The goal of this competition was to predict the category of a visual stimulus presented to a subject from the concurrent brain activity (captured with an MEG device which records 306 timeseries at 1KHz of the magnetic field associated with the brain currents). The categories of the visual stimulus for this competition was the presentation of a face or a scrambled face.

While being a pretty standard Neuroscience experiment, the originality of this challenge was that subjects from the training set and the test set were different. As a result many state-of-the art algorithm relying on spatial filtering or source separation were not as effective.

For solving this problem, I used a combination of transfer learning and unsupervised classification. A first classifier, based on Riemannian Geometry, was trained on all training subjects together, and then used as an initialization for unsupervised iterative training (self-learning) on each subject of the test data. Luckily enough (!) this procedure converged toward a viable solution for all subject, and allowed me to take a comfortable advance in the challenge (with the winning solution being made 45 day before the end of the challenge, an eternity by Kaggle standards.)