Gary Marcus and Ancient Critiques of Deep learning

I finally got around to reading a paper (and I do not have a proper reference – maybe that suggests something?) that had been lying about for over a year from an NYU professor named Gary Marcus. Given the volume of references he has to himself, it’s shocking that I haven’t blundered into him before.

I liked the way the paper is organized – motivation, points, discussion. The content, perhaps motivated by the sloppy and ego-driven style, encouraged me to want to push back on every point. I think his larger point is that Deep Learning has broad weaknesses that have kept it from informing a new race of intelligent robots. Each point has some substance, but it’s not clear to what end – they are not technical problems (and he admits as much). As an example, the point is made that there is insufficient integration of other domain expertise into solving a problem. While this is correct, that’s really a criticism about how I organize my tools to solve a problem, not in the efficacy of a particular tool. It’s unfortunate, because that particular paragraph is well written and is one of the few extended portions that has a pleasant logic and flow. I feel that if he had raised his head a bit and organized the intent of each criticism, it would have been a more interesting and engaging read.

A minor note, but one that caught my eye, was his latching onto the relationship between correlation and causation. It was, again, not well written in a prosaic sense, but the gist was clear – that this distinction is somehow ‘fundamental’. And again, there are missing words: fundamental to what? Fundamental to reproducing intelligence? Is that the real goal here?

What caught my eye here is that I know there’s at least one other recent piece I came across in the past year that hit on this idea that a weakness in Deep Learning is not mastering causation vs. correlation and it bothered me. That distinction is barely understood among humans and I’m not convinced it is understood among any random selection of conscious, living creatures among whom I am relatively inclined to label as ‘intelligent’.