Hacker News Reader

Comment by cscurmudgeon | Hacker News Reader

None of those tasks have been solved fully at human-levels. Subtasks and datasets? yes.

That isn't always true. Face recognition now exceeds human level performance. GoogleNet's image identifier is only 1.7% worse than this guy's performance:

http://karpathy.github.io/2014/09/02/what-i-learned-from-com...

Given that the author is writing a blog about AI, it's reasonable to assume he's above average in intelligence and knowledge, and therefore likely better than the average human at this task. But even if you don't make that assumption, object identification is within spitting distance of human level performance.

The same is true for many of these other tasks.

Not it is not. Learning new object classes from a single image or a few images is very hard. See

http://www.sciencemag.org/content/350/6266/1332.short

Machine translation is a joke.

Put any comment on this page through Google translate to another language and back to English and see what you get.

I did a small part of yours. Hardly human-level for just a small simple sentence.

> But even if you do not make this assumption, identifying the object involves spitting the distance from the performance of the human level.

The fact that one-shot learning is still hard does not falsify what I said. Computers are now better than humans at many tasks that were once thought to be the domain of general intelligence. How they were trained is not particularly significant.

It depends on what you mean by a task. Machines are not universally good at object detection because they fail in cases where there is too little data. We can't magically wish for non existent data (yet humans would do quite well on those data-scare tasks).

Ya, i'm happy to accept that humans are still better at generalizing from scant data. But that is something that's being actively worked on, and progress is being made.

https://en.wikipedia.org/wiki/One-shot_learning

Agreed. I would love to see quick progress in that too (that is one of my projects).

There's been lots of articles recently about reaching "human-level" performence, you should know better than to believe them. The tests are on constrained datasets and ignore many factors. You mention face recognition, you realize for every image the program correctly recognizes, there also exists an image, that you will perceive as identical to the aforementioned one, yet which the system fails for.

Yes, i'm aware of differential attacks on neural networks. That doesn't falsify the hypothesis. There are instances where you will fail to recognize a human that those NNs will not.

Reply on news.ycombinator.com