Gary Marcus

if you're not advocating for a "simple hybrid" between symbolism & connectionism, then can you please point to a single implementation that: a) is representative of the thing you have in mind, and b) succeeded on a challenging task that has escaped a pure DL system?

the more general answer is that what i want does not yet (reliably) exist: the kind of thing that lets a human recognize a zebra after being told it is a horse with stripes, or manipulate a grater once told the logic

There's an assumption here that this "thing" must be a hybrid, since the behaviour you outlined does not constrain the mechanism. It's fully possible that a purely connectionist model can accomplish this behaviour.

Also, there's a bait and switch with your argument. You advocate for a more "subtle" hybrid system when criticizing the "simple" hybrid, but you only describe this mystery system by what it can do, not its implementation. Implementation is the point of contention, not capacity

your confusing many points here. Bengio accused me of advocating something I never advocated; that’s the context here. i didn’t claim i had solved the problem.

I asked for clarification of what you *do* advocate. Your "generic answer" described a behaviour. You pointed to some other models in other tweets: can you state exactly what makes these not "simple hybrids"? Can you precisely characterize the hybrid systems you like?

if mendel has described his factors (genes) and not been able to identifying the underlying physiological basis, would you have dismissed him? that is effectively what you are doing here.

Oh come on. I wouldn't exactly call your views on mechanism agnostic pending further evidence. And a vague description of behaviour is not at all equivalent to Mendel's description of heritability, which, I should add *actually does* constrain underlying mechanism