I'm not an expert in ML. However, with what little I do know, I wonder how well "training" a neural network really mimics the brain.

Our brains grow as we mature and along the way reinforce pathways as we internalize information. Our interpretation of the world is internal and unique to our individual consciousness.

The part that is "trained" is how we relate that information externally to our environment and other entities.

A child says "Da da" because it learns that sound elicits a reaction. The internal representation of its father is not trained. It is where the neurons grew and reinforced.

Maybe poorly communicated on my part to people who know much more about it than me. Tl;dr: maybe focus less on perfecting the training of recognition and more on the training of mapping the external to what was naturally recognized where the neurons fell.