Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem.
Matthias HeinMaksym AndriushchenkoJulian BitterwolfPublished in: CVPR Workshops (2019)
Keyphrases
- high confidence
- training data
- class labels
- association rules
- training set
- learning algorithm
- network structure
- test set
- data sets
- heterogeneous networks
- test data
- decision trees
- training examples
- training dataset
- training process
- social networks
- unlabeled data
- small number
- prior knowledge
- pairwise
- training samples
- semi supervised learning
- supervised learning
- classification models
- support vector machine
- network design
- image processing