Features Disentanglement For Explainable Convolutional Neural Networks

Explainable methods for understanding deep neural networks are currently being employed for many visual tasks and provide valuable insights about their decisions. While post-hoc visual explanations offer easily understandable human cues behind neural networks’ decision-making processes, comparing their outcomes still remains challenging. Furthermore, balancing the performance-explainability trade-off could be a time-consuming process and require a deep domain knowledge. In this regard, we propose a novel auxiliary module, built upon convolutional-based encoders, which acts on the final layers of convolutional neural networks (CNNs) to learn orthogonal feature maps with a more discriminative and explainable power. This module is trained via a disentangle loss which specifically aims to decouple the object from the background in the input image. To quantitatively assess its impact on standard CNNs, and compare the quality of the resulting visual explanations, we employ metrics specifically designed for semantic segmentation tasks. These metrics rely on bounding-box annotations that may accompany image classification (or recognition) datasets, allowing us to compare both ground-truth and predicted regions. Finally, we explore the impact of various self-supervised pre-training strategies, due to their positive influence on vision tasks, and assess their effectiveness on our considered metrics.