FNA++: Fast Network Adaptation via Parameter Remapping and Architecture Search

Area of Science:

• Computer Vision
• Deep Learning
• Machine Learning

Background:

• State-of-the-art (SOTA) computer vision models often use image classification backbones pre-trained on ImageNet.
• Designing specialized architectures via neural architecture search (NAS) can improve performance in tasks like semantic segmentation and object detection.
• Pre-training large NAS search spaces or discovered networks incurs substantial computational expense.

Purpose of the Study:

• To introduce a computationally efficient method for adapting pre-trained neural networks for various computer vision tasks.
• To enable the effective application of NAS for segmentation and detection by reducing the associated computational burden.
• To demonstrate the versatility and performance improvements of the proposed adaptation technique.

Main Methods:

• Proposed Fast Network Adaptation (FNA++) method to adapt seed networks (e.g., ImageNet pre-trained) to new architectures (varying depth, width, kernel size).
• Employed a parameter remapping technique for efficient architecture and parameter adaptation.
• Applied FNA++ to MobileNetV2, ResNets, and NAS networks for semantic segmentation, object detection, and human pose estimation.

Main Results:

• FNA++ adapted networks significantly outperformed manually designed and NAS-derived networks on segmentation, detection, and pose estimation tasks.
• Demonstrated strong generalization ability when applied to diverse network architectures like ResNets and NAS networks.
• Achieved substantial reductions in computational cost compared to SOTA NAS methods (e.g., 1737x less than DPC, 6.8x less than Auto-DeepLab, 8.0x less than DetNAS).

Conclusions:

• FNA++ offers a highly efficient approach to leverage pre-trained models for specialized computer vision tasks.
• The method significantly lowers the computational barrier for applying NAS to demanding tasks like semantic segmentation and object detection.
• FNA++ shows promise for broader adoption in developing high-performance computer vision models with reduced resource requirements.