Gary Marcus

advocates of #machinelearning, I am told that you all know that (current) #ML is limited. fair enough. but which limits are you willing to *publicly* acknowledge?https://twitter.com/NotSimplicio/status/1173373706674085888 …

Many limitations in the sense that they're active topics and we don't have answers yet (out-of-distribution testing, not "understanding" at the right level of abstraction, causality, etc), but I wouldn't classify any needing a paradigmatic shift of techniques

Given it's not really falsifiable to say "a neural network can't do this", and yet there may be things out there that work much better than deep learning for much of this - at what point would you say a paradigmatic shift *is* needed? What's the threshold?

Maybe this isn't really answering your question, but like you say there are many (known) techniques that work better than deep learning for certain problems. Hence, no paradigmatic shift needed, since we already have the tools. Maybe just a new way of putting pieces together?

Perhaps another way of saying this is that I don't imagine a totally new *field* (in the likes of statistics, machine learning, etc) being invented over the next 50-100 years to address the problems we're facing now.

I do. That's the difference between you and me. The one I imagine is all about integrating prior knowledge better with current techniques.

Is that more an engineering challenge or does it require new theories, radical changes of perspective, etc? My hunch is the former, which I don't think constitutes a new field per se