Tuesday's Paris Machine Learning Meetup #3 went very well. Thank you to Frederick Demback and his crew who welcomed us at DojoEvents. Guillaume Pellerin kindly provided the TC-202 Case to record the three talks and put them on YouTube. In all, about 70 people showed up out of 120 who had registered.
Because of the heterogeneity of the group, Franck and I decided to start with a brief 20 minutes overview of Machine Learning. This presentation was followed with two interesting talks on scikit-learn and its use to classify fMRI events for mind reading. The talks were in French but the presentations slides are in English.
- Machine Learning: What is it good for ? Franck Bardol, Igor Carron
- Scikit-learn: une boite à outils de machine learning, Gaël Varoquaux, Research faculty, Parietal team, INRIA et Associate researcher, Unicog team (cognitive neuroimaging unit), INSERM
- Mind Reading with Scikit-learn / Lire dans les pensées avec le Scikit-learn, Alexandre Gramfort, Assistant Professor, Telecom ParisTech
During Gaël's and Alexandre's talks. we stumbled upon the issue of what it meant to be brain dead. I think we need to see a talk on that subject. In effect, the underlying question revolves around the following issue: what sort of algorithm can make sense of low level conscience and no conscience at all with noisy measurements. I also did not realize that the data from Gallant's lab at Berkeley test mentioned earlier here was not publicly available. As an aside, it is also interesting to note that there seems to be little competition taking place to get better reconstruction results from datasets that have already been gathered as in the case of data from [3] featured in the video below. Alexandre showed us an awesome reconstruction using scikit-learn using many fewer lines of codes than the original attempt [3]. In retrospect, the situation looks very similar to what took place in the initial development of sparse coding/dictionary learning algorithms. While in 1996 [4] it took several hours to get results, much faster algorithms have since surfaced providing better reconstructions as well. As aside, I loved the use of IPython.
Gaël and Alexandre's talks lasted a full hour and twenty minutes in all, far more than the scheduled 40 minutes. Thank you to them for being good sports while fielding our long stream of questions.
As an aide, at the end of the first talk, I was asked about what I meant about french specific datasets. One person rightly mentioned cultural biased datasets. I also meant to encompass subject areas that only local folks care for or would try to spend some time cracking. I gave a not-so-optimal example but here are four that I have thought about since:
- AFFELNET, the algorithm used by the each state's education office to assign students to different high school. Most parents are interested in figuring out how it really works.
- Price index for appartments in Paris:
You think you know some interesting time series, here is one, the price index for appartment in Paris from 1200 till 2013 [1,2]
there you can read explanation for trends like this one:
The decrease in years 1350-1450 compared to years 1200-1350 might have been caused by the Hundred Year’s War and the Black Plague, the increase from 1450 to 1550 by the return to relative peace, the pause in 1550-1600 by the Wars of Religion, and the following steep rise by the return to relative peace as well as the residence of French kings in the capital, but these are mere speculations.
Here's something you don't hear that everyday: an explanation for pricing time series fluctuations and how they are influenced by the Black Plague.
- TV programs
It is the same all around the world but one wonders if there an ML type of analysis undertaken by the different TV companies to evaluate how US TV series re-runs compete against more expensive local productions.
- Learning analytics:
This one is culturally biased for sure especially when it comes to learning the language.
Other smaller discussions after the presentations included a connection between imaging and Machine Learning (Sunday Morning Insight: A Quick Panorama of Sensing from Direct Imaging to Machine Learning), the use of count-min (see Muthu's count-min sketch & its applications page) and other sketches in the data streaming model (Sketching data structures), the series of amazing videos from GenomeTV (some selected pieces can be found here) and Moore's law.
[1] Comparing Four Secular Home Price Indices Jacques Friggit, Working paper, version 9, June 2008 , from House Prices in France : Property Price Index, French Real Estate Market Trends, 1200-201326/08/2013
[3] Visual Image Reconstruction from Human Brain Activity using a Combination of Multiscale Local Image Decoders, Yoichi Miyawaki,Hajime Uchida,Okito Yamashita,Masa-aki Sato,Yusuke Morito,Hiroki C. Tanabe,Norihiro Sadato,Yukiyasu Kamitani
[4] Emergence of Simple-Cell Receptive Field Properties by Learning a Sparse Code for Natural Images, Olshausen BA, Field DJ (1996). Nature, 381: 607-609. reprint (pdf) | abstract
[3] Visual Image Reconstruction from Human Brain Activity using a Combination of Multiscale Local Image Decoders, Yoichi Miyawaki,Hajime Uchida,Okito Yamashita,Masa-aki Sato,Yusuke Morito,Hiroki C. Tanabe,Norihiro Sadato,Yukiyasu Kamitani
[4] Emergence of Simple-Cell Receptive Field Properties by Learning a Sparse Code for Natural Images, Olshausen BA, Field DJ (1996). Nature, 381: 607-609. reprint (pdf) | abstract
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.
