I (Memming) presented Eliasmith et al. “A Large-Scale Model of the Functioning Brain” Science 2012 for our computational neuroscience journal club. The authors combined their past efforts for building various modules for solving cognitive tasks to build a large-scale spiking neuron model called SPAUN.
They built a downloadable network of 2.5 million spiking neurons (leaky-integrate-and-fire (LIF) units) that has a visual system for static images, working memory for sequence of symbols (mostly numbers), motor system for drawing numbers, and perform 8 different tasks without modification. I was impressed by the tasks it performed (video). But I must say I was disappointed after I found out that it was “designed” to solve each problem by the authors, and combined with a central control unit (basal ganglia) which uses its “subroutines” to solve. Except for the small set of weights specific for the reward task, the network has static weights for implementing specific functions, and information is gated depending on the context of the task via a simple action selection circuit. It is still impressive that the authors were able to make a working model of this size, but I cannot imagine that its operations resemble the actual brain at a macroscopic level.
For the 6 out of the 8 tasks Spaun is programmed to perform, are purely symbolic computation (the two exceptions are copy drawing and reinforcement learning). The symbolic computations are done in a “biologically plausible” manner, by first converting the problem into a high dimensional vector space operations, and approximating the operations in the Euclidean space with spiking neurons and linear projections. However, it is not necessary to use spiking neurons as far as the functionality of Spaun. Also, Spaun can add numbers using a very simple abstract number system for numbers from 0 to 9, and by design increment and decrement by 1 can be easily implemented (allowing the counting task).
In summary, Spaun is a hard-wired system that consists of human designed functions (implemented by a network of spiking neurons via least-squares—supervised learning of each part), and it works like a very basic non-universal calculator (Turing sense). Hence, it cannot learn new tasks by itself. Although it seems far from the human brain, it is definitely the first of the kind to implement various cognitive functions with spiking neurons. Hopefully in the future, we other attempts with the same goal will get us closer to understanding of the brain.