Performance-Aware Approximation of Global Channel Pruning for Multitask CNNs.

Area of Science:

• Deep learning model compression
• Computer vision
• Artificial intelligence

Background:

• Global channel pruning (GCP) aims to reduce model size by removing filters without performance loss.
• Existing GCP methods struggle with multitask learning due to task mismatch and filter interdependencies.
• Task mismatch can lead to pruning filters crucial for secondary tasks during backbone optimization.

Purpose of the Study:

• To develop an effective framework for multitask model compression using global channel pruning.
• To address the challenges of task mismatch and filter interactions in multitask pruning.
• To propose a method that preserves task-relevant filters for improved multitask performance.

Main Methods:

• Introduced a Performance-Aware Global Channel Pruning (PAGCP) framework for multitask scenarios.
• Theoretically defined an objective function considering joint intra- and inter-layer filter saliency.
• Employed a sequentially greedy pruning strategy with a performance-aware oracle criterion to evaluate filter sensitivity across tasks.

Main Results:

• Achieved over 60% reduction in FLOPs and parameters on multitask datasets.
• Demonstrated minimal performance degradation despite significant model compression.
• Reported 1.2x to 3.3x acceleration on both cloud and mobile platforms.

Conclusions:

• PAGCP effectively compresses multitask deep learning models.
• The framework successfully balances performance preservation with significant size reduction.
• The proposed method offers practical benefits for deploying deep learning models on resource-constrained devices.