BCI and Machine Learning

[ilja.kuzovkin]

I'm a student, who was working with AI techniques and suddenly became interested in BCI field.
And the very next idea was to join BCI and ML.
Surprisingly I was not able to find any practical realizations of this idea... only some theoretical articles...
So, I thought "Hmm, seems I have something to do my bachelor paper about..."

An idea:
1. Put an BCI device on head
2. Think "shovel"
3. Get the signal from the BCI and feed it to the Machine Learning algorithm saying "That was the 'shovel'"
4. Repeat step 3 N times
5. Think "right"
6. Get the signal from the BCI and feed it to the Machine Learning algorithm saying "That was 'right'"
7. Repeat step 5 N times
..
..
..
n. End training
n+1. Think "shovel"
n+2. ML algoritm will say: "Hey! I know that signal - it seems like 'shovel'"

Questions
a) How possible it is? I mean how many "shovels" such technique will be able to distinguish?
b) Wasn't is already done like 10 years ago?
c) What does "in-field-experienced" people think about this idea in general? ("in-field-experienced" people - looking forward to get answer from you )

[yrenard]

Dear ilja.kuzovkin,

thank you for your interest in OpenViBE and welcome on this forum.

Actually, BCI use a lot of machine learning techniques such as classification algorithms and / or knowledge based signal processing. I suggest that you watch the 10 mns video presenting the OpenViBE software and get more BCI background from the very good http://future-bnci.org website.

I hope this helps,
Yann

[fabien.lotte]

Dear ilja.kuzovkin,

Machine learning has indeed be used a lot in the design of BCI systems. It was mostly introduced in the BCI community by the Berlin BCI group (http://www.bbci.de), who published on number of papers on this topic. You may want to have a look at their papers to learn more. See, for instance, this paper:

MACHINE LEARNING TECHNIQUES FOR BRAIN-COMPUTER INTERFACES
K.-R. Müller, M. Krauledat, G. Dornhege, G. Curio, B. Blankertz
Biomed. Tech., 2004
pdf: http://citeseerx.ist.psu.edu/viewdoc/do ... 1&type=pdf

However, machine learning has not been used to recognize any type of neurophysiological patterns, but only some very specific ones like ERD/ERS (mostly due to motor imagery), P300 or SSVEP. Indeed, these signals are known to be detectable in EEG signals. But as far as I know, there is no paper reporting to detect "shovel" or something similarly exotic in EEG signals. Actually, it may be worth exploring such alternative mental tasks or neurophysiological patterns, since most of the BCI community is only using the patterns I mentioned above. To do so you will indeed need machine learning. But you have to realize that it won't be easy, since there are probably a lot of mental tasks that can't be recognized from EEG. But if you find one that can be recognized and that has not been explored before, then you will have something new and interesting for your Bachelor paper!