BlockQNN: Efficient Block-Wise Neural Network Architecture Generation.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Convolutional neural networks (CNNs) are highly successful in computer vision.
• Current popular CNN architectures are typically hand-crafted, requiring significant expertise and effort.
• Automating CNN architecture design is a key challenge in machine learning.

Purpose of the Study:

• To introduce BlockQNN, a novel block-wise network generation pipeline for automated CNN architecture design.
• To leverage the Q-Learning paradigm for efficient and effective network block construction.
• To reduce the complexity and computational cost associated with designing high-performance CNNs.

Main Methods:

• Utilized a Q-Learning agent with an epsilon-greedy strategy to sequentially select component layers for optimal network blocks.
• Stacked generated blocks to construct complete, auto-generated CNN architectures.
• Implemented a distributed asynchronous framework and an early stop strategy to accelerate the generation process.

Main Results:

• Achieved state-of-the-art performance on image classification tasks compared to hand-crafted networks.
• The best BlockQNN-generated network attained a 2.35% top-1 error rate on CIFAR-10.
• Demonstrated strong generalizability, with networks performing well on larger datasets like ImageNet (82.0% top-1, 96.0% top-5 accuracy).

Conclusions:

• BlockQNN offers a significant reduction in the search space for network design, completing generation in as little as 3 days with 32 GPUs.
• The block-wise generation approach provides a unique advantage in creating efficient and effective CNN architectures.
• Automated CNN design using BlockQNN yields competitive and generalizable results, outperforming traditional methods.