Accelerating Convolutional Neural Networks by Removing Interspatial and Interkernel Redundancies.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional Neural Networks (CNNs) are computationally intensive, limiting their widespread application.
• Existing methods for CNN acceleration primarily address either interspatial or interkernel redundancy, but not both.
• Redundant calculations in CNNs lead to high resource demands and reduced efficiency.

Purpose of the Study:

• To develop a unified method for accelerating CNNs by simultaneously removing interspatial and interkernel redundancies.
• To improve the accuracy and compression rate of CNN models.
• To address the computational resource limitations of current CNNs.

Main Methods:

• A novel block decomposition is proposed to convert interspatial redundancy into interkernel redundancy within convolutional layers.
• Rank-selection and pruning techniques are employed to eliminate the identified interkernel redundancy.
• A layer-wise training algorithm is utilized to manage convergence challenges, followed by full network fine-tuning.

Main Results:

• The proposed method successfully reduces computational demands by addressing both types of redundancy.
• Experiments on three widely used image classification datasets demonstrate superior performance compared to state-of-the-art methods.
• Significant improvements in both accuracy and compression rate were achieved.

Conclusions:

• The developed technique offers an effective strategy for accelerating CNNs by tackling multiple redundancy types.
• This approach enhances the practical applicability of CNNs by reducing their computational footprint.
• The method provides a competitive alternative for efficient deep learning model development.