With sufficient layers, enough training data, enough time, and often a custom tailored architecture, modern deep learning methods can be extremely successful in classification tasks. However, with real-world practical applications, a lack of training data can impede the success of such techniques, which often over-fit the sparse data. A particular task hindered by sparse data, may take advantage of an existing pre-trained network of a separate but related task trained on sufficient data, to achieve high accuracy with minimal additional training. However, such transfer-learning does not generalize well to new tasks that have little relation between their data and the pre-trained network's data. To overcome this hurdle, we introduce a novel feature (AngOri) kernel that leverages the generalized inherent richness of curvature and gradient in image edges, and that can augment any type of convolutional neural network with any arbitrary number of channels to quickly achieve accurate classification with minimal training on sparse data. The AngOri kernels can be pre-computed and thus directly implemented in a convolutional layer of a network, which is not possible with other gradient based features. Testing on the MNIST, CIFAR-10, and a satellite image database, we consistently found AngOri to aid a network to achieve more accurate classification on a small dataset when compared to the same network without the AngOri layers. Such a generalizable, lightweight kernel holds promise for using neural networks to tackle real-world problems with limited resources, such as embedded systems examining sparse data.
Paru en décembre 2019 , 14 pages