Union-net: lightweight deep neural network model suitable for small data sets

Jingyi Zhou, Qingfang He, Guang Cheng, Zhiying Lin

TL;DR

This paper introduces Union-net, a novel lightweight deep neural network designed specifically for small datasets. By employing a unique "union convolution" module that combines features from different receptive fields, Union-net achieves competitive performance on small datasets while significantly reducing model complexity and parameters compared to traditional deep learning models.

What Problem Does This Paper Address?

Deep learning models typically require vast amounts of data to achieve high performance, often leading to overfitting when trained on smaller datasets. While transfer learning (using pre-trained models) is a common solution, it faces challenges like copyright issues and the need to select appropriate pre-trained models for specific tasks. The authors aim to address the limitations of traditional deep learning models in small data scenarios by proposing a new architecture that is inherently suited for such conditions, avoiding the complexities and potential pitfalls of transfer learning.

Methods

The core innovation of Union-net is its "union convolution" module. This module consists of multiple parallel convolutional neural network units, each with a different receptive field (achieved through varying depths of stacked 3x3 convolutions). The outputs from these parallel units are merged and added, creating a richer feature representation. The Union-net architecture itself is a shallow network composed of three such union modules in series, followed by a final convolutional layer, global average pooling, and a softmax classifier. This design aims to capture diverse features without the excessive depth and long information transmission paths found in traditional deep networks, thus mitigating overfitting on small datasets. The model also incorporates residual connections (inspired by ResNet) and feature reuse (inspired by DenseNet).

Datasets

The paper evaluates Union-net on several datasets:

  • CIFAR-10: The standard CIFAR-10 dataset (60,000 images in 10 categories) was used in its entirety and also subsampled to create smaller datasets: CIFAR10_30, CIFAR10_60, and CIFAR10_100, containing 30, 60, and 100 samples per category, respectively.
  • 17 Flower Dataset: A dataset of 1360 images across 17 flower categories, known for its fine-grained classification challenges and limited samples per category (around 60 for training after splitting).
  • MNIST, Fashion-MNIST, MiniImageNet, and CUB200_2011: These datasets were used for further verification of Union-net's applicability across different domains, including image classification and fine-grained tasks.

Key Insights and Ideas

  • "Union Convolution" for Diverse Receptive Fields: The key idea is to combine features extracted from different receptive field sizes within a single module. This allows the model to capture both fine-grained details and broader contextual information simultaneously, which is crucial for small datasets where each sample is more valuable.
  • Shallow Network for Small Data: The paper argues that a shallow network structure, as implemented by Union-net, is more suitable for small datasets than deep networks. This reduces the risk of overfitting and the loss of information during transmission.
  • Parameter Efficiency: By using smaller convolution kernels (3x3) and a more compact architecture, Union-net significantly reduces the number of parameters and computational complexity (FLOPs) compared to many classic deep learning models.
  • Avoiding Dropout for Small Datasets: The authors deliberately omit dropout, a common regularization technique, arguing that it can hinder the model's ability to learn from limited data by randomly discarding acquired information.

Key Results

  • Superior Performance on Small Datasets: Union-net consistently outperformed several classic CNN models (CNN, DenseNet, Vgg16, Resnet50, MobileNetV2, Efficientnet, BReG-NeXt, ASMNet) on the CIFAR10_30, CIFAR10_60, CIFAR10_100, and 17 Flower datasets. For instance, on CIFAR10_30, Union-net achieved 89.88% accuracy, significantly higher than many other models.
  • Competitive Performance on Large Datasets: While the full CIFAR-10 dataset yielded slightly better results for Union-net than its smaller counterparts, its performance was still competitive with other models, demonstrating its generalizability.
  • Parameter and FLOPs Reduction: Table 1 shows that Union-net has significantly fewer parameters (1.72M) and FLOPs (0.003B) compared to models like VGG16 (138M, 15.5B) and ResNet50 (26M, 3.9B), making it a lightweight alternative.
  • Ablation Studies: Experiments on the 17 Flower dataset confirmed that combining multiple convolutional units within the union module generally improved performance, with the full Union-net architecture (Uabcd) achieving the best results.

Limitations and Future Work

  • Fine-Grained Classification Improvement: While Union-net performs well, the paper notes that further improvement is needed for fine-grained classification tasks, as indicated by its performance on the CUB200_2011 dataset.
  • More-Class Classification: The authors suggest that extending the research to handle datasets with a much larger number of classes (e.g., 100-class or 1000-class classification) would be a valuable future direction.
  • Model Complexity vs. Performance: The ablation studies showed that adding more complex convolutional units (like a five-layer stacked convolution) to the union module did not significantly improve performance over the designed Union-net structure, while increasing computational cost. This suggests a trade-off between complexity and marginal gains.

Summary updated

BibTeX
@article{zhou2023union,
  author = {Jingyi Zhou and Qingfang He and Guang Cheng and Zhiying Lin},
  title = {{Union-net: lightweight deep neural network model suitable for small data sets}},
  journal = {Journal of Supercomputing},
  year = {2023},
  doi = {10.1007/s11227-022-04963-w},
}
Figure