if you regard hashing as dimensionality reduction, there's certainly hashing involved -- some machine learning pipeline will consume high dimensional raw image data and map it into a lower dimensional feature space
on the other hand, for cryptographic hashing we want very similar but inequal inputs to be hashed to wildly different outputs, so a useful hash function should be highly nonlinear / discontinuous, small changes in the input leading to huge changes in the output.
if we're hashing in machine learning, we probably want the hash function to be (crudely) continuous functions -- with "similar" inputs hashed to similar values, so we can operate on the hashed values to cluster or classify things. we dont want a bit of "noise" in the input to produce a massive difference in the lower dimensional feature space.