Hammers

[Ed Pell]

When the best tool you have is a hammer make your problems in the shape of nails.

Right or wrong we are going down the neural network software and hardware and researching spending path. GPUs give us 10-100x over CPUs, GraphCore and Wave Computing give another 100-200x in compute power and power efficiency. Maximum die size GraphCore chip in 10nm tech node with TSV stacked memory on top will get us a factor of 1,000,000 over CPUs.

Folks are learning to do reasoning with nn, see reasoning with schematic loss function. Ben's video from Berlin 2015 about driving vision nn to how a more structured middle layer and Hiinton's capsules are improving vision. We know to use real world sequential data to do unsupervised training. For example books feed in a character at a time with the nn predicting the next letter trains for word knowledge. Books feed in a word at a time trains for sentence structure. The challenge seems to be to drive a rich and structured mid layer.  

[Linas Vepstas]

Ed, I've got a half-written paper here, which talks about the structure of language (and the structure of other things e.g biochemistry) which claims that the natural structure is a mathematical sheaf. That part is now written. The unwriten part is that the neural nets also are defacto working with such structures; I hope to write that real soon .   Perhaps that may clarify the role of nn better.  Perhaps as I write it, I will discover that .. well, its half finished. We shall see. https://github.com/opencog/atomspace/blob/master/opencog/sheaf/docs/sheaves.pdf

--linas

--
You received this message because you are subscribed to the Google Groups "opencog" group.
To unsubscribe from this group and stop receiving emails from it, send an email to opencog+unsubscribe@googlegroups.com.
To post to this group, send email to ope...@googlegroups.com.
Visit this group at https://groups.google.com/group/opencog.
To view this discussion on the web visit https://groups.google.com/d/msgid/opencog/f84b352f-160f-4355-9de9-a4f76369f784%40googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

--

"The problem is not that artificial intelligence will get too smart and take over the world," computer scientist Pedro Domingos writes, "the problem is that it's too stupid and already has."