burn-deblur
| Crates.io | burn-deblur |
| lib.rs | burn-deblur |
| version | 0.1.0 |
| created_at | 2026-01-23 00:19:52.415356+00 |
| updated_at | 2026-01-23 00:19:52.415356+00 |
| description | An optimizer for burn which is suitable for training on non-I.I.D data |
| homepage | |
| repository | https://codeberg.org/jeremylist/burn-deblur |
| max_upload_size | |
| id | 2063037 |
| size | 211,487 |
Jeremy List (quickdudley)
documentation
README
burn-deblur
An implementation of the technique described in Using Fast Weights to Deblur Old Memories by Hinton & Plaut as an optimizer for use with burn.
Depending on the model structure and hyperparameters used, users can expect slower initial convergence than more commonly used optimizers such as ADAM, but more robustness to non-I.I.D. training data (including reinforcement learning).