To utilize network pruning to identify optimal Unet architectures and examine the effects of gradual channel pruning.
Approach:
Network Pruning: Investigated gradual channel pruning in Unet, focusing on channel selection during pruning and comparing it with random and systematic pruning baselines.
Lean Unet Architecture: Proposed a Lean Unet (LUnet) architecture with a fixed number of channels per block, aiming for efficiency and performance.
Key Findings:
Pruned Unet architectures can perform comparably or better than dense counterparts with significantly fewer parameters.
The choice of channels during pruning optimizes architecture rather than relying solely on weights/activations.
LUnet architecture has over 30 times fewer parameters than traditional Unet while maintaining performance.
Interpretation:
Limitations:
The study primarily focuses on Unet and may not generalize to other neural network architectures.
Gradual pruning can be resource-intensive due to the need for retraining after each pruning step.
