I actually agree with this take! Deep learning models shine in problem domains where the gap between the data representation to the way how humans parse it is l

I actually agree with this take! Deep learning models shine in problem domains where the gap between the data representation to the way how humans parse it is large. In images, for example, we just have RGB matrix but what human vision does is very complex. So representation learning done by deep learning models tries to close the gap. For lot of tabular data, human's don't go from raw representation to a mental model quickly. We in fact follow a model similar to tree based classifier: look at how a variable is splitting the output, compare averages across groups etc.