Year 1, Month 7, Day 18
(6 bulan dan 18 hari telah berlalu)
======= = =======
PyMVPA Training Course in Delmenhorst, 6th-7th March 2014
Multi-variate Pattern Analysis with Python
Two weeks ago, I attended a training course in Hanse-Wissenschaftskolleg
, Institute for Advanced Study in Delmenhorst, Lower Saxony, Germany (wiki: de
). The Institute of Advance Study
in Princeton, New Jersey, USA, is famous of Paul Dirac, Albert Einstein, Noam Chomsky and more than 30 Nobel Laurates (wiki: en
). Well, I only came to the one in Delmenhorst
, instead of the one in Princeton. There I learn PyMVPA, which is Python toolbox to don Multi-variate Pattern Analysis (MVPA).
To be in the PyMVPA training course, I had to submit an application. There are many interested applicants from all over Europe and half of them are rejected. The training is more useful for scientists, doing fMRI than EEG research, since the course title is “Multivariate Pattern Analysis (MVPA) of Neuroimaging Data with PyMVPA” (link
). I don’t know why I got accepted, since I am doing EEG research. Maybe it’s because of I am a student from the University of Oldenburg
, as the organizer of this event. Well, I did write about my Python experience and something about multivariate statistic in my application form.
Before the course, I had to prepare some python ability. They told me about the prerequisites of PyMVPA training.
Then I refreshed my python skills and learned NumPy, since PyMVPA is developed on NumPy.
What I learn in the PyMVPA course can be found below.
The courses consist of lectures and hands-on. So they gave me 20-30 minutes of lecture and about 60-90 minutes hands-on, then a coffee break. So we don’t have to feel bored of lectures.
Day 1 hands-on
Day 2 hands-on
The course use debian environment, to run IPython. So virtual machine NeuroDebian is used. VirtualBox from Oracle is used to open NeuroDebian image.
In the first day, I learned fMRI data processing for the first time. The fMRI data are big but they are well structured nowadays. But different research groups use different structure. The data structure contains voxels (volume pixels) per time, events and other information. With EEG signal processing, I had experience with 128 Hz and 2048 Hz sampling frequency. But in fMRI, 2 seconds of sampling period (0.5 Hz) is considered fast and it can turn into Big Data. Well, EEG measures electric fields on scalp and fMRI measures changes of blood flow inside the head, e.g. blood oxygenation-deoxygenation. With fMRI, we can measure until 1.4 mm of voxel. The dimension of EEG are number of channels and time but the ones of fMRI are voxels (3D) and time. The data are BIG!
In the end of first day, they trained “Searchlight” method to process fMRI data. Voxel per voxel for each time are analyzed in a region of interest. This method is computationally expensive. The laptop, I used, were getting warm and I had to wait until the program finished running. On the second day, they showed their fMRI data from 20 participants, in 2-hour experiment with 7-Tesla fMRI in 2 s TR and 1.4 mm voxel. In the experiment, the participants heard audio-movie Tom Hanks, “Forrest Gump”. The experiment produced 355 GB of data, which could be freely “downloaded” from StudyForrest website
. Well, 355 GB processing need computing effort. Laptop can be burning. It is better to use supercomputer or something else that can handle Big Data.
In the second day, they trained pattern analysis, especially clustering. The statistical evaluation using PyMVPA was taught because most scientific papers and the reviewers would ask how statistically significant our experiments are. Prof. Hanke gave tips about publishing papers in Nature: “Get good data and use the same data again and again for different papers”. Well, they could make participants stay for 2 hours in fMRI scanner while other researchers could only conduct for 20-30 minutes. If the experiments are short time, get more participants (more than 50). Big Data sells!
Well, today is the era when scientists measure everything and hope that computers will do the job for them and punch you in the nose with results.
From the course, I met researchers and health-care workers from Europe: Netherlands, Italy, and so on. I also found out that Uni Magdeburg would be an interesting place to do research in neuroscience. Well, I met a female researcher from Lübeck, whom I had met in Uni Bremen. She studied Psychology and I did Electrical Engineering in Bremen. Like many other scientists, she had Panda eyes, due to sleep deprivation. I began to think whether I should reduce my sleeping hours for research or let nobody or nothing steal them.
Well, it was good experience to meet a lot of people from different institutes. I could get ideas of what they had been doing, as well as, the future challenges in neuroscience research. I got their contacts for collaboration or my future career plan in their research group. Well, I am still in the beginning of my doctoral study and I haven’t got a good picture of where my research is going, yet.
I have no idea whether I am going to use PyMVPA for my EEG analysis although Prof. Hanke told me that it was possible. I am still learning to analyze my EEG data with EEGLAB in MATLAB. I am also taking a Multi-variate Statistic class, in which R is used as statistical tool. Not to forget C++ to run some experiments. Well, I think I learn too many programming languages: MATLAB, C++, Python and R. Although none has been useful of producing good data, yet.
About my preparation before attending PyMVPA course can be read in other blog posts:
Science will find a way. Darah Juang!
Oldenburg, 18 Maret 2014
Darah Juang Doktoral http://drhdrdro.blogspot.com/2014/03/pymvpa-training-course-note-y01-m07-d18.html