Hacker News Reader

Sure, but we did figure out how to make things more robust and generalizable, at least for perceptual tasks so far. Knowledge representation and probabilistic reasoning are still non-existent, though. Moreover, nobody is even working on any of that, for fear of being compared to Doug Lenat.

View on HN

Representation learning and probabilistic methods are huge sub-areas of modern machine learning, just take a look at the proceedings of ICLR2019.

Representation learning != knowledge representation, probabilistic methods != probabilistic reasoning. I'm talking foundations of AGI, which as far as I'm aware, nobody is seriously working on at the moment.

So robust and generalizable that adding a pixel here and a pixel there, too small to be even recognized by a human, confuse image recognition system to the point it confuse panda with gibbon?

These adversarial examples are generated in a very artificial way that will not be present in natural images (and if you’re thinking of security issues, the attacker needs access to your model...)

They’re still an interesting topic to explore but hardly evidence that neural nets don’t generalize.

Wouldn't the worry be that generalizing to real world tasks involves enough variation that the same issue could arise naturally?

Reply on news.ycombinator.com