Source-linked AI summary
BrainNet: A Multi-Person Brain-to-Brain Interface for Direct Collaboration Between Brains
Linxing Preston Jiang, Andrea Stocco, Darby M. Losey, Justin A. Abernethy, Chantel S. Prat, Rajesh P. N. Rao
TL;DR
BrainNet addresses the limited interactivity, two-person scale, and physical-output requirements of earlier human brain-to-brain interfaces. It combines EEG and TMS in a three-person collaborative Tetris-like task, where Senders advise a Receiver and can provide feedback. Across five triads, average accuracy was 0.8125, and Receivers learned which Sender was more reliable when one signal was noisier.
Problem
Previous human brain-to-brain interfaces had minimal interactivity, sometimes required physical actions, and supported only two subjects.
Method
BrainNet uses EEG to decode two Senders’ decisions, TMS to deliver them to a Receiver, and a second interaction round for feedback during a collaborative Tetris-like task.
Results
Five triads achieved a mean accuracy of 0.8125, and Receivers learned which Sender was more reliable when noise was injected into one Sender’s signal.
Takeaways & Limitations
BrainNet demonstrates multi-person non-invasive brain-to-brain interaction for collaborative problem solving and supports differential trust in connected Senders.
Takeaways & Limitations
BrainNet transmits only binary information, and this low bit rate requires disproportionate technical hardware and setup.
Abstract
from arXiv · showhide
We present BrainNet which, to our knowledge, is the first multi-person non-invasive direct brain-to-brain interface for collaborative problem solving. The interface combines electroencephalography (EEG) to record brain signals and transcranial magnetic stimulation (TMS) to deliver information noninvasively to the brain. The interface allows three human subjects to collaborate and solve a task using direct brain-to-brain communication. Two of the three subjects are "Senders" whose brain signals are decoded using real-time EEG data analysis to extract decisions about whether to rotate a block in a Tetris-like game before it is dropped to fill a line. The Senders' decisions are transmitted via the Internet to the brain of a third subject, the "Receiver," who cannot see the game screen. The decisions are delivered to the Receiver's brain via magnetic stimulation of the occipital cortex. The Receiver integrates the information received and makes a decision using an EEG interface about either turning the block or keeping it in the same position. A second round of the game gives the Senders one more chance to validate and provide feedback to the Receiver's action. We evaluated the performance of BrainNet in terms of (1) Group-level performance during the game; (2) True/False positive rates of subjects' decisions; (3) Mutual information between subjects. Five groups of three subjects successfully used BrainNet to perform the Tetris task, with an average accuracy of 0.813. Furthermore, by varying the information reliability of the Senders by artificially injecting noise into one Sender's signal, we found that Receivers are able to learn which Sender is more reliable based solely on the information transmitted to their brains. Our results raise the possibility of future brain-to-brain interfaces that enable cooperative problem solving by humans using a "social network" of connected brains.
Introduction
BrainNet addresses key limitations of earlier human brain-to-brain interfaces by enabling multi-person, interactive, fully brain-mediated collaboration and allowing Receivers to learn which Senders are more reliable.
- Motivation: Earlier human BBIs supported minimal interactivity, often relied on physical actions, and were limited to two subjects.In the 20 Questions BBI, the Receiver’s performance did not affect the Sender’s decision, and the Receiver selected questions by touching a screen.
- Contribution: BrainNet connects two Senders with one Receiver, who integrates their independent decisions while the system is designed to scale to more Senders.The Senders observe the task and convey decisions; the Receiver integrates those decisions and selects an action.
- Contribution: A second interaction round lets Senders perceive the Receiver’s first action and provide potentially corrective decisions.
- Contribution: BrainNet uses EEG for recording and TMS for stimulation, eliminating the need for physical movements to convey information.
- Experiment: Receivers can learn to trust the more reliable Sender when noise is injected into one Sender’s signal.The manipulation varies one Sender’s signal reliability, and the Receiver learns which signal is unaffected.
Results
Across five triads, BrainNet supported above-chance Tetris-like task performance, reliable EEG decision decoding, and information transfer between Senders and the Receiver. Receivers also learned to distinguish more reliable Senders during repeated interactions.
- Overall Performance: 0.8125 mean accuracy corresponded to 13 correct block-rotation trials out of 16 across the five triads, exceeding chance (p = .002).Overall performance was measured as the proportion of correct block rotations, lines cleared, or maximum score achieved.
- SSVEP Performance: SSVEP decoding separated correct from incorrect "Rotate" and "Do Not Rotate" signals during the task.The corresponding power differences were significant for both signals: t(15) = 9.709 and t(15) = 10.725, respectively, with p < 0.0001.
- ROC Performance: 0.83 mean AUC across triads exceeded chance, with overall AUC lower than good-Sender AUC but significantly higher than bad-Sender AUC.The AUC framework distinguishes true-positive from false-positive decisions; chance is 0.5 and an ideal observer is 1.0.
- Mutual Information: MI was 0.336 for good Senders and 0.051 for bad Senders, and the difference was statistically significant.Both values exceeded chance MI, while good Senders transmitted more information on average.
- Mutual Information: The estimated mutual-information bias was negligible (b = −0.045) and did not affect any statistical tests.The estimate used NR = 2 possible responses and NS = 32 samples per participant pair.
- Learning Sender Reliability: Beta weights and correlation coefficients rose steeply across blocks for good Senders but not bad Senders, with trend differences significant for both measures.The differences were Z = 5.87, p < 0.001 for beta weights and Z = 7.31, p < 0.001 for correlations, suggesting Receiver learning of Sender reliability.
Discussion
BrainNet demonstrates multi-person non-invasive brain-to-brain collaboration while extending prior interfaces through networked interaction and reliability learning. The authors identify binary information complexity, specialized hardware, deliberately injected unreliability, and a task-specific server as improvement areas.
- BrainNet is presented as the first successful demonstration of multi-person non-invasive direct brain-to-brain interaction for collaborative problem solving.
- The interface expands brain-to-brain collaboration to multiple human subjects, combines EEG recording with TMS stimulation, and enables Receivers to learn Sender reliability.
- Earlier collaborative BCIs pooled brain signals but lacked direct stimulation-based information transfer and involved tasks performed on different days.
- BrainNet transmits only binary information per communication iteration, requiring a disproportionate amount of technical hardware and setup.
- The reliability experiment deliberately introduced a bad Sender, leaving open whether Receivers can learn reliability under more natural unreliability sources.
- The current server supports BrainNet’s experimental task rather than general-purpose routing, motivating a cloud-based BBI server for broader operation.
Methods
BrainNet connected two EEG-based Senders and one TMS/EEG-based Receiver across repeated Tetris-like decisions. Sender signals were transmitted as occipital TMS pulses, and the Receiver made and repeated decisions across two interaction rounds.
- Experimental setup: Fifteen healthy participants formed triads that played a simplified Tetris-like game in controlled laboratory sessions.
- Experimental task: Each session contained 16 randomized trials, with eight requiring rotation and eight requiring the block to remain unchanged.
- Experimental task: Each trial used two interaction rounds, allowing the Receiver to act after initial Sender input and then receive another opportunity for validation or feedback.
- BrainNet architecture: Two Senders viewed the task and encoded rotate or do-not-rotate decisions through EEG-based SSVEP cursor control.
- BrainNet architecture: Sender decisions traveled through TCP/IP and became sequential occipital TMS pulses for the Receiver, with pulse intensity encoding yes or no.
- Receiver and reliability manipulation: The Receiver used phosphene perception and the same SSVEP procedure to choose the block action, while one Sender was deliberately made less accurate in 10 of 16 trials.
- Signal processing: EEG preprocessing filtered incoming data from 0 to 30 Hz, divided it into 1-second epochs, and compared 17-Hz and 15-Hz power.
- Receiver and reliability manipulation: Receiver sessions included TMS safety screening and stimulation calibration, with TMS delivered over the left occipital lobe.
Author Contributions Statement
The authors divided responsibility across experiment conception, data collection, software implementation, analysis, and manuscript preparation.
- RPNR, AS, and CP conceived the experiment; LJ and JA conducted it; LJ and DL implemented the software; LJ and AS analyzed the results.
- All authors participated in writing and reviewing the manuscript.
Corresponding Author
The figures depict BrainNet’s multi-person architecture, iterative task interaction, and evaluation measures. Together, they show how EEG-based decisions reach a Receiver through TMS and how performance, signal discrimination, information transfer, and reliability learning are assessed.
- Architecture: BrainNet connects two EEG-based Senders to a TMS-stimulated Receiver, who integrates their inputs and sends an EEG-based action back into the task.The architecture includes internet transmission and feedback from the updated game state to the Senders.
- Iterative interaction: The two-round interaction lets the Receiver act first and then receive further Sender information after the game state updates.The Receiver’s second-round action can correct the first-round choice.
- Performance: Triad performance is plotted as correct block-rotation accuracy against a theoretical chance level of 0.5.The plot compares five participant triads.
- Signal decoding: SSVEP task spectra compare one-second EEG power during the task with power three seconds before and after it.During the task, the frequency corresponding to the correct answer has significantly greater power.
- Decision quality: ROC plots compare overall triad performance with Good and Bad Sender performance using AUC, with the dashed line marking chance.Each dot is associated with a triad, and shaded regions represent ROC-curve AUC.
- Information transfer and learning: Mutual information is significantly higher between the Receiver and Good Sender than between the Receiver and Bad Sender, while reliability measures rise over time only for the Good Sender.The reliability-learning plots track regression weights and Pearson correlations across 4-trial blocks.