Usually there are clear advancements in ML methodologies in research especially, where I can say X method is essentially better than Y method for most datasets.

However, I recently accidentally stumbled upon SOMs, and they seem pretty cool, like an interpretable form of UMAP or tSNEs, with the ability to offer perhaps a little more topological consistency in a way that is trained in a non-error correction way. Seems like a neat idea!

But when I look around for research on this: I hardly ever see people talk about it, and there isn't too much in the way of python libraries. And I can't see some "clear" reason as to why this is so. Is it truly outclassed by competing methods? If so, why is it used competitively within the flow cytometry community?

Just wondering if anyone here can shed light on this apparent "drop-off" wrt SOMs in regards to usage, or general research directions/applications. It's just not so clear for me.

  • 1
    $\begingroup$ I don't know much about it but was curious about the data. Google scholar returns 10k results for "self-organising map" and 1.8k since 2020: scholar.google.com/… . The corresponding numbers for tSNE are 22k and 13k, which seems consistent with your proposition. $\endgroup$
    – mkt
    Commented May 25, 2023 at 5:36
  • 3
    $\begingroup$ You focus on possible technical explanations (e.g. superiority of other methods), while there might be sociological explanations too (re: research funding, the choice of what is taught to students, etc., which may not always be based on purely technical considerations, and may impact the popularity of a method -a historical example of this kind of phenomenon in statistics is the Barnard's test). If it's the case here (I don't know if it is), it might be difficult to get an answer on this website, as it may require some sociological research in addition to specialized domain knowledge of SOM. $\endgroup$
    – J-J-J
    Commented May 27, 2023 at 6:28
  • 2
    $\begingroup$ @J-J-J That could be true! I am basically operating under the principle of "no harm in asking" :) Just in case someone maybe from the SOM community sees this post and could give an insight etc $\endgroup$ Commented May 27, 2023 at 9:02
  • 1
    $\begingroup$ They were never "great" and when DNNs came to the fore, auto-encoders blew them out of the water. "Theoretically" a SOM emphasizes topological mapping and clustering, practically an AE (or VAE) can do that even better. $\endgroup$
    – usεr11852
    Commented Jun 2, 2023 at 3:06
  • 1
    $\begingroup$ "practically an AE (or VAE) can do that even better." --- Can you prove that statement? It is not very obvious imo. Because you especially do not have a stronger guarantee of topological consistency with VAEs as compared to the set up of.a SOM.And if VAEs / AEs are objectively superior to SOMs, why is it that they are the "state of the art" in the field of cytometry --- where I am sure they have definitely heard of VAEs and AEs given how common they are. There is definitely some objective superiority to using SOMs in certain cases it appears $\endgroup$ Commented Jun 5, 2023 at 18:05


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.