Katharina Morik, TU Dortmund University, Germany, will be joining us on Thursday May 29th, to give a special talk on “Resource-Aware Graphical Models.” We would love for you to join us and this special one-off presentation in SF. To save your spot you just need to register here. The evening will start with drinks, snacks and networking. Please note the talk will start promptly at 6:30PM.
AbstractMachine learning can help to enhance small devices. For instance, keeping the energy consumption of smart phones low is one of the major concerns of the users, as is well illustrated by various “charge your mobile” stations at public places. Where the operating systems of smart phones already offer heuristics and battery apps show consumption profiles, machine learning can do more. Predictions allow better optimizations of the operating system, prepare for particular app usages at certain points in time, or manage services such as GPS or WLAN in a context-aware and adaptive manner. This challenges learning algorithms to real-time application of their models. Moreover, it demands the models to run on the resource-restricted device without consuming more energy themselves than they save! In the talk, graphical models are presented that face these challenges. Using Conditional Random Fields (CRF) for the prediction of files that the user will fetch next on her smart phone can be used by the operating system for organizing the memory. Analyzing groups of apps running on the smart phone may estimate the energy consumption over time. A novel spatio-temporal random field (STRF) has been implemented, smoothing the temporal changes and distributing the optimization. This graphical model has been used to predict app usage over time. In another application, it has been combined with a trip planner resulting in smart routing for smart cities. In order to run graphical models on very restricted devices, even those without floating point calculation, one computing with integer values only has been developed. The integer approximation of graphical models shows good accuracy and speed-up and opens up novel applications on resource-restricted devices.
For those interested here is Katharina’s suggestion for further reading
- Nico Piatkowski, Sangkyun Lee, Katharina Morik (2013). Spatio-Temporal Random Fields: Compressible Representation and Distributed Estimation, in: Machine Learning Journal, Vol. 93, No. 1, S. 115 – 139.
- Jochen Streicher, Nico Piatkowski, Katharina Morik, Olaf Spinczyk (2013). “Open Smartphone Data for Mobility and Utilization Analysis in Ubiquitous Environments” in: Mining Ubiquitous and Social Environments (MUSE) workshop at ECML PKDD
- Thomas Liebig, Nico Piatkowski, Christian Bockermann, Katharina Morik (2014). Predictive Trip Planning – Smart Routing in Smart Cities, in: Mining Urban Data Workshop at 17th International Conference on Extending Database Technology.
- Nico Piatkowski, Sangkyun Lee, Katharina Morik (2014). The Integer Approximation of Undirected Graphical Models, in: 3rd Int. Conf. on Pattern Recognition Applications and Methods (ICPRAM).