[alok-g]

Understanding, for the brain, is ultimately memorization too. Memorization need not be understanding. More below:

Going by a reasonable approximation/assumption that the the information in the brain is stored as synaptic weights, both understanding and rote memorization would end up as those weights.

The difference is how much semantic structuring happens in those neurons/weights. Rote memorization would be closer to relatively unlinked pieces of information whereas understanding would have more cross-linking.

The same thing arises for machine learning too. Certain models basically just store all data. When a good model generalizes, it is able to interpolate/extrapolate better to unseen situations, which is going beyong rote memorization.

The double-descent phenomenon in large neural networks however is surprising and is something I do not currently understand enough.

Upon observing myself, something similar or related seems to happen for the brain too. I've noticed that when memorizing by rote (e.g., practicing a specific song on a piano), I tend to make mistakes when starting out, become better, but then again start making mistakes, and finally improve and settle.