BNAS: Efficient Neural Architecture Search Using Broad Scalable Architecture.

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Efficient Neural Architecture Search (ENAS) uses parameter sharing and reinforcement learning (RL) for efficient architecture learning.
• ENAS's deep scalable architectures incur significant training costs, limiting search efficiency.
• Reducing layers in ENAS accelerates search but degrades performance.

Purpose of the Study:

• To propose Broad Neural Architecture Search (BNAS) using a novel Broad Convolutional Neural Network (BCNN) architecture.
• To improve search efficiency and maintain high performance in neural architecture search.
• To address the performance drop associated with layer reduction in ENAS.

Main Methods:

• Designed a shallow BCNN architecture for faster training.
• Employed RL and parameter sharing, similar to ENAS, for optimization.
• Developed two BCNN topology variants for BNAS.

Main Results:

• BNAS achieved a search cost of 0.19 days, 2.37x faster than ENAS.
• Learned architectures achieved state-of-the-art test errors of 3.58% and 3.24% on CIFAR-10 for small and medium models.
• Achieved 25.3% top-1 error on ImageNet with only 3.9 million parameters.

Conclusions:

• BNAS significantly enhances search efficiency compared to ENAS.
• The BCNN architecture enables high performance with reduced computational cost.
• BNAS offers a promising approach for efficient and effective neural architecture search.