| #!/usr/bin/env python3 |
| """ |
| Evolutionary JEPA Masking Search (Evo-JEPA) |
| ============================================ |
| Uses CMA-ES to evolve optimal masking parameters for I-JEPA self-supervised pretraining. |
| Fitness is evaluated via kNN accuracy on CIFAR-100 after short pretraining runs. |
|
|
| Paper motivation: |
| - I-JEPA (arxiv:2301.08243) showed masking params swing accuracy by 45+ points |
| - FER paper (arxiv:2505.11581) showed evolution produces better representations than SGD |
| - This combines both: evolving the masking strategy that guides SGD-based JEPA training |
| """ |
|
|
