Scene Novelty Prediction from Unsupervised Discriminative Feature Learning

Safety Monitoring of Neural Networks Using Unsupervised Feature Learning and Novelty Estimation

“…Using the experimental setup from [40] on CIFAR-10, [20], expanded upon in [19] for CIFAR-100 [20], one class is considered to be in distribution and the rest anomalies. In total, each dataset D is split into N D classes, and corresponding N D experiments are performed.…”

“…For CIFAR-100 the 20 superclasses are used as labels, instead of the 100 regular classes, to reduce the number of experiments. The benchmarking is done against a mix of conventional methods, autoencoders, GANs and metric learning based approaches, in addition to the original approach proposed in [19]. The benchmarked methods used are:…”

“…In the third experiment the novelty estimation will instead be used as an accuracy prediction and confidence estimation for a typical driving function. Using a similar experimental setup as in [19], the correlation of the novelty estimation is evaluated against a basic segmentation network, SegNet [22].…”

“…This paper gives a detailed description of the novelty detection framework. The main contribution is an extension of the method developed by the authors in [19] providing in-depth analysis and a better clustering approach to the original framework used to estimate the novelty of the current driving environment.…”

“…• A novelty estimation framework using unsupervised feature learning [19], further developed utilizing a better clustering approach and metric for the novelty estimation, increasing the performance. This also allows the method to be extended into finding novel segments of an image.…”