Spaun sees a series of digits: 1 2 3; 5 6 7; 3 4 ?. Its neurons fire, and it calculates the next logical number in the sequence. It scrawls out a 5, in legible if messy writing.
This is an unremarkable feat for a human, but Spaun is actually a simulated brain. It contains 2.5 million virtual neurons — many fewer than the 86 billion in the average human head, but enough to recognize lists of numbers, do simple arithmetic and solve reasoning problems.
Described for the first time in Science, Spaun — the Semantic Pointer Architecture Unified Network — is the brainchild of Chris Eliasmith, a theoretical neuroscientist at the University of Waterloo in Canada, and his colleagues. It stands apart from other attempts to simulate a brain, such as the ambitious Blue Brain Project (see ‘Brain in a box‘), because it produces complex behaviours with fewer neurons. “Throwing a lot of neurons together and hoping something interesting emerges doesn’t seem like a plausible way of understanding something as sophisticated as the brain,” says Eliasmith.
More here.
I have recently published a paper that deals with Eliasmith’s NEF and structured representations, though semantic pointers were not covered. There is an extensive discussion of them in his forthcoming book. Anyway, if anyone’s interested, I can share the paper (it’s not available yet, see https://en.ccpress.pl/produkt/Philosophy_in_Neuroscience_32).
It is interesting to compare the S.P.A.U.N. brain model (*Science*, 2012, 38) vs the model detailed in *The Cognitive Brain* (MIT Press 1991) in terms of their different theoretical approaches to understanding how the brain works. SPAUN models 2.5 million neurons but offers no biologically plausible mechanism for setting the synaptic connection weights among these millions of neurons. Instead, in SPAUN, the connection weights are determined by a standard mathematical procedure for least-squares optimization to perform coding and decoding. Since patterns of synaptic transfer weights are absolutely critical to the performance of cognitive brain mechanisms, this is, in my view, a serious flaw. In contrast, the structure and dynamics of relevant brain mechanisms that are modeled in *The Cognitive Brain* include a biologically plausible mechanism for forming the necessary patterns of synaptic transfer weights for learning, recognition, and action. In this theoretical approach, the object is not to model as many neurons as possible, but rather to model the minimal credible neuronal mechanisms that can perform the essential cognitive tasks. Description of this cognitive brain model and some tests of its competence can be seen here:
https://people.umass.edu/trehub/thecognitivebrain/chapter2.pdf
https://people.umass.edu/trehub/thecognitivebrain/chapter3.pdf
https://people.umass.edu/trehub/thecognitivebrain/chapter4.pdf
https://people.umass.edu/trehub/thecognitivebrain/chapter7.pdf
https://www.people.umass.edu/trehub/thecognitivebrain/chapter12.pdf
https://people.umass.edu/trehub/sparscodtre.pdf
I hadn’t seen Chris’s article, I will definitely take a look and post some comments here once I’ve had the time to digest it.
Note: the article is actually published in Science (available here), not Nature, and the above is an excerpt of a blurb from Nature, and is written by Ed Yong, not Robert Briscoe.