Register now After registration you will be able to apply for this opportunity online.
Explainable Transformer Pipelines for Imagined-Speech and Motor-Imagery EEG BCIs
Noisy signals, scarce labels, and black-box models hinder EEG-based BCIs for imagined speech and limb movement. We will tackle these issues with hybrid convolution-transformer networks—CTNet (github.com/snailpt/CTNet) and MSCFormer (github.com/snailpt/MSCFormer)—augmented by transfer learning and few-shot adaptation. Attention heat-maps and SHAP explanations (github.com/slundberg/shap) will expose which channels and time windows drive each decision. The work aligns with ViTFOX’s goal of low-power, explainable neuro-AI (vitfox.eu).
We will train the models on the multi-session imagined-speech / grasp-imagery dataset published in Frontiers in Human Neuroscience (10.3389/fnhum.2022.898300).
1. Architectures: CTNet couples an EEGNet-style CNN front-end with a transformer encoder; MSCFormer adds parallel multi-scale convolutions before the attention block.
- Transfer & Few-Shot: Models will be pre-trained on larger EEG corpora, then fine-tuned; meta-learning will enable < 10-trial personalisation for new users.
- Explainability: Transformer attention, gradient saliency, and SHAP will highlight task-relevant electrodes and time segments.
- Efficiency: Pruning and token-length reduction will keep inference light for future wearable deployment on ViTFOX hardware.
We will train the models on the multi-session imagined-speech / grasp-imagery dataset published in Frontiers in Human Neuroscience (10.3389/fnhum.2022.898300).
1. Architectures: CTNet couples an EEGNet-style CNN front-end with a transformer encoder; MSCFormer adds parallel multi-scale convolutions before the attention block.
- Transfer & Few-Shot: Models will be pre-trained on larger EEG corpora, then fine-tuned; meta-learning will enable < 10-trial personalisation for new users.
- Explainability: Transformer attention, gradient saliency, and SHAP will highlight task-relevant electrodes and time segments.
- Efficiency: Pruning and token-length reduction will keep inference light for future wearable deployment on ViTFOX hardware.
- Generalisation – maintain high cross-subject accuracy and enable < 10-trial few-shot adaptation.
- Explainability – provide attention maps and SHAP reports that reveal physiologically plausible EEG features.
- Community Impact – release code, pretrained weights, and low-power deployment guidelines to accelerate ViTFOX-compatible BCIs.
- Generalisation – maintain high cross-subject accuracy and enable < 10-trial few-shot adaptation.
- Explainability – provide attention maps and SHAP reports that reveal physiologically plausible EEG features.
- Community Impact – release code, pretrained weights, and low-power deployment guidelines to accelerate ViTFOX-compatible BCIs.
Dr. Nikhil Garg (nigarg@ethz.ch) (H69, ETZ). Available as bachelor/Semester Thesis (milestones and project goals will be adjusted accordingly). Supervisor: Prof. Dr. Laura Bégon-Lours
Dr. Nikhil Garg (nigarg@ethz.ch) (H69, ETZ). Available as bachelor/Semester Thesis (milestones and project goals will be adjusted accordingly). Supervisor: Prof. Dr. Laura Bégon-Lours
