A large scale model of a human brain has been created by a team of scientists at the Centre for Theoretical Neuroscience, University of Waterloo, Ontario. This is a virtual model, inside a computer, that involves 2,5 million virtual neurons structures in a pattern resembling the overall human brain’s anatomy, including cortical regions, motor control regions, etc. There are two components of the model: Visual processing including input and visual memory, and motor control sufficient to make a relatively simple, but 3D, arm move so it can draw things. The brain is called Semantic Pointer Architecture Unified Network, or, rather creepily, “Spaun.”
This is a pretty intense project but Spaun is sufficiently complex and integrate that it begins to approach an actual human brain in one important way: The external manifestation of its function is orders of magnitude simpler than its internal workings, so the best way to get a feel for how it works is to just watch it in action. The researchers, clearly understand this, have produce a paper (published in Science Magazine) that comes along with rather impressive videos of their spawn, er, I mean, of Spaun, in action. Finally, a complex research paper that explains itself!
Let’s start with the Introductory Movie:
There are eight tasks that Spaun has been assigned. Let’s look at some of them.
Copy drawing, Image Recognition, Counting, Question Answering, Serial Working Memory
Given a randomly chosen handwritten digit, Spaun should produce the same digit written in the same style as the handwriting.
Now let’s look at something a bit more complex, Fluid Reasoning, which gives us a look at the creature’s, I mean, model’s IQ.
Here is the abstract from the paper:
A central challenge for cognitive and systems neuroscience is to relate the incredibly complex behavior of animals to the equally complex activity of their brains. Recently described, large-scale neural models have not bridged this gap between neural activity and biological function. In this work, we present a 2.5-million-neuron model of the brain (called “Spaun”) that bridges this gap by exhibiting many different behaviors. The model is presented only with visual image sequences, and it draws all of its responses with a physically modeled arm. Although simplified, the model captures many aspects of neuroanatomy, neurophysiology, and psychological behavior, which we demonstrate via eight diverse tasks.
Christian Machens writes a very helpful perspective on the project.
… Spaun performs all … tasks based on the activity of 2.5 million simulated neurons that are organized into subsystems resembling different brain areas, and wired up to provide the necessary functionality…
In each of Spaun’s areas or modules, the actual information is processed through populations of spiking (active) neurons. The link between the high-level computations performed by the networks and the lowlevel computations performed by individual neurons is constructed by means of the “neural engineering framework” (9), which specifies how to implement arbitrary mathematical vector operations in spiking neural networks. The framework assumes that information is read out linearly from neural firing rates and is transformed through the nonlinearities of neural activation functions. Consequently, the information processed by each area is distributed across neurons, and thereby roughly matches some of the known electrophysiological features of the areas, such as the diverse tuning of neural firing rates to sensory stimuli or motor outputs.
Given the scope of the model, … it is not surprising that many aspects of Spaun deviate from real brains. For instance, the spiking activity within many areas differs … from that measured in real brains. To what extent this problem can be remedied in future work, or to what extent these discrepancies point toward fundamentally different computations in the brain, is currently unclear. Spaun’s principal shortcoming is that it is essentially hard-wired and cannot learn any new tasks. Its architecture is quite flexible and not bound to particular tasks, and several parts of Spaun are learned (such as the visual hierarchy or the values of actions). However, learning in its broadest sense, such as learning completely new tasks, is one of the issues that the authors have deliberately—and perhaps wisely—side-stepped. … By assembling a large amount of brain know-how into one model, Eliasmith et al. have provided a coherent theory of how the brain works (with the exception of learning). To paraphrase the statistician George Box, their model is likely to be wrong, but it is certainly useful. …
Spaun was created using the Nengo simulation software package, and run on clusters Orca and Kraken of the SharcNet High Performance Computing Consortium. In other words, it is run on a Linux based supercomputer.
It takes up about 24GB of ram and takes about 2.5 hours of processing for one second of simulated time.
Eliasmith, C., Stewart, T., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., & Rasmussen, D. (2012). A Large-Scale Model of the Functioning Brain Science, 338 (6111), 1202–1205 DOI: 10.1126/science.1225266