Hub-LoRA
Hub-LoRA improved hub sensitivity but its accuracy gains were not statistically significant. Hub-LoRA improved Brain-JEPA's biomarker metrics while maintaining other metrics. Hub-LoRA applies low-rank adaptation to the first layer so adapt…
1 sources - 5 claims
Hub-LoRA improved hub sensitivity but its accuracy gains were not statistically significant. Hub-LoRA improved Brain-JEPA's biomarker metrics while maintaining other metrics. Hub-LoRA applies low-rank adaptation to the first layer so adaptation directly interacts with input features. Hub-LoRA constrains adaptation toward hub nodes by freezing a one-hot matrix initialized from nodes whose ambivert degree exceeds 1.0. Hub-LoRA was proposed because standard foundation-model fine-tuning did not sufficiently capture hub structures.