Symbiotic Brain-Machine Drawing via Visual Brain-Computer Interfaces

Gao Wang, Yingying Huang, Lars Muckli, Daniele FaccioarXiv 2025 · DOI · arXiv

Introduction

Traditional Brain-Computer Interfaces (BCIs) for image reconstruction often face a "Goldilocks" problem: they are either highly invasive, require expensive fMRI equipment, or rely on rigid deep learning models that struggle to generalize. These systems typically suffer from low Information Transfer Rates (ITR), capped at around 10 bits per second. This research introduces a low-cost, non-invasive BCI that allows users to "draw" with their minds using a single-channel EEG. By treating the computer as an active collaborator that adaptively places visual probes, the system achieves a significant increase in communication bandwidth.

Methods and Materials

The system utilizes a custom-built, single-channel EEG device targeting the Oz position (primary visual cortex). It operates on Steady-State Visual Evoked Potentials (SSVEPs), where brain activity synchronizes to the frequency of flickering lights.

The "Symbiotic" workflow involves an iterative loop:

  1. Visual Probes: 10 flickering discs are presented on a screen.
  2. Selection: The user focuses on the disc that overlaps most with their imagined shape.
  3. Decoding: Canonical Correlation Analysis (CCA) identifies the target frequency from the EEG signal.
  4. Iterative Update: A Policy Function (Gabor-inspired or Data-driven) decides the next probe placement based on the emerging sketch. The Gabor-inspired policy mimics human edge perception, while the data-driven policy uses Non-negative Matrix Factorization (NNMF) on datasets like MNIST.
  5. Generative Refinement: The final abstract sketch is processed via Stable Diffusion to generate high-fidelity images.

Results

The symbiotic interaction between the human user and the AI policy function resulted in a 5x increase in BCI bit-rates compared to non-adaptive methods. Measured bit-rates reached 1.31 bits/s, exceeding theoretical maximums for one-way decoding.

Key performance metrics include:

  • Efficiency: Simple shapes were reconstructed in under two minutes, typically requiring only 25 iterations.
  • Accuracy: The system achieved an average Cosine Similarity (COSS) of 0.76 across eight subjects.
  • Generalization: The Gabor-inspired policy successfully reconstructed novel geometric shapes without prior training, demonstrating a model-free capability.

Discussion and Conclusion

The core innovation lies in the "Symbiotic" approach, where the computer actively assists the user by predicting intent and optimizing probe placement. This offloads cognitive burden and bypasses traditional bandwidth limits. While currently limited to relatively simple shapes and requiring roughly two minutes per drawing, the system demonstrates a viable path toward portable, high-bandwidth neural creative tools. Future work will explore multi-channel EEG and more complex spatial content to move closer to a true "telepathic" creative tool.

Summary updated

BibTeX
@misc{wang2025symbiotic,
  author = {Gao Wang and Yingying Huang and Lars Muckli and Daniele Faccio},
  title = {{Symbiotic Brain-Machine Drawing via Visual Brain-Computer Interfaces}},
  year = {2025},
  eprint = {2511.20835},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.20835},
}
Figure