Submitted by **olmec-akeru** t3_z6p4yv
in **MachineLearning**

…and why?

So the science has moved on quite considerably since the linear methods of PCA and others; about 5±1 years back we had t-SNE and later on VAEs then UMAP. I appreciate that each of these methods is taking a subtly different (ok ok ok, sometimes its not that subtle) view of the problem, but I wonder what approaches are SOTA now?

Where to now?

Deep-Station-1746t1_iy2k52e wroteIt fully depends on assumptions. Assumptions about the data, and the model. Without assumptions, you can't do dim reduction.

So, what are you willing to assume? e.g. Assuming you have vast quantities of text data, arguably the current best dim reducers are generative transformers.