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 …

Lack of generalization outside of the training set makes DL (maybe not all *possible* ML) inappropriate for many problems. Ultimately, the limitation of ML is our own: failure to architect solutions that lever it but also (will) use other tech to advantage

that of course has been my central point since 1998; it rarely appears in discussions of deep learning results, though at this point it is widely known within the field.

This point has always been known. But again, it is a limitation of supervised learning, not of the architecture (deep or not). Geoff Hinton's focus on unsupervised learning for the last 40 years (and me for the last 20) stems from this.

I agree it's a limitation of sup learning *as conceived today in ML*. But schools teach wetware pupils every day and that's sup learning, which implies to me that it's the choice of algorithms and how you put them together (architecture).

School learning is only partially supervised. Schooling teaches a tiny amount of high-level knowledge. The vast majority of our knowledge is acquired early and in a self-supervised manner. Without this background knowledge, schooling would be ineffective. See The Cake.