Searching for Network Width With Bilaterally Coupled Network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Channel pruning is crucial for deploying convolutional neural networks (CNNs) on hardware with constraints.
• Current one-shot supernet methods for width search suffer from training unfairness due to unilateral augmentation (UA).

Purpose of the Study:

• To address training unfairness in CNN width search using supernets.
• To develop a novel supernet architecture for fair and accurate evaluation of network widths.

Main Methods:

• Introduction of Bilaterally Coupled Network (BCNet) where channels are trained fairly across network widths.
• Proposal of BCNetV2 to reduce search space and ensure rigorous training fairness.
• Leveraging a stochastic complementary strategy and prior initial population sampling for evolutionary search.
• Development of Channel-Bench-Macro, an open-source benchmark for macro structure width search.

Main Results:

• BCNet ensures each channel is responsible for an equal amount of network widths, leading to more accurate width evaluations.
• BCNetV2 enhances training fairness and reduces the redundant search space.
• The proposed methods, including evolutionary search enhancements, achieve state-of-the-art performance on benchmark datasets.

Conclusions:

• BCNet and its enhanced version BCNetV2 offer a fair and accurate approach to CNN width search.
• The developed benchmark facilitates better comparisons of width search algorithms.
• The findings contribute to more efficient deployment of CNNs under hardware constraints.