Siamese-VAE Framework: Enhancing Adaptive BCI Models for Motor Imagery Classification

Opportunity

With the recent advances in deep learning and machine learning, the future of brain-computer interfaces (BCIs) has much potential in offering a plethora of opportunities across diverse real-world applications. These opportunities encompass transformative impacts on healthcare and medicine. For instance, BCIs can improve and supplement existing medical diagnosis and mental health monitoring in terms of brain health monitoring. Furthermore, we may be able to better predict and formulate preventive plans for high-risk individuals.

BCIs also hold promise in enhancing accessibility and inclusivity, enabling individuals with disabilities to control devices, communicate and navigate the world more independently. Moreover, they can aid neurorehabilitation for individuals who have survived traumatic brain experiences.

In consumer electronics and entertainment, BCIs can redefine user experiences, introducing a novel and an intuitive control of devices and immersive interactions in virtual reality. These interfaces will also have strong implications in research as well as neuroscience advancements in many other promising fields such as defence strategies and revolutionising the way users interact with digital machines.

Going forward, interdisciplinary collaboration across neuroscience, engineering, computer science, and related fields may be applied to harness the full potential of BCI technologies and ensuring their ethical and equitable integration into society, considering aspects such as data privacy and accessibility for all.

Technology

The work introduces a novel unsupervised adaptive framework, the Siamese-VAE, that offers significant improvements from the current state-of-the-art in motor imagery classification from electroencephalography signals. This is done via fine-tuning the model upon a selected subset of previously known data in a process dubbed as model re-learning.

The proposed framework considers both the stationary and non-stationary features of motor imagery including ill-defined features, achieving greater performance via selection of data with closest feature representations. This method can be applicable beyond the chosen task in this study towards other similar tasks, and customised to suit other use cases such as improving classifiers of muscle or heart signals, thus broadening the potential applications of the proposed adaptive framework.

The advantage of the Siamese-VAE is that it enables unsupervised adaptive learning framework that leverages on already known data to greatly increase the quality and amount of data that is used during the model adaptation process. This method achieves greater adaptation improvements across all subjects without the need for additional target data. This allows for the adaptation framework to be used for long-term unsupervised system re-calibrations which was previously not possible.

Figure 1: Siamese-VAE unsupervised few-shot adaptive framework. The similarity between latent distributions of the target and known data are estimated through comparing VAE model losses to measure the distance between learned target features to the input data features. Known data samples which have low contrastive distance values are utilised in the fine-tuning stage for model adaptation. 

Applications & Advantages

• Rehabilitation for stroke patients or individuals with motor impairments.
• Screening for highlighting individual risk levels of developing stroke.
• Building of neuroprostheses with direct control between brain and end effector.
• Communication for users via selecting words on a screen through mental movements.