Towards classification trustworthiness: one-class classifier ensemble

Many autonomous safety-critical systems rely on neural networks for image classification. While they achieve high accuracy, their decisions are hard to interpret. Also, a known issue with neural networks is that they tend to provide high probabilities for unknown images. Uncertainty on how neural networks will behave is a challenge to safety. To address these issues, this paper presents a different approach by using an ensemble of one-class autoencoder. This architecture can both classify images and detect unknown images. The experiments show that, for certain datasets, it achieves a similar accuracy compared to state-of-the-art neural networks, while being robust to unknown images.