Gary Marcus

I don’t argue that people can’t use formalisms that involve derivatives just because most people can’t explicitly explain what a derivative is. I think you are confusing formal, conscious use of a certain kind of machinery with what brain does unconsciously.

If you see the reasoning engine as *separate* (and qualitatively different) from the deep learning system that provides it with inputs, then we disagree. Unless this reasoning "system" is a pen and paper to do math/logic. And much of human intelligence functions without it.

That may distill a second point of disagreement; I see nothing wrong with (for some purposes) having separate systems for (eg) image classification vs reasoning. Certainly it is possible in principle to engineer systems that way; what’s your objection? Efficiencies of training?

I wonder if we are mixing different methodological questions here. If I need to build a robust and reliable system today, I would mix deep learning, probabilistic programming, and symbolic methods. 1/

But at the same time, I think it is critically important to see if we can push DL methods to provide a unified solution to perception, reasoning, and action (with robustness and safety). 2/

Indeed. I think @ylecun's position is that to reach human-level AI/AGI, pipeline systems (such as @GaryMarcus's GQ3) will never suffice. In other words, training DL like systems to use logic/symbols/reasoning via gradient descent is likely to be more robust than two-stage system.

What’s the argument that it would actually be robust? No fair overrelying on data from image recognition to explain how such techniques would solve fundamentally different problems (language, common sense, everyday reasoning) that so far haven’t yielded to same approaches.

The ML community is working hard on the robustness question. We know that stochastic gradient descent sometimes confers robustness and other times fails miserably. We are exploring many avenues: causal models, adversarial training, improved models, etc.