@BenjaminHan and rephrasing my earlier statement:

Using #knowledge as engineering short-hand for #inductivebias is fine. Use #KG as a type of implementation of #datastorage for #DNN, but not as #representation of #knowledge.

Another view point from which to consider the obvious deficit of #KG is the work of #Vervaeke et al, esp. the interpretation of “4 types of knowledge”.

https://www.psychologytoday.com/us/blog/theory-knowledge/202101/john-vervaeke-s-brilliant-4p3r-metatheory-cognition

@BenjaminHan also, coming back to ur OP ;-)

In practice, we find #KG useful to loosely couple databases across different “subcontexts”, bridging (not breaking) #organizational #silos. Yet, the shared context remains narrow and brittle.

#Transformers we find most useful to catch the longtail of at least a subcontext in a practically robust way.

And yes, #inductivebias here is informed by #KG, top-down, yet as always #datamodel of input is the determinant of predictive viability.