Learning from mistakes

Application of neural activity based error detection for improvement of a continuous BMI control (further info at end of page).

Scientists identify error signals and help to improve movement detection for brain-machine interfaces

Brain-machine interfaces (BMIs) translate the activity of nerve cells that accompany a user’s movement intention into action, for instance that of a robotic arm. In spite of continuous improvements, errors that are made while the computer tries to decode an intended movement still pose a major problem for current BMI systems. If the decoded movement veers off noticeably from the intended movements, scientists call this an “execution error”. Another type of error that is important to consider with BMIs is the so-called outcome error. In this case, a subject fails to reach a certain movement goal (for instance, securely grasping a cup). Operating BMIs would greatly benefit from detecting such errors: The computer could correct them after detecting that intention and execution do not match any more.  And furthermore, an adaptive BMI decoding algorithm could update itself and learn to avoid such errors in the future.

In a study published in the journal PLoS ONE, Tomislav Milekovic and colleagues from the Bernstein Center at the University of Freiburg and Imperial College London demonstrate that such errors can in fact be detected from human brain signals. As a source for the data in which they looked for error signals, they used a method called electrocorticography (ECoG). This procedure measured changes in electrical potential at the surface of the brain. Because electrodes are not implanted into human skulls merely for the purpose of a scientific study, the researchers depended on the help from patients who had such electrodes implanted for medical reasons. When the subjects carried out a continuous movement, the computer programme detected errors with high precision within less than half a second. Even when the researchers focused on the data output of only 4 electrodes covering a small square of brain surface, they were able to extract from these locations 82% of detection information for outcome error and 74% of detection information for execution error.

In future, the error detection method presented by the team from London and Freiburg could correct errors that occur during BMI operation, and furthermore to adapt a BMI algorithm to make fewer errors. The results also show that even very small implants are sufficient to detect these errors. As smaller implants mean a reduced medical risk during implantation, their findings may help mass-produced brain-machine interfaces to become a reality. In Freiburg, it is the newly founded Cluster of Excellence BrainLinks-BrainTools that pursues this goal.
 

Original publication (open access)

Milekovic T, Ball T, Schulze-Bonhage A, Aertsen A, Mehring C (2013) Detection of Error Related Neuronal Responses Recorded by Electrocorticography in Humans during Continuous Movements. PLoS ONE 8(2): e55235. doi:10.1371/journal.pone.0055235

 

Image legend

If a BMI decodes an intended movement correctly, no neuronal error signal is elicited. If a so-called execution error is detected, the algorithm can be adapted to reduce the number of errors in decoding in the future. If the unwanted movement causes the cursor to reach an unwanted target, an outcome error signal may be evoked. If this is detected by the BMI system, it can change the decoding algorithm as well, this time in a different way.