ONLINE: Self-aware AI: How can a neural network learn to say „I do not know“?

The duration of the presentation is about 45 minutes

Abstract: 

Deep learning stands at present as the most powerful generic tool for developing intelligent systems. Deep neural nets have reached unprecedented prediction accuracies, to the extent that they have outperformed the best bioinformatics techniques in protein folding prediction and the best language understanding methods in text analysis without any domain knowledge. Neural networks owe their success to their capability of memorizing data in a parametric way.
A successful neural network architecture typically has a larger number of parameters than observations. This property causes a severe artifact: Neural nets are not capable of propagating uncertainty through their layers properly, thus they are not capable of attributing sensible uncertainties to their predictions. This causes a neural net trained on handwritten digits to assign the picture of a cat to a random digit with high confidence, making its safe and reliable use in industry scenarios questionable.

This talk will introduce the basics for uncertainty quantification in neural nets, walk through a compilation of recent research challenges in this area, and provide insights on key industry use cases of uncertainty-aware AI methods.