@amoeba: Regarding:
this. I mostly agree with Anony and Erich (Sorry!). I think you are an extremely well-versed dimensionality reduction practitioner, perfectly capable of working with really complex datasets and thus you are showing a mild (reverse) Dunning-Kruger effect. As a principal, using t-SNE outputs for clustering is disputable for me. Yes, it can be "successful" and yes, I have used t-SNE/DBSCAN but... I wouldn't do it light-heartedly again.