CBraMod: A Criss-Cross Brain Foundation Model for EEG Decoding
Authors: Jiquan Wang, Sha Zhao, Zhiling Luo, Yangxuan Zhou, Haiteng Jiang, Shijian Li, Tao Li, Gang Pan
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We evaluate CBra Mod on up to 10 downstream BCI tasks (12 public datasets). CBra Mod achieves the state-of-the-art performance across the wide range of tasks, proving its strong capability and generalizability. (Abstract) and Table 2: The results of different methods on emotion recognition. |
| Researcher Affiliation | Collaboration | 1State Key Laboratory of Brain-machine Intelligence, Zhejiang University 2College of Computer Science and Technology, Zhejiang University 3Alibaba Group |
| Pseudocode | No | The paper describes the methodology using prose and figures (Figure 2, 3) but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | The source code is publicly available at https: //github.com/wjq-learning/CBra Mod. |
| Open Datasets | Yes | CBra Mod is pre-trained on Temple University Hospital EEG corpus (TUEG) (Obeid & Picone, 2016). (Section 3.1) and Table 1: Overview of downstream BCI tasks and datasets. BCI Tasks Datasets Rate # Channels Duration # Samples Label I. Emotion Recognition FACED (Chen et al., 2023) 250Hz 32 10s 10,332 9-class SEED-V (Liu et al., 2021b) 1000Hz 62 1s 117,744 5-class... |
| Dataset Splits | Yes | We use subject 1 to 80 for training, 81 to 100 for validation, and 101 to 123 for test in FACED. (Section 3.3) and We divide the fifteen trials of each session into three equal parts (5:5:5) as the training, validation and test sets, respectively. (Section 3.3) and Subject 1 70, 71 89, 90 109 are used for training, validation and test, respectively. (Section 3.3) and We use subject 1 to 15 for training, subject 16 to 20 for validation, and 21 to 25 for test. (Section 3.3) and In our experiment, we set subject 1 to 80 as training set, subject 81 to 90 as validation set and subject 91 to 100 as test set. (Appendix E.1) and We use subject 1 to 19 for training, subject 20, 21 for validation, and subject 22,23 for test. (Appendix E.2) and 24 MDD patients and 19 NCs are used for training, 5 MDD patients and 4 NCs are used for validation, and 5 MDD patients and 5 NCs are used for test. (Appendix E.4) and In our experiment, we use subject 1 to 15 for training, subject 16 to 19 for validation and subject 20 to 23 for test. (Appendix E.5) and The original dataset provide the training and test splits. We further divide the training subjects into training and validation set by 80%:20%, consistent with BIOT. (Appendix E.7) and The original dataset provide the training and test splits. Be same as BIOT, we further divide the training subjects into training and validation set by 80%:20%. (Appendix E.8) and Subject 1-5, 6-7, 8-9 are used for training, validation, and test, respectively. (Appendix E.9) |
| Hardware Specification | Yes | CBra Mod was pre-trained on one machine with Intel Xeon Gold 6226R CPU and four NVIDIA RTX A5000 GPU for about 5 days. |
| Software Dependencies | Yes | We implemented CBra Mod based on the Python 3.11.7 and Py Torch 2.1.2 + CUDA 12.1. |
| Experiment Setup | Yes | We set the time duration of each EEG patch as 1 second (200 data points), and one 30-second EEG sample is segmented to 19 * 30 = 570 EEG patches. For the model configurations, the patch encoder consists of 3-layer 1D convolution with group normalization and GELU activation. The positional encoder is an one-layer 2D depthwise CNN. The backbone of CBra Mod is a 12-layer criss-cross transformer with 200 hidden dimensions, 800 inner dimensions (feed-forward), and 8-head criss-cross attention (4 heads for S-Attention and 4 heads for V-Attention). 50% of mask ratio is used to randomly mask the patches. The batch size was set to 128 and the number of epochs was set to 40. The model is trained using the Adam W optimizer with default settings, the learning rate is set to 5e-4 and the weight decay is set to 5e-2. Cosine Annealing LR was used to dynamically adjust the learning rate during the pre-training. |