Introduction
Deep learning has traditionally been a "big data" endeavor. State-of-the-art models like ResNet and EfficientNet typically require massive datasets such as ImageNet to achieve high accuracy. However, in many real-world applications—ranging from medical imaging to niche industrial inspections—collecting millions of labeled images is often impossible or prohibitively expensive. While transfer learning is the standard workaround, it faces challenges related to copyright and domain gaps between source and target data.
This paper introduces Union-net, a lightweight convolutional neural network (CNN) architecture specifically designed to learn effectively from scratch using small datasets. By employing a novel architectural shift, Union-net avoids the overfitting common in standard deep models when data is scarce.
Methodology: The Union Module
The core innovation of this architecture is the Union module. Instead of a single sequential stack of convolutions, each module consists of four parallel paths that capture features at different spatial scales:
- Path A: One 3x3 convolution.
- Path B: Two stacked 3x3 convolutions (equivalent to a 5x5 receptive field).
- Path C: Three stacked 3x3 convolutions (equivalent to a 7x7 receptive field).
- Path D: Four stacked 3x3 convolutions (equivalent to a 9x9 receptive field).
Unlike models that concatenate parallel paths, Union-net utilizes element-wise addition for information fusion. This keeps the model lightweight and preserves the feature map shape. The full network is relatively shallow, utilizing only three Union modules in series, followed by Global Average Pooling (GAP) to drastically reduce the parameter count by eliminating dense layers.
Experimental Results and Efficiency
Union-net was evaluated against several "small data" scenarios, including subsampled versions of CIFAR-10, MNIST, and the 17Flower dataset. On the CIFAR10_30 benchmark (only 30 images per class), Union-net achieved 89.88% accuracy, significantly outperforming standard architectures like ResNet50 (82.88%) and MobileNetV2 (85.49%).
A key strength of Union-net is its extreme efficiency. It achieves competitive results with just 1.72 million parameters, compared to 138 million for VGG16. Its computational cost is also remarkably low, with FLOPs sitting at a mere 0.003B.
Conclusion
Union-net demonstrates that a shallow, multi-scale approach can overcome the limitations of deep learning on small datasets. By focusing on parameter efficiency and receptive field diversity, it provides a robust alternative to transfer learning for specialized domains where data is a luxury rather than a commodity.
Summary updated