Using adversarial networks to extend brain computer interface decoding accuracy over time
Xuan Ma1, Fabio Rizzoglio1, Kevin L Bodkin1
1Department of Neuroscience, Northwestern University, Chicago, United States.
Elife
|August 23, 2023
Summary
Cycle-Consistent Adversarial Networks (Cycle-GAN) offer a robust solution for stabilizing brain-computer interfaces (BCIs). This method improves decoder accuracy over time by aligning neural data distributions, reducing the need for frequent recalibration.
Area of Science:
- Neuroscience
- Biomedical Engineering
- Machine Learning
Background:
- Intracortical brain-computer interfaces (iBCIs) restore movement for individuals with paralysis by translating neural activity into control signals.
- Decoder accuracy in iBCIs degrades over time due to changes in recorded neurons, necessitating recalibration.
- Recalibration is time-consuming and requires user effort to relearn new neural dynamics.
Purpose of the Study:
- To develop and evaluate unsupervised methods for stabilizing iBCI decoders without requiring frequent recalibration.
- To address the challenge of shifting neural representations by aligning coordinate systems of neural activity.
- To compare the efficacy of Cycle-GAN against existing methods like ADAN and Procrustes alignment.
Main Methods:
- Proposed a novel method using Cycle-Consistent Adversarial Networks (Cycle-GAN) to align full-dimensional neural recording distributions.
- Compared Cycle-GAN with a previously proposed Generalized Adversarial Network (GAN) method, Adversarial Domain Adaptation Network (ADAN), and Factor Analysis-based Procrustes alignment.
- Evaluated methods on data from multiple monkeys and diverse behaviors, focusing on unsupervised learning with minimal data requirements.
Main Results:
- Cycle-GAN demonstrated superior performance compared to ADAN and Procrustes alignment in stabilizing iBCI decoders.
- Cycle-GAN proved easier to train and more robust than ADAN.
- All tested methods were unsupervised and required minimal data, indicating practical applicability.
Conclusions:
- Cycle-GAN is an effective and practical method for stabilizing iBCI systems by mitigating decoder drift caused by neural turnover.
- The findings suggest Cycle-GAN can significantly enhance the long-term usability and reliability of brain-computer interfaces.
- This approach reduces the burden on users by minimizing the need for recurrent recalibration sessions.
