Over the course of organizing the Paris Machine Learning Meetup, we saw Kaggle competitions turning the crowdsourcing of ML challenges as a way to identify interesting algorithms: deep learning became mainstream because of its ability to do well on image classification competitions, but we could also witness the effectiveness of Random Forests and more recently the rise of VowPal Wabbit. Yet there is still a need to convey many of these changes to the rest of the community around us. That community includes scientific colleagues, engineers and even VCs if you are trying to raise money. Here are two blog entries on the subject:
- Balazs Kegl talks about The Data Science Ecosystem
- Laurent Duval introduces us to Dédoménologie : la science du traitement de données (signal, images, etc.) - it's in French, sorry -
Go read these entries, I'll wait.
I think Laurent [2] and Balazs are also seeing the great convergence in action and want to make sure everyone knows about it.
The great convergence will speed up as we gain a better theoretical understanding of the different parts. However, it shouldn't stop us from moving forward as it may take some time: In the mid-1990s, while NMF and sparse coding changed the way we dealt with data, some of their theoretical explanation did not emerge until recently and it certainly did not stop practitioners from using them extensively. With good reason, in compressive sensing, we even had a 200 year gap !
So when Pavel Hála did an outstanding job [3] for his MS thesis and his advisor gave him a bad grade because, you know, it's a black box and unproven.... Data Science is wild, I had to respond in some unceremonious manner.
[1] A group that has more than 2250 members and growing - archives of the meetup are here
[3] Thesis: Spectral classification using convolutional neural networks, Pavel Hála - implementation -
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.
No comments:
Post a Comment