PathNet: a novel multi-pathway convolutional neural network for few-shot image classification from scratch

In recent years, advanced computer vision models have trended toward deeper and larger network architectures, and model depth is often considered an important feature for achieving superior performance. While deeper networks can help solve complex vision tasks, they also raise issues such as model space complexity and parallelization of long tandem block structures. Therefore, we revisit the network design space by using a shallower depth to reduce the complexity of the vertical spatial structure, horizontally extending multiple computational pathways to improve model capacity and scalability, and highly optimizing the internal modeling to fully exploit the inherent inductive biases in ConvNet and the intrinsic benefits of global attention to improve the overall performance. In this paper, we propose a novel 16-layer shallow depth multi-pathway parallel convolutional neural network, called PathNet, which can be used as a generic backbone for few-shot image classification. We evaluate the effectiveness of PathNet by training from scratch. Experimental results show that PathNet achieves a top-1 accuracy of 54.06% on the Oxford Flowers-102 dataset, 95.16% on the Cifar10 dataset, 76.02% on the Cifar100 dataset, and 61.85% on the TinyImageNet dataset, providing a competitive advantage and great potential in terms of accuracy, scalability, and parallelization compared to the state-of-the-art models.