Modulated Convolutional Networks.

Area of Science:

• Computer Vision
• Machine Learning
• Deep Learning

Background:

• Deep convolutional neural networks (DCNNs) excel in vision tasks but are computationally intensive.
• Model compression, particularly binarization, is crucial for resource-constrained devices.
• Existing binarization methods often struggle to maintain high performance.

Purpose of the Study:

• To propose Modulated Convolutional Networks (MCNs) for high-performance binarized DCNNs.
• To develop a generic framework applicable to various DCNN architectures.
• To significantly reduce model storage overhead without performance degradation.

Main Methods:

• Reformulated MCN calculation as a discrete optimization problem for binarized DCNNs.
• Introduced a unified framework incorporating filter loss, center loss, and softmax loss.
• Developed modulation filters (M-Filters) to recover filters from binarized ones.

Main Results:

• Achieved comparable performance to full-precision models with binarized DCNNs.
• Reduced storage cost of convolutional filters by a factor of 32.
• Outperformed other state-of-the-art binarized models.

Conclusions:

• MCNs offer an effective solution for compressing DCNNs via binarization.
• The proposed method enables efficient deployment of DCNNs on resource-limited hardware.
• MCNs provide a promising direction for high-performance, low-storage deep learning models.