Full fine-tuning and larger CardioState-JEPA encoders

Investigate the performance of full fine-tuning and larger encoder sizes for CardioState-JEPA beyond the frozen-encoder linear-probing setting evaluated in the paper.

Background

CardioState-JEPA is evaluated exclusively as a frozen encoder with task-specific linear probes, rather than through end-to-end adaptation to downstream tasks. The paper also uses a ViT-B shared encoder, and does not assess whether increasing the encoder size improves transfer across ECG, PPG, and PCG tasks. These unresolved evaluation directions are identified as future work because they would test the model under broader adaptation and scaling regimes.