BNAIC 2017 Session: Machine learning, Wednesday 8 November 2017, 09.10h

Session Chair: Martijn van Otterlo

////BNAIC 2017: Session report “Machine Learning”

After the opening remarks by the conference organizers and an introduction to the scientific context in Groningen, this session featured the first four paper presentations of the conference in the main lecture room. It had an unfortunate start due to a mix of cable and hardware issues, but speakers and chair together solved these cooperatively and very quickly and then the session started with a presentation by Claudio Reggiani titled “Feature selection in high-dimensional dataset using MapReduce”, which is joint work with Yann-Aël Le Borgne and Gianluca Bontempi. Claudio presented results of a MapReduce implementation for feature selection (in this case the minimal Redundancy Maximal Relevance filtering technique). The (open source code) results showed good scalability performance.

The second presentation was given by Dirk van der Hoeven and was titled: “Is Mirror Descent a Special Case of Exponential Weights?” which is joint work with Tim van Erven. The presentation managed to deliver a clear summary of fairly technical work in online convex optimization. One of the goals of the work is to relate online gradient descent, mirror descent, and exponential weights. Possible applications of the main results and future work include efficient sampling in linear bandits, priors for learning rates, and scale free algorithms.

The third presentation was delivered by Siamak Mehrkanoon (joint work with Johan Suykens) and titled: “Regularized Semi-Paired Kernel CCA for Domain Adaptation”. The method presented can be placed in the context of transfer learning algorithms and is aimed at generalizing a model trained on a source task to a target domain by taking into account how many labeled instance in both tasks are available. Experimental results showed that a joint representation of the data set across different domains can be learned and utilized.

The final presentation in this session was given by Lynn Houthuys who presented joint work with Zahra Karevan and Johan Suykens on “Multi-View LS-SVM for Temperature Prediction”. In this work, applied to a temperature prediction task for cities in Belgium, multiple views of the data (representing each city) are combined in a support vector machine setting. The proposed method enforces alignment of error variables over multiple views, distributed over multiple LS-SVMs, such that when training one SVM on one view, the other views are taken into account. The experiments show that the new approach outperforms both existing weather predictions as well as predictions based on naively concatenating all available features. This session had four speakers who all managed very well to deliver their presentation in time, which resulted in a lot of time for discussion after each talk, which was filled with many questions from the audience and the chair, even on this early hour of the conference.

By |2018-03-01T20:11:02+00:00March 1st, 2018|Event reports|0 Comments