PaperlayerView exact revision on arXiv

Source-linked AI summary

BlockQNN: Efficient Block-wise Neural Network Architecture Generation

Zhao Zhong, Zichen Yang, Boyang Deng, Junjie Yan, Wei Wu, Jing Shao, Cheng-Lin Liu

arXiv:1808.05584v1cs.CVcs.LG

TL;DR

Automated CNN design faces large architecture search spaces and limited transfer across datasets. BlockQNN learns a network block with Q-learning and stacks it, reporting strong classification results, faster search, and transfer to ImageNet.

  • Problem

    Automated architecture design faces a large layer-and-hyperparameter search space, while networks designed for one dataset or task may transfer poorly to others.

    Source: §1 · p. 1

  • Method

    BlockQNN uses Q-learning to select block layers sequentially, then stacks the learned block into a network, with distributed asynchronous search and early stopping.

    Source: §1 · p. 1, §6 · p. 12

  • Results

    BlockQNN networks outperform modern hand-crafted and auto-generated networks on image classification; the best CIFAR-10 model reaches 2.35% error.

    Source: §1 · p. 2, §6 · p. 12

  • Takeaways & Limitations

    The faster version produces a comparable CIFAR result with 1 GPU in 20 hours, making the search more accessible to researchers with limited computing resources.

    Source: §1 · pp. 1–2

    The generated network remains dataset-specific and cannot transfer well to tasks with different input sizes; a universal transferable block-connection formulation remains open.

    Source: §3.4 · p. 5

Abstract

from arXiv · show

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 to choose component layers sequentially. 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 yields state-of-the-art results in comparison to the hand-crafted networks on image classification, particularly, the best network generated by BlockQNN achieves 2.35% top-1 error rate on CIFAR-10. (2) it offers tremendous reduction of the search space in designing networks, spending only 3 days with 32 GPUs. A faster version can yield a comparable result with only 1 GPU in 20 hours. (3) it has strong generalizability in that the network built on CIFAR also performs well on the larger-scale dataset. The best network achieves very competitive accuracy of 82.0% top-1 and 96.0% top-5 on ImageNet.

1 INTRODUCTION

BlockQNN addresses the cost and transfer challenges of automated network design by learning a block with Q-learning and stacking it into a network. The paper reports strong image-classification results, faster search, and transfer to ImageNet.

1 INTRODUCTION

  • The paper frames automated design as a large layer-and-hyperparameter search problem, while block-wise search limits structural exploration to one block before stacking.The stated challenges also include weak transfer of networks designed for one dataset or task; block-wise generation is motivated by generalization across datasets and tasks.Source: §1 · p. 1, Figure 1 · p. 2
  • BlockQNN uses Q-learning with experience replay and epsilon-greedy exploration to select component layers sequentially, then stacks the learned block into a network.An early-stop strategy and reward function aim to accelerate search while keeping early-stopped performance positively correlated with converged performance.Source: §1 · p. 1
  • This paper adds depthwise convolution, analyzes inter-block connection styles, and introduces faster BlockQNN using network-performance prediction.The faster version is reported to cost 20 hours on one GPU for CIFAR.Source: §1 · p. 2
  • BlockQNN outperforms hand-crafted and existing automatically generated networks on image classification, with 2.35% error on CIFAR-10.The paper identifies this as a leading CIFAR-10 result.Source: §1 · p. 2
  • CIFAR search converges in 3 days with 32 GPUs, versus NAS at 800 GPUs for 28 days; a faster version reaches comparable results with 1 GPU in 20 hours.The paper attributes faster search to block-wise generation and an early-stop strategy.Source: §1 · p. 2
  • The best CIFAR-generated network transfers to ImageNet with 82.0% top-1 and 96.0% top-5 accuracy.The paper describes the transfer as requiring little modification.Source: §1 · p. 2

2 RELATED WORK

Prior automated network design includes evolutionary search and reinforcement-learning methods, but the paper identifies cost and transfer limits in these approaches. It distinguishes block-structure search from efficiency methods and meta-learning focused on tuning existing algorithms.

2 RELATED WORK

  • Evolutionary and genetic algorithms were earlier architecture-search approaches, but the authors say they did not compete with hand-crafted networks.Source: §2 · p. 2
  • NAS and MetaQNN use reinforcement learning, but direct application to ImageNet is computationally expensive and their generated networks may not transfer across datasets or input sizes.Source: §2 · p. 2
  • BlockQNN searches block architectures with distributed asynchronous Q-learning and early stopping, drawing on Inception,, and ResNet, block designs.The passage notes that NASNet concurrently learns blocks, called cells, with an RNN-based controller.Source: §2 · p. 2
  • Regression-based performance prediction and SMASH [37] target search efficiency, but the cited methods are described as not matching state-of-the-art performance and remaining task- or dataset-specific.Source: §2 · p. 2
  • Table 1 describes a search space with seven common layer types, kernel-size options, and predecessor indices for representing block connections.Source: Table 1 · p. 3
  • The paper distinguishes its goal of learning network-block topology from hyperparameter optimization and meta-learning, which use meta-data to tune existing algorithms.Source: §2 · p. 3

3 METHODOLOGY

BlockQNN uses Q-learning to design CNN blocks and stacks them into networks, then adds early stopping, automated block connections, and architecture-based accuracy prediction to accelerate search.

3.1 Convolutional Neural Network Blocks

  • NSC encodes each layer as a five-value code, while pre-activation convolution cells and concatenated terminal layers support complex blocks that are stacked into networks.Pooling performs downsampling; when pooling halves feature-map size, block weights double, and ImageNet networks use more pooling layers than CIFAR networks because their input sizes differ.Source: §3.1 · pp. 3–4

3.2 Designing Network Blocks With Q-Learning

  • The agent sequentially selects finite NSCs as block layers, trains the stacked network, and uses validation accuracy as reward; shaped intermediate rewards accelerate convergence.The authors contrast shaped rewards with prior work that set intermediate rewards to zero.Source: §3.2 · p. 4, Figure 5 · p. 5

3.3 Early Stop Strategy

  • 3 days with 32 GPUs complete the search, compared with NAS’s 28 days and 800 GPUs for similar performance.Early-stop accuracy can rank blocks poorly against final accuracy, so the redefined reward incorporates FLOPs and graph density to better track final accuracy.Source: §3.3 · p. 5

3.4 Connection Style Between Blocks

  • Automated block connections can improve performance over sequential stacking, but the resulting network is dataset-specific and does not transfer well across tasks with different input sizes.The authors identify transferable universal block-connection formulation as an open problem.Source: §3.4 · p. 5

3.5 Predicting Network Performance Before Training

  • Faster BlockQNN predicts accuracy at a specified epoch from architecture alone, embedding layer codes, aggregating them with a topology-aware LSTM, then combining the structural feature and epoch in an MLP.Unlike methods that observe an initial training curve, it predicts the entire curve using only the architecture.Source: §3.5 · pp. 5–6, Figure 7 · p. 6
  • 1 GPU and 20 hours suffice for Faster BlockQNN to obtain a comparable result.The earlier search took 3 days with 32 GPUs and achieved performance similar to NAS, which used 800 GPUs over 28 days.Source: §3.3 · p. 5, §3.5 · p. 7

4 TRAINING DETAILS

BlockQNN training combines Q-learning exploration and replay with early screening, distributed network training, and separate full-training and transfer-evaluation stages.

Distributed Asynchronous Framework

  • A master, controller, and compute-node framework trains generated networks in parallel and returns validation accuracy as the agent's reward.The experiments used 32 NVIDIA 1080Ti GPUs and took about 3 days; the faster version used 1 GPU for 20 hours for a comparable result.Source: §4 · pp. 7–8

Epsilon-greedy Strategy

  • The agent shifts from random to greedy actions as ϵ falls from 1.0 to 0.1, while replay memory supplies 64 blocks per iteration for Q-value updates.The paper reports better results with longer exploration because it exposes the agent to more block structures.Source: §4 · p. 7

BlockQNN Generation

  • 178 iterations sampled 11,392 blocks, with 64 NSC sets per mini-batch and a maximum block layer index of 23.The Q-learning update uses a learning rate of 0.01, discount factor 1, and reward weights µ = 1 and ρ = 8.Source: §4 · p. 7

CIFAR-100 training

  • Generated networks receive 12 CIFAR-100 epochs during block search, then selected stacked-block networks train for 300 epochs with augmentation and cutout.The search stage uses no data augmentation; the full-training stage uses random crops, horizontal flips, and cutout.Source: §4 · p. 7

Transferable BlockQNN

  • The CIFAR-100-searched block is evaluated on CIFAR-10 and ImageNet, with ImageNet training on 224×224 images and evaluation at 224×224 or 320×320.The CIFAR-10 settings match those used for CIFAR-100; ImageNet training uses batch size 512, SGD, cosine learning-rate decay, label smoothing, an auxiliary classifier, and dropout.Source: §4 · p. 7

Faster BlockQNN

  • Faster BlockQNN trains a predictor on 2,000 CIFAR-100 block structures, then substitutes predictor-based estimates for training sampled structures during search.The predictor uses 40-dimensional layer and epoch embeddings, a 160-unit LSTM hidden state, and a three-layer MLP with 200 hidden units per layer.Source: §4 · pp. 7–8

5 EXPERIMENTAL RESULTS

The experiments analyze block and connection search on CIFAR-100, including the effects of the PCC search unit and cross-resolution block connections.

5 EXPERIMENTAL RESULTS

  • The experiments evaluate block-search and block-connection training, then compare the proposed models with other state-of-the-art approaches on CIFAR and ImageNet.They also examine evolutionary-search properties and architecture-generation efficiency.Source: §5 · p. 8

5.1 Block Searching Analysis

  • 65% final-stage early-stop accuracy on CIFAR-100, versus 56% during random exploration; both top block candidates appeared in the final Q-learning stage.The authors report that the top candidates were verified by training their respective networks.Source: §5.1 · p. 8
  • PCC search outperforms searching ReLU, convolution, and batch normalization separately, which is more likely to generate poor blocks and requires more search space and time.PCC combines the three components into one searchable unit.Source: §5.1 · p. 8

5.2 Block Connection Analysis

  • The generated connection style favors links between blocks at different resolutions rather than sequential connections, potentially combining features across resolutions.The paper notes that this connection pattern has also been used in segmentation, and reports it works for classification with an appropriate style.Source: §5.2 · pp. 8–9
  • 62% final-stage early-stop accuracy on CIFAR-100, versus 56% during random exploration; the top connection model appeared in the final stage.The connection search used 46 batches, and its training process was shorter than standard BlockQNN because its search space was smaller.Source: §5.2 · p. 8, Figure 12 · p. 9
  • The Block-QNN-Connection model is dataset-specific and handles only 32×32 images; transferable block-connection design remains open.The authors identify this as a limitation of the generated connection style.Source: §5.2 · p. 9

5.3 Faster BlockQNN with 1 GPUs in 20 Hours

Faster BlockQNN uses a predictor during block search to reduce computation, though its search curve is shakier than standard BlockQNN's.

5.3 Faster BlockQNN with 1 GPUs in 20 Hours

  • 60% final-stage mean accuracy on CIFAR-100, versus 56% during random exploration; the top block was found in the final search stage.The search curve was shaky, and its mean accuracy was lower than standard BlockQNN's; the authors suggest predictor and early-stop training errors may explain the difference.Source: §5.3 · p. 9
  • The predictor-based search produced a block unlike standard BlockQNN's top blocks: it lacked multi-branch structure and resembled a DenseNet/ResNet variant.The authors describe the generated block as less complex than standard BlockQNN blocks.Source: §5.3 · p. 9
  • The faster model required very little computing resource while achieving comparable performance, according to the authors.The paper attributes the approach's lower resource requirement to using the predictor during search.Source: §5.3 · p. 9

5.4 Results on CIFAR

On CIFAR, BlockQNN variants outperform hand-crafted and automatically generated networks, with results depending on block connections, parameter constraints, and convolution design.

Comparison with hand-crafted networks

  • With CIFAR’s 32 × 32 images, N = 4; Block-QNN-A and Block-QNN-B outperform most hand-crafted networks, while DenseNet-BC’s compression design was not adopted.The authors state that adopting DenseNet-BC’s prior knowledge could further improve their performance.Source: §5.4 · p. 9

Comparison with auto-generated networks

  • 32 GPUs for 3 days yielded state-of-the-art performance versus NAS’s 800 GPUs for 28 days; the faster version used 1 GPU for 20 hours.The faster version is reported as 100x less expensive than standard BlockQNN and achieved a comparable result.Source: §5.4 · p. 9

Transfer block from CIFAR-100 to CIFAR-10

  • 3.60% error rate on CIFAR-10 is reported after transferring top blocks learned from CIFAR-100, with experiment settings unchanged.The authors describe this as evidence of transferable ability.Source: §5.4 · p. 9

Analysis on network parameters

  • A 10M parameter cap with adaptive block numbers produced Block-QNN-S, which outperformed NASv3 with fewer parameters; more filters yielded 2.80%.The higher-filter configuration changes convolutional-layer filters from [32,64,128] to [80,160,320].Source: §5.4 · p. 9

Block connection

  • 2.35% error rate follows adding more filters to the block-connection design, which improves on sequential Block-QNN-S from 3.30% to 3.01% with fewer parameters.The passage reports 2.35% with more filters in each convolutional layer.Source: §5.4 · p. 9

Advanced depthwise convolution operation

  • 2.58% CIFAR-10 error rate is achieved by BlockQNN-Depthwise with 3.3M parameters, outperforming networks with similar parameter counts.Each convolution uses ReLU, depthwise convolution, pointwise convolution, and batch normalization, with batch normalization and ReLU removed between the depthwise and pointwise operations.Source: §5.4 · p. 9

5.5 Transfer to ImageNet

The authors transfer CIFAR-learned blocks to ImageNet without fine-tuning and report competitive results, including strong single-model accuracy under stated image-size conditions.

Transfer setup

  • CIFAR-100-learned blocks were used directly without fine-tuning; the ImageNet network used N = 3, additional downsampling before blocks, and filters [64,128,256,512].The blocks were transferred from CIFAR to ImageNet.Source: §5.5 · pp. 9–10

ImageNet comparison

  • Block-QNN-Depthwise outperformed the cited hand-crafted Xception and ResNext networks on ImageNet.The passage attributes those networks’ parameter reduction and performance improvements to depthwise convolution.Source: §5.5 · p. 10

Transferability and scope

  • The authors frame block learning as a way to transfer architectures from smaller datasets to large-scale ImageNet, where most earlier automatic-generation work had not reported competitive results.They identify SENet’s channel-response recalibration as a design feature outside their search space and future work consideration.Source: §5.5 · p. 10

ImageNet results

  • 82.0% top-1 and 96.0% top-5 accuracy are reported for Block-QNN-Depthwise, trained on 224×224 and tested on 320×320 images under single-crop, single-model conditions.For comparison, Block-QNN-Depthwise and NASNet-A(6 @ 4032) were trained on 224×224 images with the same parameter setting.Source: §5.5 · p. 10

5.6 Evolutionary Process of Auto-Generated Blocks

Across sampled iterations, generated blocks shift from simpler exploration-stage structures toward more complex exploitation-stage designs, with recurring connection patterns.

Evolution across iterations

  • At iterations 1, 30, 60, 90, 110, 130, 150, and 170, median-performance block structures sampled to illustrate evolution were much simpler in random exploration than in exploitation.Source: §5.6 · p. 11

Exploitation-stage structures

  • Multi-branch structures appear frequently in exploitation, with increasing connections, concatenation often last, and shortcut connections and elementwise addition common.These patterns are reported as structures generated in the exploitation stage.Source: §5.6 · pp. 11–12

Limited-parameter variant

  • BlockQNN-L generates fewer 5 × 5 convolution layers, consistent with its parameter limitation.The passage identifies these as “Conv,5” layers.Source: §5.6 · p. 12

Interpretation

  • The authors interpret the evolving structures as evidence that BlockQNN learns reusable design concepts and produces more explicable networks than other automatic architecture-design methods.Source: §5.6 · p. 12

5.7 Efficiency of BlockQNN

Against random search under equivalent sampling, BlockQNN produces stronger exploitation-stage models, including when both methods use parameter limits and adaptive block counts.

Evaluation criterion

  • Effectiveness is measured as the increase in top auto-generated network performance from initial random exploration to exploitation, with comparisons using equivalent training iterations.Source: §5.7 · p. 12

BlockQNN versus random search

  • Over 1% better best-model performance was observed for BlockQNN than random search during exploitation after 11,392 sampled blocks on CIFAR-100.The reported mean performance of the top five models also favored BlockQNN; both methods began with the same exploration phase.Source: §5.7 · p. 12

Limited-parameter comparison

  • Over 1% better exploitation-phase performance was reported for BlockQNN-L than for random search with limited parameters and adaptive block numbers.The authors conclude that BlockQNN can learn to generate better architectures than random search.Source: §5.7 · p. 12

5.8 Additional Experiment

BlockQNN is evaluated for key-point estimation and against random search for architecture-search efficiency. Its model trained on MPII transfers to COCO and exceeds the two-stack hourglass result there.

5.8 Additional Experiment

  • Figure 18 compares BlockQNN-L and RS-L by mean accuracy of top-ranked models against training iterations, using early-stop performance after 12 epochs on CIFAR-100.Source: Figure 18 · p. 12
  • 70.5 AP versus 70.1 AP for the two-stack hourglass on the COCO validation dataset; the BlockQNN model was searched on MPII and transferred to COCO in five days.Source: §5.8 · p. 12

6 CONCLUSION AND FUTURE WORK

The conclusion reports strong image-classification performance, transfer to ImageNet, and a faster, less costly BlockQNN version. It points to broader task support and unified automation as future work.

6 CONCLUSION AND FUTURE WORK

  • BlockQNN networks outperform hand-crafted and other auto-generated networks on image classification; the best CIFAR block transfers to ImageNet with competitive performance against hand-crafted networks.Source: §6 · p. 12
  • The faster BlockQNN version is 100x less expensive than standard BlockQNN while giving a comparable result.Source: §6 · p. 12
  • Empirical results indicate that sequential connection between blocks is not optimal, and the generated architectures are described as more elegant and explicable.Source: §6 · p. 12
  • Future work includes generating backbones directly for detection, segmentation, and tracking, and unifying input preprocessing, network design, and training-hyperparameter setting.Source: §6 · p. 12
Loading 1808.05584v1…