Machine Learning for Signal Processing (MLSP) TC Announces the Winners of the 6th Annual Data Analysis Competition

Contributed by Kenneth E. Hild II (TC Liaison to eNews)

For each of the past 6 years, the Machine Learning for Signal Processing (MLSP) TC has hosted a data analysis competition. The specific application varies from year to year. In previous competitions, the applications included: blind source separation (BSS) of instantaneously and convolutively-mixed sources; denoising of evoked magnetoencephalographic (MEG) data; processing of functional magnetic resonance image (fMRI) data; artifact rejection for event-related electroencephalographic (EEG) data; maximizing the annual rate of return by trading stocks over a six-month period; and applying computer vision methods to find a specific person in a set of images. The goal for this year's competition, MLSP 2010 Competition: Mind Reading, was to develop a feature extraction method and classifier to detect the neural signatures in EEG data that occur when the subject detects an instance of a pre-defined target during an image presentation task (using the oddball paradigm).

The competition this year was a resounding success. We received and evaluated entries from a total of 35 different teams. Performance was gauged using the area under the receiver operating curve (AUC), which was estimated from a disjoint test set (for which the entrants did not have prior access; follow this link for more details). The affiliations of the top three teams are: (1) Universidad Carlos III de Madrid and Max Plank Institute for Biological Cybernetics, (2) Istanbul Technical University, and (3) INSA de Rouen and Universite de Rouen. The entries of each of these three teams had an AUC of approximately 0.82. At the upcoming MLSP conference, in Kittila, Finland starting on 29 August 2010, we plan to give awards to these three teams. The awards include several N900 high-performance mobile computers, courtesy of Nokia, and a total of 1500 Euros to be distributed as travel grants, courtesy of the PASCAL2 Challenge Program.