| --- |
| thumbnail: "https://github.com/joel99/ndt3/blob/main/static/ndt3_thumbnail_square.png" |
| tags: |
| - neuroscience |
| - bci |
| license: "cc-by-nc-4.0" |
| metrics: |
| - Decoding $R^2$ |
| --- |
| # Neural Data Transformer 3 (NDT3) |
|
|
| NDT3 is a multimodal Transformer modeling neural population spiking activity from motor cortex and low-dimensional behavioral covariates. It is trained on up to 2000 hours of human and monkey neuroscience data. |
|
|
| We upload 45M and 350M NDT3 checkpoints for each of 200h and 2kh pretraining. These are pytorch-lightning checkpoints with model configurations wrapped into the checkpoint files. The accompanying codebase and full model card is available at: [joel99/ndt3](https://github.com/joel99/ndt3). |
|
|
| The configs for these models are the corresponding files on `v5` [experiment branch](https://github.com/joel99/ndt3/tree/main/context_general_bci/config/exp/v5). The Weights and Biases report with pretraining curves are available [here](https://api.wandb.ai/links/joelye9/m8waz8e7). |
|
|
| - 350m 200h: jg3skdsx |
| - 350m 2kh: 900t21lf |
| - 45m 2kh: l2ulo75e |
| - 45m 200h: 753jmg4u |
