Practical Block-wise Neural Network Architecture Generation

1. Introduction

• CNN architecture design remains challenging because layer configurations create a vast search space and successful hand-crafted designs rely heavily on expertise.Automated designs also face high search costs and limited transfer across datasets or tasks.Source: §1 · p. 1
• Block-wise architectures use flexible stacks of personalized blocks, following the repeated-block pattern found in architectures such as Inception and ResNet.The paper presents block-wise generation as a way to improve search efficiency and generalization across datasets or tasks.Source: §1 · p. 1
• 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 are proposed to accelerate search while correlating early-stopped accuracy with converged accuracy.Source: §1 · pp. 1–2
• 3.54% error rate on CIFAR-10 is reported as state-of-the-art among auto-generated networks, with BlockQNN also comparable to hand-crafted networks.The CIFAR network is reported to transfer to ImageNet with little modification and achieve competitive performance.Source: §1 · p. 2
• The CIFAR search converges in 3 days with 32 GPUs, compared with NAS taking 28 days with 800 GPUs for the same performance.Block-wise search reduces the structures searched to one block, which is then stacked to build the network.Source: §1 · p. 2, Figure 1 · p. 2

2. Related Work

• NAS and MetaQNN use reinforcement learning to search architectures, but direct application to ImageNet is computationally expensive and their generated networks are described as dataset-specific.Earlier evolutionary approaches are reported as not competing with hand-crafted networks.Source: §2 · p. 2
• The block design follows CNN precedents: Inception uses multi-level convolution branches, while ResNet uses shortcut connections; generated blocks can exhibit similar structures.Figure 2 presents representative multi-branch and shortcut blocks with their Network Structure Codes.Source: §2 · pp. 2–3, Figure 2 · p. 3
• Network Structure Codes represent blocks as layer-operation vectors with kernel and predecessor information, using a finite set of choices to define the search space.The code space includes seven commonly used layer types, with specified kernel-size and predecessor-index options.Source: Table 1 · p. 3
• BlockQNN targets network-block topology, unlike hyperparameter optimization and meta-learning methods that improve existing algorithms or choose settings such as learning rates and layer counts.The related-work discussion cites hyper-parameter optimization, meta-learning, and learning-to-learn methods.Source: §2 · p. 3

3.1. Convolutional Neural Network Blocks

• Generated networks begin with convolution layers, then stack repeated blocks; pooling layers downsample, with more pooling for ImageNet’s larger inputs.Blocks can be repeated any N times or placed in other arrangements, including within Network-in-Network or with shortcuts between blocks.Source: Figure 3 · p. 3, §3.1 · p. 4
• Each block is represented as a DAG encoded by five-dimensional Network Structure Code vectors, including layer type, kernel size, and predecessor positions.Layers without successors are concatenated to form the block output.Source: §3.1 · p. 3
• A Pre-activation Convolutional Cell bundles ReLU, convolution, and batch normalization, reducing the search space compared with searching those components separately.The paper says this provides better initialization and a quicker training process for finding block structures.Source: §3.1 · p. 4

3.2. Designing Network Blocks With Q-Learning

• Q-learning selects NSC layer choices sequentially as a trajectory, with validation accuracy used as the reward to update Q-values.The state and action spaces are finite and discrete; the method assumes that a well-performing layer in one block should also perform well in another.Source: §3.2 · pp. 4–5
• Shaped intermediate rewards make the learning process converge faster than ignoring those rewards.The comparison uses the same exploration; the paper notes that ignoring intermediate rewards may slow early convergence and favor stopping at small blocks.Source: §3.2 · p. 5, Figure 5 · p. 5

3.3. Early Stop Strategy

• The early-stop reward uses early-stop accuracy alongside FLOPs and block density, making the reward more relevant to final accuracy.FLOPs estimate block computational complexity, while density is the number of DAG edges divided by its number of nodes.Source: §3.3 · p. 5, Figure 6 · p. 5
• 3 days on 32 GPUs, versus ’s 28 days on 800 GPUs for the same performance, completed block search using early stopping and a reduced search space.Source: §3.3 · p. 5

4.1. Distributed Asynchronous Framework

• Master samples block structures, the controller builds and dispatches networks, and compute nodes train them in parallel and return validation rewards.The framework uses 32 NVIDIA TitanX GPUs and took about 3 days to complete searching.Source: §4.1 · pp. 5–6, Figure 7 · p. 6, §4.2 · p. 6

4.2. Training Details

• The agent decreases ϵ from 1.0 to 0.1 to shift from random exploration toward greedy actions, while replay memory supplies 64 blocks and accuracies per iteration.The authors report better results with longer exploration because it lets the agent see more block structures.Source: §4.2 · p. 6
• Search used 178 iterations and 11,392 blocks, with 64 NSC sets per mini-batch and candidate networks trained for 12 epochs on CIFAR-100.The Q-learning rate was 0.01, the discount factor was 1, and reward-function hyperparameters µ and ρ were 1 and 8.Source: §4.2 · p. 6
• During search, CIFAR-100 candidates train without augmentation using Adam and batch size 256, with up to three restarts if first-epoch accuracy is below random guess.The learning rate starts at 0.001 and is reduced by a factor of 0.2 every five epochs.Source: §4.2 · p. 6
• After selecting a block, the authors train the stacked network for 300 epochs with augmentation, then evaluate transferred blocks using a 90-epoch SGD schedule on CIFAR-10 and ImageNet.Final CIFAR models use batch size 128; transfer evaluation uses center-crop test images.Source: §4.2 · p. 6

5.1. Block Searching Analysis

• 65% final-stage CIFAR-100 early-stop accuracy at ϵ = 0.1, versus 56% during random exploration; both top blocks appeared in the final Q-learning stage.Source: §5.1 · pp. 6–7, Figure 8 · p. 7

5.2. Results on CIFAR

• Block-QNN outperforms most hand-crafted networks and beats MetaQNN and NASv3 on CIFAR; Table 3 compares error rates on CIFAR-10 and CIFAR-100.The authors report that NAS used 800 GPUs for 28 days, whereas BlockQNN used 32 GPUs for 3 days.Source: Table 3 · p. 7, §5.2 · p. 8

5.1. Block Searching Analysis

• PCC’s combined ReLU, convolution, and batch-normalization search outperforms separate component search from the beginning and reduces search space and time for finding good blocks.Source: §5.1 · p. 7, Figure 9 · p. 7

5.2. Results on CIFAR

• 3.60% error rate on CIFAR-10 using top blocks transferred from CIFAR-100.Source: §5.2 · p. 8
• 3.54% result with more convolutional filters; under a 10M-parameter cap, Block-QNN-S outperforms NASv3 with fewer parameters.Figure 8(d) shows the limited-parameter block.Source: Figure 8 · p. 7, §5.2 · p. 8

5.3. Transfer to ImageNet

• Block structures learned on CIFAR-100 transfer directly to ImageNet without fine-tuning and yield competitive results against human-designed models; Table 5 reports single-crop error rates.The authors caution that comparison with Xception and ResNeXt is not fair because Block-QNN does not use their depth-wise convolution operation.Source: Table 5 · p. 8, §5.3 · p. 8

• BlockQNN networks outperform modern hand-crafted and other auto-generated networks in image-classification tasks.Source: §6 · p. 8
• The best CIFAR block transfers to ImageNet and yields competitive performance compared with the best hand-crafted networks.Source: §6 · p. 8
• The authors plan to search for blocks with limited FLOPs and experiment on detection or segmentation.Source: §6 · p. 9

• Effectiveness is defined as the increase in top auto-generated-network performance from initial random exploration to exploitation, not average performance.Source: A. Efficiency of BlockQNN · p. 9
• Over 1% in exploitation, BlockQNN’s best model outperforms random search after 11,392 sampled blocks on CIFAR-100.The mean performance of BlockQNN’s top five models also exceeds random search; both methods share the same exploration phase.Source: A. Efficiency of BlockQNN · p. 9
• Over 1% in exploitation, BlockQNN-L outperforms RS-L when both use limited parameters and adaptive block numbers.Source: A. Efficiency of BlockQNN · p. 9

• Blocks sampled at iterations 1, 30, 60, 90, 110, 130, 150, and 170 progress from simpler exploration-stage structures to more elaborate exploitation-stage structures.The sampled blocks are median-performing examples; Figures 12 and 13 show BlockQNN and BlockQNN-L, respectively.Source: B. Evolutionary Process of… · p. 9, Figure 12 · p. 10, Figure 13 · p. 10
• Exploitation-stage blocks commonly have multi-branch structures, increasing connections, shortcut connections, elemental add layers, and Concat as the last layer.BlockQNN-L has fewer Conv,5 layers because of its parameter limitation.Source: B. Evolutionary Process of… · p. 9
• The authors interpret the generated blocks as learning universal design concepts and describe the resulting networks as more elegant and model-explicable than other automatic architecture-design methods.Source: B. Evolutionary Process of… · p. 9

• 70.5 AP versus 70.1 AP for the state-of-the-art hourglass 2 stacks network on the COCO validation dataset.Source: C. Additional Experiment · p. 10
• The search was conducted on MPII, and the best model found there was transferred to the COCO challenge.Source: C. Additional Experiment · p. 10
• The person key-points search took 5 days to complete.Source: C. Additional Experiment · p. 10