Abstract—Decoding motor behavior from electrocorticography (ECoG) signals is difficult due to strong variability across recording sessions and the limited availability of labeled data. The goal of this semester project was to improve a supervised baseline model by exploring self-supervised learning (SSL) approaches and alternative training losses. The baseline model is a transformer-based architecture that takes wavelet-transformed ECoG signals as input and predicts the wrist position of a monkey. TS-TCC is applied as a self-supervised pretraining method, using contextual and temporal contrastive losses without relying on labels. After pretraining, only the encoder is retained and used for downstream regression tasks. The results show that the self-supervised pretrained model generalizes better to future recording sessions compared to the fully supervised baseline, achieving higher mean R^2 scores across recordings. Label fraction experiments further demonstrate that SSL allows the model to reach reasonable performance with as little as 1% of labeled data. Additional losses based on soft temporal and instance-level contrastive learning were also evaluated. While these losses showed promising behavior on smaller datasets, they tended to degrade performance when applied to large-scale pretraining, often smoothing predictions and reducing performance on well-performing sessions. Overall, this project highlights the potential of self-supervised learning to improve generalization and label efficiency for ECoG-based motor decoding.
👉 Read the full project's report here.
Project by Hugo Demule supervised by Yuhan Xie and Prof. Shoaran from the Integrated Neurotechnologies Laboratory, Campus Biotech, Geneva, Switzerland