[Paper Review] Practical Block-wise Neural Network Architecture Generation

Zhao Zhong, Junjie Yan+3 more|arXiv (Cornell University)|Aug 18, 2017

Advanced Neural Network ApplicationsComputer Science43 references20 citations

TL;DR

This paper proposes BlockQNN, a Q-learning-based framework for automatic, block-wise neural network architecture generation that achieves state-of-the-art performance on CIFAR-10 with a 3.54% top-1 error rate. By training an agent to sequentially select optimal block components and using early stopping and distributed training, BlockQNN reduces search space and computational cost—achieving convergence in just 3 days with 32 GPUs, while also demonstrating strong transferability to ImageNet.

ABSTRACT

Convolutional neural networks have gained a remarkable success in computer vision. However, most usable network architectures are hand-crafted and usually require expertise and elaborate design. In this paper, we provide a block-wise network generation pipeline called BlockQNN which automatically builds high-performance networks using the Q-Learning paradigm with epsilon-greedy exploration strategy. The optimal network block is constructed by the learning agent which is trained sequentially to choose component layers. We stack the block to construct the whole auto-generated network. To accelerate the generation process, we also propose a distributed asynchronous framework and an early stop strategy. The block-wise generation brings unique advantages: (1) it performs competitive results in comparison to the hand-crafted state-of-the-art networks on image classification, additionally, the best network generated by BlockQNN achieves 3.54% top-1 error rate on CIFAR-10 which beats all existing auto-generate networks. (2) in the meanwhile, it offers tremendous reduction of the search space in designing networks which only spends 3 days with 32 GPUs, and (3) moreover, it has strong generalizability that the network built on CIFAR also performs well on a larger-scale ImageNet dataset.

Motivation & Objective

• To address the high computational cost and complexity of manually designing high-performance convolutional neural networks.
• To reduce the search space in neural architecture search by focusing on block-wise rather than per-layer architecture generation.
• To develop an efficient, scalable, and generalizable method for automatic neural network architecture generation.
• To enable fast convergence and transferability of generated networks across datasets like CIFAR-100 and ImageNet.

Proposed method

• A Q-learning agent with experience replay and epsilon-greedy exploration is trained to sequentially select optimal component layers for a single network block.
• The optimal block is generated through reinforcement learning, with a custom reward function that correlates early-stopped performance with converged model accuracy.
• An early stop strategy accelerates training by identifying high-performing blocks early, reducing total training time.
• A distributed asynchronous framework enables efficient scaling across 32 GPUs, significantly reducing search time.
• The final network is constructed by stacking the generated optimal blocks, enabling modular and generalizable architecture design.
• The method supports adaptive block numbers and parameter constraints, improving efficiency and model compactness.

Experimental results

Research questions

• RQ1 Can a block-wise neural architecture search framework achieve state-of-the-art performance with significantly reduced computational cost?
• RQ2 How effective is Q-learning with early stopping in accelerating the discovery of high-performing network blocks?
• RQ3 To what extent can a network auto-generated for CIFAR-10 generalize to larger-scale datasets like ImageNet?
• RQ4 Can the proposed method outperform random search and existing NAS approaches in both accuracy and search efficiency?

Key findings

• BlockQNN achieves a 3.54% top-1 error rate on CIFAR-10, outperforming all existing auto-generated networks.
• The method converges in only 3 days using 32 GPUs, compared to 28 days with 800 GPUs in NAS, demonstrating a significant efficiency gain.
• The best model generated by BlockQNN surpasses random search by over 1% in top-5 accuracy on CIFAR-100 after equivalent training iterations.
• BlockQNN-L, with parameter and block number constraints, still outperforms its random search baseline by over 1% in the exploitation phase.
• The auto-generated network for person keypoint detection on MPII achieves 70.5 AP on COCO, surpassing the SOTA hourglass-2-stacks model (70.1 AP).
• The evolutionary process shows that BlockQNN learns to generate complex, multi-branch structures with skip connections and concatenation, indicating effective learning of general design principles.