Brain signal classification remains one of the most exciting and complex challenges in human-machine interaction. Electroencephalography (EEG)-based brain-computer interface (BCI) systems make it possible to translate neural activity into digital commands, opening doors in rehabilitation, device control, and augmentative communication. However, the noisy, non-stationary and highly variable nature of EEG signals demands models that are both accurate and extremely efficient in terms of parameters and computation. Traditional approaches such as linear discriminant analysis, support vector machines, or multilayer neural networks often require a large number of weights, increasing the risk of overfitting and the need for large volumes of labeled data.
In this context, an innovative proposal emerges: Variational Fasorial Circuits (VPCs). Inspired by variational quantum circuits, VPCs operate exclusively on the manifold of the unit circle S¹, using trainable phase shifts and unit mixing operations instead of matrices of dense weights. This purely classical and deterministic approach offers an alternative representation of information: each feature is encoded as a phase in the complex plane, and the interference between these phases—constructive or destructive—generates a linear decision function in a cosine/sine lifting space. That is, the model learns to align and cancel phase vectors to separate classes, without resorting to the traditional nonlinearity of artificial neurons.
The results obtained with real EEG signals of motor imagination —from the PhysioNet Motor Movement/Imagery base— are especially revealing: an average accuracy of 0.60, the highest among conventional classifiers (LDA, logistic regression, SVM with RBF kernel and multilayer perceptron), but using an order of magnitude fewer parameters and showing the least variance between subjects. This suggests that the phase structure naturally captures power information in frequency bands – a predominant feature in EEG – while avoiding unnecessary complexity. However, the authors honestly acknowledge a fundamental limitation: VPCs cannot represent parity functions, a ceiling that depth cannot overcome, which makes them ideal for problems with near-linear separability in phase space, but not for tasks that require XOR logic or other non-hyperplane-separable patterns in that representation.
For the business and technological development field, this architecture has a great attraction: low resource consumption, ease of deployment in edge hardware and high interpretability. A company like Q2BSTUDIO, which specialises in the development of bespoke applications and bespoke software, can integrate this type of classifier into portable BCI systems or neurophysiological monitoring solutions. In addition, the ability to run inference with few parameters makes these models perfect for environments with compute or power limitations, such as wearables. The combination of artificial intelligence with biomedical signal processing techniques is a rapidly expanding field, and Q2BSTUDIO offers services ranging from consulting to implementation, including AWS and Azure cloud services to scale training and inference, as well as cybersecurity to protect sensitive patient data.
Another relevant aspect is the possibility of hybridizing VPCs with real or simulated quantum modules, creating classic front-end systems that power quantum processors. In this sense, AI agents can orchestrate the workflow: from EEG signal acquisition and filtering to VPC classification and subsequent decision-making. Q2BSTUDIO also develops business intelligence and Power BI service solutions that allow real-time visualization of brain activity and the evolution of models, facilitating decision-making in clinical or research environments. All of this is naturally integrated under the umbrella of AI for enterprises, where computational efficiency and personalization are key.
In conclusion, the Variational Fasorial Circuits represent a step forward towards lighter, more interpretable and robust classifiers for physiological signals. Their success in the motor imagination task proves that sometimes less is more: with few parameters and a well-defined geometric structure, they can outperform traditional dense models. For organizations looking to implement real BCI solutions, having a technology partner like Q2BSTUDIO – which offers enterprise AI and custom software development – allows you to explore these cutting-edge architectures and adapt them to specific contexts, whether in neurological rehabilitation, prosthetic control, or human-machine interfaces of the future.
