Machine learning in brain-computer interfacing lecture
This is obviously fascinating stuff. The lecture is by Professor Klaus-Robert Mueller, who works at the technical university of Berlin and has studied mathematical physics and theoretical computer science, in addition to his work on brain computer interfacing (also known as BCI). Here is a link to the lecture, and although the streaming is a bit choppy the work they are doing is worth the wait. Here are my notes:
The general structure for their non-invasive BCI is a brain, some measuring device, some feature abstraction, some classifier, and some feature output. The goal could be to control some device using neural transmissions.
One challenge is that if you measure electrical activity, you can only measure a certain superposition of electroactivity at given brain sources. The problem is that there is not a single brain but many quasi-independent pathways working together. This is known as the cerebral cocktail party problem.
One way to sort between sources of interest and non-interest (and “solve” the cerebral cocktail problem) is to use independent component analysis (aka ICA, info here). It attempts to tell which independent compents are the source of the electrical acitivity data.
In their research, they wanted to both predict motor movements before they occured, and predict motor movements based on humans imagining the movement.
Their “imagining movement” study including a 20 minute training phase (where they collected ~ 200 data points), a 5-10 minute machine learning phase, and then a feedback phase where the information learned is tested. After this point, the subjects were able to play computer pong with their brains!
Peak bit rates for the feedback phase reached 37 bits per minute, although the overall bit/rate varied substantially within subjects (there were 6).
They were also able to have the subjects spell words using imagined movements by their left or right hands, and one innovate technique was non-binary, which expedited the process.
This data showed that the techniques can be non-invasive (they used EEGs), and still produce high transfer results.
This lecture was in 2007, before all of the hoopla over the prosthetic arm movements by monkeys in Japan, but this field is (IMNSHO) about to explode. Get ready.