* Code and data are not published at the moment due to legal issues. We are working on it!
Abstract
Real-world datasets collected with sensor networks often contain incomplete and uncertain labels as well as artefacts arising from the system environment. Complete and reliable labeling is often infeasible for large-scale and long-term sensor network deployments due to the labor and time overhead, limited availability of experts and missing ground truth. In addition, if the machine learning method used for analysis is sensitive to certain features of a deployment, labeling and learning needs to be repeated for every new deployment. To address these challenges, we propose to make use of system context information formalized in an information graph and embed it in the learning process via contrastive learning. Based on real-world data we show that this approach leads to an increased accuracy in case of weakly labeled data and leads to an increased robustness and transferability of the classifier to new sensor locations.
(tl/dr)
- Real-world datasets are messy (e.g. those coming from sensor networks)
- Complete labeling is infeasible (labor and time overhead, experts are required)
- Models are usually not robust against data distribution changes
Contributions
- We present a method which uses system context information to counteract the negative impact of few and weak labels by combining contrastive learning with an information graph.
- We present a unified learning process in which annotations are encoded as Gaussian random vectors to treat them similar to data.
- We demonstrate on a dataset gathered from a real-world deployment in the Swiss alps, how the method can be used to train a classifier with improved generalization performance across sensors with diverging characteristics.
Method
Results
Further Information
More information about the actual deployment at Illgraben, Switzerland including impressive drone footage, can be found in the video made by Werkstattgespräche ETH Zürich. (Swiss German only, sorry) and in this paper by Wenner et al., 2021. Quoting from the paper
The Illgraben catchment in southwest Switzerland is one of the most active mass wasting sites in the European Alps [...] Yearly precipitation is controlled by summer rainstorms with high rainfall intensity during which mass wasting with rock-slope failures and debris flows occur regularly [...]. From its highest point at the Illhorn (2716 m a.s.l.), the Illgraben catchment reaches down to the Rhone Valley (600 m a.s.l.), where its main torrent flows into the Rhone River.
Media Sources
Illgraben 29.05.2017 - Lave torrentielle, Murgang, (Video)
Der Bergsturz von Bondo - Wie die Klimaerwärmung die Schweizer Alpen bedroht | SRF Einstein (Video)
Bondo Rockfall (Video)