Why ReLU networks yield high-confidence predictions far away from the training data and how to mitigate the problem.
Matthias HeinMaksym AndriushchenkoJulian BitterwolfPublished in: CoRR (2018)
Keyphrases
- high confidence
- training data
- class labels
- association rules
- learning algorithm
- training set
- decision trees
- data sets
- domain knowledge
- social networks
- software defect
- training process
- test data
- test set
- training samples
- semi supervised learning
- social network analysis
- supervised learning
- complex networks
- computer networks
- noisy data
- network analysis
- network model
- classification accuracy
- heterogeneous networks
- training instances
- learned models
- prior knowledge
- network size