A Structurally Regularized CNN Architecture via Adaptive Subband Decomposition.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are powerful tools for image classification.
• Traditional CNNs process full-band signals, which can be computationally intensive.
• Existing architectures may lack robustness to noise and quantization.

Purpose of the Study:

• To propose a generalized CNN architecture that decomposes input signals into subbands.
• To improve computational efficiency and robustness of CNNs.
• To maintain or enhance classification accuracy compared to existing methods.

Main Methods:

• A novel CNN architecture employing an adaptive filter bank for subband decomposition.
• Independent feature extraction from each subband using convolutional layers.
• Feature combination via fully connected layers for classification.
• End-to-end learning of subband decomposition from data.

Main Results:

• The proposed architecture demonstrates structural regularization by limiting subband CNNs to specific signal spectra.
• Achieved state-of-the-art classification accuracy on MNIST, CIFAR-10/100, Caltech-101, and ImageNet-2012 datasets.
• Attained top-5 (86.91%) and top-1 (69.73%) accuracy on ImageNet-2012.
• Significant reduction in computational costs: over 90% in inference and ~75% in back-propagation (single-layer decomposition).
• Robustness to input and weight-and-bias quantization noise.

Conclusions:

• The proposed subband decomposition CNN architecture offers a computationally efficient and robust alternative to traditional full-band CNNs.
• It achieves superior or competitive accuracy across various benchmark datasets.
• The method provides substantial computational savings, particularly beneficial for large-scale deep learning applications.