>Traditional AI researchers will be horrified by the view that thoughts are merely the hidden states of a recurrent net and even more horrified by the idea that reasoning is just sequences of such state vectors. That's why I think its currently very important to get our critics to state, in a clearly decideable way, what it is they think these nets won't be able to learn to do. Otherwise each advance of neural networks will be met by a new reason for why that advance does not really count. So far, I have got both Garry Marcus and Hector Levesque to agree that they will be impressed if neural nets can correctly answer questions about "Winograd" sentences such as "The city councilmen refused to give the demonstrators a licence because they feared violence." Who feared the violence?
>A few years ago, I think that traditional AI researchers (and also most neural network researchers) would have been happy to predict that it would be many decades before a neural net that started life with almost no prior knowledge would be able to take a random photo from the web and almost always produce a description in English of the objects in the scene and their relationships. I now believe that we stand a reasonable chance of achieving this in the next five years.
>I think answering questions about pictures is a better form of the Turing test. Methods that manipulate symbol strings without understanding them (like Eliza) can often fool us because we project meaning into their answers. But converting pixel intensities into sentences that answer questions about an image does not seem nearly so prone to dirty tricks.
3 comments
[ 2.5 ms ] story [ 12.8 ms ] thread>Traditional AI researchers will be horrified by the view that thoughts are merely the hidden states of a recurrent net and even more horrified by the idea that reasoning is just sequences of such state vectors. That's why I think its currently very important to get our critics to state, in a clearly decideable way, what it is they think these nets won't be able to learn to do. Otherwise each advance of neural networks will be met by a new reason for why that advance does not really count. So far, I have got both Garry Marcus and Hector Levesque to agree that they will be impressed if neural nets can correctly answer questions about "Winograd" sentences such as "The city councilmen refused to give the demonstrators a licence because they feared violence." Who feared the violence?
>A few years ago, I think that traditional AI researchers (and also most neural network researchers) would have been happy to predict that it would be many decades before a neural net that started life with almost no prior knowledge would be able to take a random photo from the web and almost always produce a description in English of the objects in the scene and their relationships. I now believe that we stand a reasonable chance of achieving this in the next five years.
>I think answering questions about pictures is a better form of the Turing test. Methods that manipulate symbol strings without understanding them (like Eliza) can often fool us because we project meaning into their answers. But converting pixel intensities into sentences that answer questions about an image does not seem nearly so prone to dirty tricks.
Anyway this issue has been explained and fixed somewhat in this paper: http://arxiv.org/abs/1412.6572
And karpathy has a post which explains it: https://karpathy.github.io/2015/03/30/breaking-convnets/