EEG2Rep is a self-supervised approach for EEG representation learning that addresses three key challenges in EEG data: low signal-to-noise ratio, wide amplitude ranges, and lack of explicit segmentation. The method predicts masked inputs in latent representation space and uses a novel semantic subsequence preserving (SSP) method to generate informative masked inputs. EEG2Rep outperforms state-of-the-art methods on six diverse EEG tasks with subject variability. It is robust to noise and shows that preserving 50% of EEG recordings yields the most accurate results. The model's architecture includes input embedding, context-driven target prediction, and predictor networks. The SSP method ensures sufficient semantic information for reconstruction. EEG2Rep is trained to reconstruct abstract features in latent space, which reduces noise and improves semantic quality. The model's efficiency is enhanced through multiple masking and prediction strategies. The method is evaluated on multiple datasets, including Emotiv and TUH datasets, and shows superior performance in linear probing and fine-tuning tasks. EEG2Rep also demonstrates robustness to noise, with superior performance compared to other models. The results indicate that EEG2Rep is effective in enhancing EEG representations and improving classification accuracy.EEG2Rep is a self-supervised approach for EEG representation learning that addresses three key challenges in EEG data: low signal-to-noise ratio, wide amplitude ranges, and lack of explicit segmentation. The method predicts masked inputs in latent representation space and uses a novel semantic subsequence preserving (SSP) method to generate informative masked inputs. EEG2Rep outperforms state-of-the-art methods on six diverse EEG tasks with subject variability. It is robust to noise and shows that preserving 50% of EEG recordings yields the most accurate results. The model's architecture includes input embedding, context-driven target prediction, and predictor networks. The SSP method ensures sufficient semantic information for reconstruction. EEG2Rep is trained to reconstruct abstract features in latent space, which reduces noise and improves semantic quality. The model's efficiency is enhanced through multiple masking and prediction strategies. The method is evaluated on multiple datasets, including Emotiv and TUH datasets, and shows superior performance in linear probing and fine-tuning tasks. EEG2Rep also demonstrates robustness to noise, with superior performance compared to other models. The results indicate that EEG2Rep is effective in enhancing EEG representations and improving classification accuracy.