【Adaptation】 TTA to Federated and Semi-supervised learning

- TTA to Federated and Semi-supervised learning # Test time adaptation - Notable properties 1. Out-of-distribution Generalization with target images in testing 2. **Self-training** anyway 3. Online update (optional) 4. Small parts of parameters (optional) 5. Fast optimization (empirical finding) 6. If update BN, the performance impact is due to \[norm para > de-norm para\] (empirical finding) - Problems - Self-training for multi epochs → catastrophic failure → So then, Overfitting detector? - Learning what the model is already good at. → Even if any complex loss (ex, Consistency Training), no significant performance improvement by optimized de-norm. → Need to self-train what the model is bad at. - Local optimization - Small label efficiency compared to the large risk from on-device self-training. - Problems with my work 1. Domain division - Dividing into domains by weather and time was not proper since they have a small domain shift. (proof: [WILDS paper](https://arxiv.org/abs/2012.07421)) 2. Small margin performance improvement # Federated learning # Semi-supervised learning