13

A few years ago I wrote a research paper for college on neural networks. At the time IBM's Blue Brain was the clear winner. Some rumor went around that they were close to emulating a brain the complexity of a cat; this has since been debunked.

At the time their research was groundbreaking and previously even emulating the brain of a C. Elegans seemed far-fetched.

Describing the complexity of neural networks is complex but I'll define number of synapses or neurons as an easily measurable scale of complexity.

What is the most complex working artificial neural network? What is the relative complexity of this network compared to an animal's brain?

tcrosley
  • 248
  • 1
  • 5
Ben Brocka
  • 9,836
  • 5
  • 53
  • 89
  • define complex... – Jeff Feb 02 '12 at 23:42
  • 2
    @ArtemKaznatcheev what does neurobiology have to do directly with either in this context? They're both different concepts but neural-network is usually used to refer to artificial ones these days so just that one of the two does make sense. – Ben Brocka Feb 03 '12 at 00:31
  • @Jeff for the purposes of this question number of synapses should be sufficient as that (and count of neurons, which correlates with it) is how artificial ones are generally gauged. I've edited the question. Some better reading on neural net complexity is here: http://math.bu.edu/people/mkon/pppp11.pdf – Ben Brocka Feb 03 '12 at 00:34
  • @Ben I don't think number of synapses or neurons (termed neural complexity in the paper you cite) is an accurate way to measure complexity in the way you are thinking about it. I can build an ANN with a billion hidden units that does nothing but compute a XOR function. The paper's other form of complexity (informational complexity) is very specific to the function you are trying to implement. Perhaps you are after the question of biological realism? That seems to be one of Blue Brain's primary objectives, though admittedly I don't know a lot about the project. – Jeff Feb 03 '12 at 01:20
  • @Jeff practical ANNs like Blue Brain always go for biological realism, obviously an ANN that isn't representative of an NN isn't acceptable but that's not a good usable metric. The metrics I always hear to compare ANNs is always either neuron/synapse count or complexity compared to an animal (simple = C. elegans, complex = cat, Holy Grail = Human). If there's a better measure I'm not aware of it but would love to know of it – Ben Brocka Feb 03 '12 at 01:58
  • @Jeff I think perhaps now this is the more useful question: http://cogsci.stackexchange.com/questions/252/what-is-an-effective-metric-of-complexity-for-an-artificial-neural-network – Ben Brocka Feb 03 '12 at 02:13
  • @Ben good question! hehe – Jeff Feb 03 '12 at 02:44
  • 1
    @BenBrocka you might be interested in this article about what-if-anything we should expect from building bigger and bigger models. It came to me via this G+ page which might also interest you. – Artem Kaznatcheev Feb 06 '12 at 22:10
  • @ArtemKaznatcheev sounds very interesting, added to my reading list :) – Ben Brocka Feb 06 '12 at 22:31
  • heres a survey late 2013. note google is very far ahead with "deep learning" visual images. – vzn Apr 08 '14 at 17:33

1 Answers1

8

A (probably incomplete) list of ongoing whole brain simulation projects can be found at http://www.artificialbrains.com/. However, based on the information reported on this site, it is sometimes hard do distinguish what already has been achieved and what is still in the planning stage. Nevertheless it gives a good overview to start with.

H.Muster
  • 2,625
  • 20
  • 30