Gary Marcus

Think this might also just be a diff of positions due to diff end goals? Those who might view AI as way to understand our brain & intel & leverage it have Graces position. Versus those who see AI as a blanket term wrt to data science+ML & focused on solns to specific problems.

Certainly there are people who use & design ANNs w/o any interest in interating on models of the brain. And that's fine. But that doesnt mean their artificial neurons arent abstractions of real neurons (unless they really take them far off in some other direction)

And people who do want to iterate and make better brain models are on solid enough ground when they start with ANNs

I think I understand where you're coming from. I'm less concerned with the model of neurons in play than the logic implemented in the 'circuits' in use today. The only reason I argue about this topic is that the abstractions in use don't appear to be converging on the subject.

The success of CNNs in predicting the response of real visual neurons (and similar findings for audition) suggest they are going in the right direction wrt getting the logic right. But yea just cuz you hook some artificial neurons up doesnt mean they'll always work like the brain

agreed with @neurograce, e only additional point I am making here is that we ought have some humility here, 33 years after White et al's 1986 landmark worm connectome, which in honesty we must recognize we have not yet fully understood. read it, and weep: https://royalsocietypublishing.org/doi/abs/10.1098/rstb.1986.0056 …

this is a good QA but the query about what visual system has that CNNs lack also include compositionality, ability to recognize line drawings, silhouettes, and other abstractions even without explicit training data, ability to infer relationships between objects, etc