From what I've seen (not much), a large obstacle in the field of brain/computer control seems to be signal interpretation. All our methods for analyzing electrical signals must be non invasive, a requirement which greatly increases noise. For example, to read EEG data from the brain we need to strap a net of ~128 electrodes around the skull and use FFT techniques to analyze 128 wave segments. The electrical signal bounces around in the convex skull, generating noise that pollutes the EEG signal. The fact is, it's hard to read electrical signals without a direct connection -- there is only so much we can do from outside the body.

(This is why consumer grade EEG devices can do little more than react in a binary fashion to relaxed/excited emotional state. The signal difference between relaxed/excited is easily detectable in aggregate, but more nuanced signals are hidden amongst the noise of EEG waves bouncing around the skull.)

For this reason, any non invasive method of reading (or in this case, apparently even controlling) electrical signals will advance the field of neuroscience exponentially.

This obstacle of noisy data leads to a second obstacle: what is the signal? Can we duplicate an arbitrary signal and send it through the nervous system on demand? That's a hard question to answer, especially when we can barely listen to the signal in the first place.

My question is, what role can machine learning play in this field? The problem of identifying signal seems more of a computational one than a biological one. How much data loss do EEG machines suffer from noise? Is all the data intact, and we just need to figure a way to avoid the noise?

Has there been any research on feeding neural signals into ML algorithms, and labeling them with results, e.g. "Moved finger," "blinked eyes," etc?

This field strikes me as one with a lot of untapped potential. With the coming VR revolution, thought-based interaction seems a natural next step.