually constrained, with weights being geometric transformations (rotation, mirroring, negation) of each other.
• The layer's performance is analyzed theoretically and empirically.

Main Results:

• The coupled convolution layer is found to be particularly effective in lower network layers where feature maps retain geometric properties.
• Experimental results on CIFAR-10, CIFAR-100, and PlanktonSet 1.0 datasets show a slight performance improvement over standard convolution layers.
• The proposed layer achieves this improvement with a reduced number of parameters compared to unconstrained layers.

Conclusions:

• The coupled convolution layer offers a promising approach to more efficient and effective deep learning models.
• Biologically inspired constraints on filters can lead to improved performance and parameter efficiency in CNNs.
• Further research can explore broader applications and variations of these constrained convolution layers.