ACLI: A CNN Pruning Framework Leveraging Adjacent Convolutional Layer Interdependence and $\gamma$γ-Weakly...

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Current convolutional neural network (CNN) pruning methods often rely on manual heuristics, limiting their generality and performance.
• Existing techniques may lack robustness and guaranteed performance due to their heuristic nature.

Purpose of the Study:

• To propose a novel theoretical framework for CNN pruning using gamma-weak submodularity.
• To develop a data-free, low-complexity algorithm for filter selection in convolutional layers.

Main Methods:

• Leveraging gamma-weak submodularity with a new importance function derived from error bounds.
• Formulating filter importance as a gamma-weakly submodular function.
• Developing a data-free oblivious algorithm for filter pruning.

Main Results:

• The proposed method outperforms state-of-the-art networks across datasets, achieving 76.52% accuracy.
• A 25.5% reduction in network parameters was achieved with competitive accuracy.
• The ACLI approach demonstrated orders-of-magnitude higher resource efficiency compared to baselines.

Conclusions:

• The gamma-weak submodularity framework provides an effective and efficient approach to CNN pruning.
• The developed algorithm offers a data-free, low-complexity solution with superior resource efficiency and accuracy.
• This method represents a significant advancement in optimizing CNNs for practical applications.