Teaching brain-machine interfaces as an alternative paradigm to neuroprosthetics control.

Iñaki Iturrate,Javier Minguez,Ricardo Chavarriaga,José Del R Millán,Luis Montesano

doi:10.1038/srep13893

Iñaki Iturrate, Javier Minguez + Show 3 more

Open Access

https://doi.org/10.1038/srep13893

Copy DOI

Abstract

Brain-machine interfaces (BMI) usually decode movement parameters from cortical activity to control neuroprostheses. This requires subjects to learn to modulate their brain activity to convey all necessary information, thus imposing natural limits on the complexity of tasks that can be performed. Here we demonstrate an alternative and complementary BMI paradigm that overcomes that limitation by decoding cognitive brain signals associated with monitoring processes relevant for achieving goals. In our approach the neuroprosthesis executes actions that the subject evaluates as erroneous or correct, and exploits the brain correlates of this assessment to learn suitable motor behaviours. Results show that, after a short user’s training period, this teaching BMI paradigm operated three different neuroprostheses and generalized across several targets. Our results further support that these error-related signals reflect a task-independent monitoring mechanism in the brain, making this teaching paradigm scalable. We anticipate this BMI approach to become a key component of any neuroprosthesis that mimics natural motor control as it enables continuous adaptation in the absence of explicit information about goals. Furthermore, our paradigm can seamlessly incorporate other cognitive signals and conventional neuroprosthetic approaches, invasive or non-invasive, to enlarge the range and complexity of tasks that can be accomplished.

Highlights

—i.e., the error-related potential (ErrP)[14,15], a time-locked potential elicited when actions do not match users’ expectations[16,17,18,19,20,21]
Classification performance (73.8%, 72.5%, 74.3% on average for Experiments 1 to 3 respectively) exceeded the chance level—a necessary condition for a reinforcement learning system to acquire an optimal control policy[25]. This decoding performance remained similar to the overall accuracy (FDR-corrected two-tailed independent t-test, p > 0 .05) during the whole experiment despite the fact that the neuroprotheses move randomly at the beginning of an experiment and the error rate decreases as the devices learn an optimal motor behaviour
The final number of actions per target for these experiments was 1.81 ± 0 .48 and 1.47 ± 1 .12. This confirms that the ErrP does not depend on targets, as the ErrP classifier maintains its performance without needing to be retrained for unseen targets. These experiments illustrate a number of appealing properties associated with the use of error-related brain signals to allow a brain-machine interfaces (BMI) to teach neuroprostheses suitable motor behaviours

Summary

Introduction

—i.e., the error-related potential (ErrP)[14,15], a time-locked potential elicited when actions do not match users’ expectations[16,17,18,19,20,21]. ErrPs are evoked by actions that the user considers wrong to achieve his/her desired goals, decoded online, and employed as a reward signal for a reinforcement learning algorithm (RL)[25] that improves the neuroprosthesis behaviour. We tested this approach in three closed-loop experiments of increasing real-life applicability involving twelve subjects (Fig. 2a). Subjects, but not the neuroprosthesis, knew the target location and monitored the performance of the device. For Experiments 2 and 3, there were two targets (practice targets) that were used during ErrP calibration and online operation; and two targets (new targets) that were only used during online operation

Methods

Results

Conclusion