Question/Issue:
My datum is a 6-channels time serie. I want to train separately a SVM on the 2-dim. embedded space that is shown in the “feature explorer”, to then add the trained model as a custom block. It seems that UMAP works with single-channel data only, thus my question is how EdgeImpulse processes the data to make them feasible to be fed to UMAP. Thanks
Project ID:
139815
Context/Use case:
Hello @jwing,
So we used to use UMAP, I think now we are using t-SNE or PCA as the dimensionality reduction.
Depending on where you see the feature explorer in your project, we use either embeddings of a pre-trained model, embeddings of your NN or the processed features of your DSP block.
Please see Data explorer - Edge Impulse Documentation for more info.
Best,
Louis