Young Ho Kang
Abed Khorasani
Robert D. Flint
Behraz Farrokhi
Sang Wan Lee- 1Department of Brain Cognitive Science, Korea Advanced Institute of Science and Technology (KAIST), Daejeon, Republic of Korea
- 2Kerman Neuroscience Research Center, Kerman Medical University, Kerman, Iran
- 3Department of Neurology, Northwestern University, Evanston, IL, United States
- 4Department of Neurology, University of California, San Francisco, San Francisco, CA, United States
Editorial on the Research Topic
Neural computations for brain machine interface applications
Along with a flowering of deep learning, there has been a renaissance of brain-computer interface (BCI) research. One of the most active research areas of the BCI aims attention at a wide range of clinical endeavors, such as studies on prosthetic limb muscle control (Vilela and Hochberg, 2020), neurorehabilitation (Bamdad et al., 2015) or epilepsy (Vidyaratne and Iftekharuddin, 2017; Alkawadri, 2019). A typical clinical application of the BCI aids an individual with central nervous system (CNS) injuries or disabilities, such that it compensates for associated dysfunctions resulting from the impairment. Beyond the scope of clinical purposes, there has also been various research on controlling an external system solely via neural signals (Trejo et al., 2006; Khaliliardali et al., 2015). In this editorial, with insights from carefully selected studies, we explore recent progress in the computational aspects of BCIs. Further, we discuss how such studies can make convergent contributions to benefitting the human-machine alignment.
All five studies on this topic astonishingly broadened our knowledge of computational tools for BCIs. Three of them, in particular, shed light on employing deep learning techniques in various BCIs. Cui et al. contemplated interpretability issues of the BCIs relying upon deep learning. Interpretability is a desirable trait for BCIs; as an interpretable BCI may provide fruitful insight into cognitive and/or neural mechanisms related to kinematics, learning and memory in particular. Deep learning, whilst often boosting decoding performance of the BCI, can be difficult to interpret. This study thus evaluated various interpretation schemes for deep BCI and demonstrated the best practice of utilizing such techniques. Also, Sun et al. introduced the EEG-completion-informer (EC-informer). The authors investigated how the number of EEG acquisition channels can be finely reduced and successfully demonstrated how virtually generated channels can compensate for the information gap. Hence, this approach eliminates extensive computing costs while preserving key factors with minimal loss—exhibiting robust applicability. Yang et al. presented a novel patient-specific approach to predict epileptic seizures based on multimodal neural data and an adversarial model. The domain-adversarial training with the multimodal data enables the model to extract invariant features of individuals and improves the model's stability, which can be otherwise reduced by subject variability. Thereby, this work successfully demonstrated how the BCI can be generalized across the subjects particularly in the clinical domain.
Meng et al. rigorously examined the effect of various gaze fixation positions and covert attention on BCI performance. The work revealed that subjects' performance on the given BCI task was not affected by the position of gaze fixation and covert attention, suggesting, at least for the motor imagery BCIs, the system can be gaze-independent. As a result, the authors provided precise instructions on the extent to which the end-users of BCIs can behave freely. A precise user guide relaxes some burdens on users, granting them a better BCI experience. Unlike the first four non-invasive BCI studies, Wan et al. demonstrated the effects of non-stationary neural signals on an intracortical-BCI (iBCI) decoding performance. Considering recording degradation and neuronal property variation, the study served as a reference for a model and its training scheme in chronic iBCI. The study gave a decent guideline for developing the BCI system with non-stationary neural signals as its input, which is commonly encountered.
Despite a large volume of contributions to advancement of the BCI decoders, in this topic improving interpretability (Cui et al.), generalizability (Yang et al.) and robustness (Sun et al.; Meng et al.; Wan et al.) of the decoder, most BCIs potentially suffer from the human-machine value alignment issues, namely sharing human values with the machine. The alignment theory views both a human and a machine as learning systems, such that two learners actively learn to accomplish a given task through a sequence of interactions within a closed loop. One study (Müller et al., 2017) established a mathematical model for the coadaptation to formulate and solve an asymmetric nature of a communication paradigm of the BCI. They suggested the error rate of human-to-machine communication can minimized by an interactive joint adaptation process via noise-free machine-to-human feedback. Not to mention the significance of the machine's decoding performance to which a large volume of studies have devoted (Craik et al., 2019; Glaser et al., 2020), how crucial it is for users to adequately explore and exploit the BCI has been constantly emphasized as the key factor of a successful brain-actuated system (Perdikis and Millán, 2020).
As a potential solution for the human-machine value alignment problem, deploying a cognitive model in the loop can minimize the burden on two learners. A cognitive model can be defined as a computational model of cognitive processes of interest built based on neural and/or behavioral data. The model's variables reflect various aspects of the cognitive function under investigation from which an inference on an individual's cognitive status can be made, at least within the given context. One good example is the temporal difference (TD) model, initially inspired by animal reward-seeking behaviors (Sutton and Barto, 1987; Sutton, 1988). There has been an attempt to leverage critical variables of the computational model of human reinforcement learning for predicting user intention, showing that the decoder's performance with the given cognitive model can be significantly improved (Kim and Lee, 2018). This model can also be supplemented with a model of human intuition on physics, called an intuitive physics engine (IPE). The IPE makes physical inferences based on human-like physical understanding (Smith et al., 2019). In the context of the BCI system serving a user as a physical aid, for instance, the IPE would generate several possible human-like inferences based on simulations of committing human-like actions. Further, decoded signals and the inferences can be combined to align the machine with the human's demanding objective more precisely. By harnessing the cognitive model, machines in BCIs would become more human-like. As a result, we anticipate the cognitive models to aid machines in understanding intentions with an exponential increase in precision, thereby easing the learning burden of both the machine and the user in the loop.
Author contributions
YK: Writing—original draft, Writing—review & editing. AK: Writing—original draft, Writing—review & editing. RF: Writing—review & editing. BF: Writing—review & editing. SL: Writing—review & editing.
Funding
The author(s) declare that no financial support was received for the research, authorship, and/or publication of this article.
Conflict of interest
The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.
Publisher's note
All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
References
Alkawadri, R. (2019). Brain–computer interface (BCI) applications in mapping of epileptic brain networks based on intracranial-EEG: an update. Front. Neurosci. 13, 191. doi: 10.3389/fnins.2019.00191
Bamdad, M., Zarshenas, H., and Auais, M. A. (2015). Application of BCI systems in neurorehabilitation: a scoping review. Disabil. Rehabil. Asset. Technol. 10, 355–364. doi: 10.3109/17483107.2014.961569
Craik, A., He, Y., and Contreras-Vidal, J. L. (2019). Deep learning for electroencephalogram (EEG) classification tasks: a review. J. Nueral Eng. 16, 031001. doi: 10.1088/1741-2552/ab0ab5
Glaser, J. I., Benjamin, A. S., Chowdhury, R. H., Perich, M. G., Miller, L. E., and Kording, K. P. (2020). Machine learning for neural decoding. eNeuro. 7, 1–16. doi: 10.1523/ENEURO.0506-19.2020
Khaliliardali, Z., Chavarriaga, R., Gheorghe, L. A., and Millán, J. R. (2015). Action prediction based on anticipatory brain potentials during simulated driving. J. Neural Eng. 12, 066006. doi: 10.1088/1741-2560/12/6/066006
Kim, D., and Lee, S. W. (2018). Model-based BCI: a novel brain-computer interface framework for reading out learning strategies underlying choices. Int. Conf. Proc. IEEE Int. Conf. Syst. Man Cybern (Miyazaki: IEEE). doi: 10.1109/SMC.2018.00091
Müller, J. S., Viadaurre, C., Schreuder, M., Meinecke, F. C., von Bünau, P., and Müller, K. R. (2017). A mathematical model for the two-learners problem. J. Neural Eng. 14, 036005. doi: 10.1088/1741-2552/aa620b
Perdikis, S., and Millán, J. R. (2020). A tale of two learners. IEEE Trans. Syst. Man Cybern. 6, 12–19. doi: 10.1109/MSMC.2019.2958200
Smith, K., Mei, L., Yao, S., Wu, J., Spelke, E., Tenenbaum, J., et al. (2019). “Modeling Expectation Violation in Intuitive Physics with Coarse Probabilistic Object Representations,” in Advances in Neural Information Processing Systems, eds H. Wallach, H. Larochelle, A. Beygelzimer, F. d'Alché-Buc, E. Fox, and R. Garnett (Curran Associates, Inc.). Available online at: https://proceedings.neurips.cc/paper/2019/hash/e88f243bf341ded9b4ced444795c3f17-Abstract.html
Sutton, R. S. (1988). Learning to predict by the methods of temporal differences. Mach. Learn. 3, 9–44. doi: 10.1007/BF00115009
Sutton, R. S., and Barto, A. G. (1987). “A temporal-difference model of classical conditioning,” in Proc. 9th Ann. Conf. Cogn. Sci. Soc. (Hillsdale, NJ: Erlbaum), 355–378.
Trejo, L. J., Rosipal, R., and Matthews, B. (2006). Brain-computer interfaces for 1-D and 2-D cursor control: designs using volitional control of the EEG spectrum or steady-state visual evoked potentials. IEE Trans. Neural Syst. Rehabil. Eng. 14, 225–229. doi: 10.1109/TNSRE.2006.875578
Vidyaratne, L. S., and Iftekharuddin, K. M. (2017). Real-time epileptic seizure detection using EEG. IEEE Trans. Neural Syst. Rehabil. Eng. 25, 2146–2156. doi: 10.1109/TNSRE.2017.2697920
Keywords: human-AI alignment problem, brain-computer interfaces, BCI, BMI, restorative technologies
Citation: Kang YH, Khorasani A, Flint RD, Farrokhi B and Lee SW (2023) Editorial: Neural computations for brain machine interface applications. Front. Hum. Neurosci. 17:1334636. doi: 10.3389/fnhum.2023.1334636
Received: 07 November 2023; Accepted: 10 November 2023;
Published: 23 November 2023.
Edited and reviewed by: Gernot R. Müller-Putz, Graz University of Technology, Austria
Copyright © 2023 Kang, Khorasani, Flint, Farrokhi and Lee. This is an open-access article distributed under the terms of the Creative Commons Attribution License (CC BY). The use, distribution or reproduction in other forums is permitted, provided the original author(s) and the copyright owner(s) are credited and that the original publication in this journal is cited, in accordance with accepted academic practice. No use, distribution or reproduction is permitted which does not comply with these terms.
*Correspondence: Sang Wan Lee, c2FuZ3dhbiYjeDAwMDQwO2thaXN0LmFjLmty
†These authors share first authorship