Out-of-distribution Detection and Generation using Soft Brownian Offset Sampling and Autoencoders
Authors:
Felix Möller,
Diego Botache,
Denis Huseljic,
Florian Heidecker,
Maarten Bieshaar,
Bernhard Sick
Abstract:
Deep neural networks often suffer from overconfidence which can be partly remedied by improved out-of-distribution detection. For this purpose, we propose a novel approach that allows for the generation of out-of-distribution datasets based on a given in-distribution dataset. This new dataset can then be used to improve out-of-distribution detection for the given dataset and machine learning task…
▽ More
Deep neural networks often suffer from overconfidence which can be partly remedied by improved out-of-distribution detection. For this purpose, we propose a novel approach that allows for the generation of out-of-distribution datasets based on a given in-distribution dataset. This new dataset can then be used to improve out-of-distribution detection for the given dataset and machine learning task at hand. The samples in this dataset are with respect to the feature space close to the in-distribution dataset and therefore realistic and plausible. Hence, this dataset can also be used to safeguard neural networks, i.e., to validate the generalization performance. Our approach first generates suitable representations of an in-distribution dataset using an autoencoder and then transforms them using our novel proposed Soft Brownian Offset method. After transformation, the decoder part of the autoencoder allows for the generation of these implicit out-of-distribution samples. This newly generated dataset then allows for mixing with other datasets and thus improved training of an out-of-distribution classifier, increasing its performance. Experimentally, we show that our approach is promising for time series using synthetic data. Using our new method, we also show in a quantitative case study that we can improve the out-of-distribution detection for the MNIST dataset. Finally, we provide another case study on the synthetic generation of out-of-distribution trajectories, which can be used to validate trajectory prediction algorithms for automated driving.
△ Less
Submitted 4 May, 2021;
originally announced May 2021.
Rooting Formal Methods within Higher Education Curricula for Computer Science and Software Engineering -- A White Paper
Authors:
Antonio Cerone,
Markus Roggenbach,
James Davenport,
Casey Denner,
Marie Farrell,
Magne Haveraaen,
Faron Moller,
Philipp Koerner,
Sebastian Krings,
Peter Olveczky,
Bernd-Holger Schlingloff,
Nikolay Shilov,
Rustam Zhumagambetov
Abstract:
This white paper argues that formal methods need to be better rooted in higher education curricula for computer science and software engineering programmes of study. To this end, it advocates (i) improved teaching of formal methods; (ii) systematic highlighting of formal methods within existing, `classical' computer science courses; and (iii) the inclusion of a compulsory formal methods course in…
▽ More
This white paper argues that formal methods need to be better rooted in higher education curricula for computer science and software engineering programmes of study. To this end, it advocates (i) improved teaching of formal methods; (ii) systematic highlighting of formal methods within existing, `classical' computer science courses; and (iii) the inclusion of a compulsory formal methods course in computer science and software engineering curricula.
These recommendations are based on the observations that (a) formal methods are an essential and cost-effective means to increase software quality; however (b) computer science and software engineering programmes typically fail to provide adequate training in formal methods; and thus (c) there is a lack of computer science graduates who are qualified to apply formal methods in industry.
This white paper is the result of a collective effort by authors and participants of the 1st International Workshop on "Formal Methods, Fun for Everybody" which was held in Bergen, Norway, 2-3 December 2019. As such, it represents insights based on learning and teaching computer science and software engineering (with or without formal methods) at various universities across Europe.
△ Less
Submitted 12 October, 2020;
originally announced October 2020.
