About UMAP Visualization on FACED Dataset

[susameddin]

Thank you for this great work, it is extremly exciting to see EEG Foundation models working well! I also appreciate how clean the code and documentation is.

I have a question about the UMAP visualizations in Appendix P.2. I’m trying to reproduce the FACED plots, but my “w/o fine-tuning” visualization isn’t as well clustered. Once I have my raw data and embedding tensors in shape (B, C, S, P), did you simply flatten them to (B, C×S×P) before running UMAP, or did you apply average-pooling over some dimensions first? And did you perform PCA (or any other preprocessing) before UMAP?

Thanks in advance for the help!