Resumen de Active Image Data Augmentation

Flávio Arthur O. Santos, Cleber Zanchettin, Leonardo Nogueira Matos, Paulo Novais

• Deep neural networks models have achieved state-of-the-art results in a great number of different tasks in different domains (e.g., natural language processing and computer vision). However, the notions of robustness, causality, and fairness are not measured in traditional evaluated settings. In this work, we proposed an active data augmentation method to improve the model robustness to new data. We use the Vanilla Backpropagation to visualize what the trained model consider important in the input information. Based on that information, we augment the training dataset with new data to refine the model training. The objective is to make the model robust and effective for important input information. We evaluated our approach in a Spinal Cord Gray Matter Segmentation task and verified improvement in robustness while keeping the model competitive in the traditional metrics. Besides, we achieve the state-of-the-art results on that task using a U-Net based model.