Comment by comex | Hacker News Reader

Comment by comex | Hacker News Reader

Parent Full thread

Interesting. I'd love to see the output of this 25-way machine translation neural network.

Still... I'm not very knowledgeable about AI, but the part of this that describes biological mechanisms really feels handwavy to me.

What we have in our heads is not a cleaned up version of the input. There are no pixels or symbol strings in the head.

All we have in our heads is big activity vectors that cause more big activity vectors.

Big activity vectors means nothing. What is the structure of these vectors? Even if the brain's image processing turns out to have a lot of random complications and exceptions that fit the data your eyes have seen before (even if these innumerable complications are necessary for it to work well), surely some god with perfect understanding of it could still summarize it as generally following such-and-such algorithms. Even if you don't know what the neurons in your model are for, I bet most of them are in fact "for" something comprehensible by mere humans, and understanding that would make it easier to design good starting conditions for models. Not that I can prove it...

(Note: I suspect that the author of this presentation has a more nuanced view than what I'm complaining about - or maybe I'm just misinterpreting it. This comment is just intended to start discussion about what's written on the slide.)

The 'big activity vector' language is partly a reference to the Vector Word Embedding stuff (see [1] for an explanation). The surprising thing about that is that it is possible to learn an embedded of individual words in a multi-dimensional vector space, such that (for example) vector(Queen)-vector(Woman) ~= vector(King)-vector(Man). Which is to say that there's a general 'royalty' direction within the vector space - and all this can be learned purely from seeing large amounts of English text (no 'traditional' supervised training). Perhaps a 'god' could identify the meaning of each direction in the space (or region of words), but the big Machine Learning labs (Google, Baidu, Montreal, Stanford, Facebook, etc) are proving that purely manipulating the vectors in the abstract works really well.

In addition to the 'word vectors' as inputs, the RNNs illustrated are also iterating over an internal state (flowing from left to right though the same network for each new word) - and this internal state is also an embedding of some kind. But it's going to be very difficult to decipher what each dimension here represents, as it's being built purely as a function of the input word vectors, its own previous state and a NN with initially random weights.

Now, although actual 'brain experiments' have shown that individual neuron (or local clusters) apparently light up when particular thoughts are had (alternatively, cause thoughts to be had), each cluster seems likely to be just one aspect of (say) 'dogginess'. So, one area will correspond to the smell of dogs, others to wet noses, others to being outdoors (i.e. all aspects of the overall 'dogginess' concept) - but these things will all overlap in multiple ways with other concept 'vectors'. Which is how huge spaces of ideas are searched in parallel, rather than sequentially (using, say, an is_doggy_quality symbol).

There are also parallels here with the Numenta Sparse Distributed Representations [2].

Overall, this presentation seems to be probing at the frontier of what works, and how to leverage that up into something that's more about 'general thinking' rather than pattern matching. It also appears to be a thought-piece, rather than a conference presentation (though, of course, Hinton deserves to be heard on just about anything in NNs, IMHO).

[1] http://colah.github.io/posts/2014-07-NLP-RNNs-Representation... [2] https://github.com/numenta/nupic/wiki/Sparse-Distributed-Rep...

Reply on news.ycombinator.com