Sabaki has a pattern matching for stone correlation (a stone related to a neighboring stone), not actual “shapes” (like table shape, bamboo joint, bulky five, or even just empty triangle). And a stone correlation is easy to program, so no actual machine learning required (maybe in the area, where multiple definitions are matched, human has some preferences to perceive one over the others based on context which is still quite difficult and not necessarily solvable, just find the “hane” definition post and see people’s different perceptions regarding a terminology. It’s a hard job if a training target is subjective).
And even if we can get this low-level feature mostly right, most of the time, this information isn’t that important. Imagine a combination of a table shape and a bamboo join (2 stones and 3 stones with one space apart), in pattern matching it would fit both, but we don’t usually call them with either of them, and what they are called aren’t even the most important thing, the key concept (the perceptual idea) is about “connectivity”, they are all good connected shape (but not fully). For teaching purposes, especially for DDK, it was the later perceptual and subjective understanding that is important and needs to be “commented on” (like under what conditions and context this “connectivity” could be broken, even if they met the definition)