Computer Vision News - September 2016

more than happy to find the name of Luisa Zintgraf (one of the authors of the research which we reviewed in Computer Vision News of July 2016 ) as high as the 7th place among the overall 688 participants . So we asked her about her experience participating in this challenge. It seems that she actually agrees on some points with the winner of the challenge; here is what she told us off the record (but kindly agreed to let us publish): “ It was a really fun challenge, I must say (and the first one I ever did, so I was quite happy about the 7th place). I spent most of my time on feature engineering and feature selection, actually. I talked to some other people who also participated in the challenge who spent a lot of time and effort on training deep neural networks, but that didn't work out too well: accuracies were below 70%! There were just not enough data points to train a good neural net. So I used a logistic regression classifier (because it's fast and I could test many different things) and focused on the features. One thing that really boosted performance was for example to use only a subset of the electrodes, and this subset was specific to the patient. Again, I think the biggest challenge was that there were so many more features than data points, which is why I spent most of my time reducing the features/noise (like taking a smaller time frame than 800ms, etc.). I did everything per-patient, and had some discussion about this with the other participants I talked to (we actually teamed up in a Kaggle competition now!). They said that it's unrealistic to have specific features per patient, and that in reality it would be better to generalize, and have the same preprocessing+model for all patients. But I strongly disagree! The performance is so much worse, also because the electrodes are at different places in the brain and therefore the signals don't have the same meaning across patients . I think it should be totally feasible to do some feature engineering and individual model training per patient (also since the potential benefits of decoding brain signals for example of paralyzed patients are huge). And even if that's too expensive, it would be possible to automate the feature selection process to some extent.” “A really fun challenge, and the first one I ever did!” “There were just not enough data points to train a good neural net” Challenge Computer Vision News Challenge 29 Luisa M. Zintgraf is graduate of Artificial Intelligence at the University of Amsterdam, Netherlands.