The fact that we call the curve fitting/optimization/compression that we do to fit machine learning models to input data "learning" is really unfortunate and le

The fact that we call the curve fitting/optimization/compression that we do to fit machine learning models to input data "learning" is really unfortunate and leads to this kind of conflation. If we trace the path of how we ended up here, it's similar to how people incorrectly refer to loci of DNA as genes. We have behavior analysis where we speak of learning as the conditioning via the antecedant-behavior-consequence loop. There was the Hebbian theory of how the ABC loop manifested physically in neurons. Early neural net papers took inspiration from that that mechanism and called it learning. Meanwhile, actual learning is far, far richer than the Hebbian theory of synaptic strengthening, and has a lot more going on than just operant conditioning. So, please, it's time for everyone to stop pretending that the fact that ML inherited the word "learn" as a term of art for curve fitting has any philosophical weight.