In recent years, deep learning techniques have been extensively applied to the analysis of electroencephalography (EEG) signals in various contexts, including motor imagery (MI) decoding. MI decoding has the potential for controlling devices or assisting in patient rehabilitation. A primary obstacle in implementing these algorithms in practical settings is the cost associated with model calibration, given their tendency to struggle with generalization to unseen subjects. To address this challenge, our work explores a new technique for representing EEG signals using embeddings. These embedding-based methods enable faster calibration without the need for model re-training. Experiments conducted on the Brain-Computer Interface Competition IV dataset 2a demonstrate the competitiveness of our approach compared to the state-of-the-art techniques in this field. It is also shown how embeddings offer additional opportunities to tackle EEG analysis challenges.