Hacker News Reader

Comment by codesushi42 | Hacker News Reader

The AI isn't "learning" anything. It's being supplied with more observations so that it can better approximate whatever "curve" that is being fit to the data.

View on HN

What is your definition of learning? I’ve generally seen it being defined as getting better performance on a task, given data.

I think true learning requires generalization performance- prediction on outgroups which were not represented in the training set (https://en.wikipedia.org/wiki/Generalization_error). It's very similar to when a new theory is able to predict experimental observations that previous theories cannot (general relativity and QM being two great examples).

It is learning. Better performance at a task over time (i.e. after observing more data) is how learning is defined in the community. Unless you have a better definition.

Not in the same sense as the article.

Also, "optimizing"?

do you think humans do something differently?

Yes. Perhaps you are mistaken and are taking the terminology "neural networks" too literally?

Do you think a bird uses its wings any differently than an airplane's to fly?

I work in ML and have a phd in biophysics; I asked a legitimate question and it's not at all clear that humans "learn" differently than good ML models generalize.

Nothing is really "clear" about the human brain yet in that regard because neuroscience is such a nascent field. What kind of answer were you expecting?

Also while we are on the subject, is there a general intelligence AI you would like to introduce me to?

No need to be snide

I do. I think humans are pretty bad at fitting data to curves consistently, but we're pretty good at learning to fit data to curves. AI doesn't work until you tell it which data to use and what output you expect, and then it functions quite consistently. Humans are also capable of trying new things and hypothesizing results.

The gap is certainly closing, but it's nowhere near where it needs to be to replace the need for humans. Maybe we'll get there soon, but it won't happen without some new approaches IMO.

Then again, I'm no longer active in the research aspect of the field, so it's hard to gauge how close we might be.

"AI doesn't work until you tell it which data to use and what output you expect, and then it functions quite consistently." -> you mean supervised ML, not AI.

Unsupervised ML is different. From what I can tell (I work in ML and have extensive biology experience), human brains use a combination of approaches that combine unsupervised and supervised learning and oneshot learning and it's not clear that a true "learning AI" requires more technology than what we have today. I think that's an open question in the field but nobody is really addressing it head-on.

Reply on news.ycombinator.com