A Practical Mind/Brain Model

(The following is a messy draft. For technical reasons - a glitch in blogger software - the text was trashed. However, it seems that this idea is worth a more formal treatment).

The following is based on correspondence with an old friend, John Heerema. We are both addicted to thinking about problems by imagining what a solution would look like, then "reverse engineering" the solution into a practical program. 

What would a silicon brain look like?

Cortical Column Model

A 10,000 x 10,000 in-memory matrix (100 megabytes)  representing (i,j) connection strength would seem to be feasible for a column modelled on a single CPU), Matrix structure may work better in 3 dimensions, x,y,z where z is the vertical layer, between 4 and 10. In this structure, we may get away with a sparse matrix which supplies non-zero interaction only for neurons that are "close" in 3 dimensions or even a network  (rather than a matrix) that allows communication only with daughter cells of the same stem cell. In any case, it would seem that the memory "budget" for this structure would not exceed 100 megabytes.

The cerebral cortex seems to be a layered structure (like the skin) with vertical properties similar to the pores of the skin, called cortical columns (google cortical column). As with the skin, elements of the cortex are probably daughter cells of a single stem cell. The number of layers varies within the cortex. As daughter cells mature, they communicate richly with neighbor cells so they "know" what to do, what shape to take, when to stop growing, when to divide etc.. This process is extensively studied in skin cells. 

Cortical column structure seems to be "fractal", an observation that is apparently common (google "fractal cortical column"). All this hints at what a computer model of the brain should look like. 

1. Model a "well-studied" region like the visual cortex

2. Simulate cortical columns, perhaps one CPU each. 

3. Simulate lower level columns recursively to model the fractal internal column logic. Each cortical column should have the same software driven by a "similar" table (see below)

4. Simulate column interaction with another layer of CPU's. Column interaction "wiring" should resemble the way the skin is wired to the spinal chord. This should reflect neural plasticity.

Such a model would need to reflect the strength of neural connections in several ways. Column-to-column connections could perhaps be simulated in a different way than internal connections between the 10,000 or so neurons within the column. 

The fractal model of columns indicates that neurons (and columns) tend to respond to the same stimulus in a "similar" way (as in the skin)

The model of column interaction could be a sparse matrix (or network) ignoring interaction between far-off columns, so if we wanted a million columns your interaction matrix would model a trillion interactions with a matrix that only bothered to have non-zero elements for |i-j| < K. Just a few gigabytes of memory, maybe spread out over 1,000 CPU's. 

The Blue Brain project was not aiming at real time simulation (although you'd never guess that from the sales talk). However, the structure above could shoot for real-time performance, especially once the column logic has settled down so it can be burned in to silicon. The computational power of the whole thing would reside in applying its fractal, recursive nature to a narrow universe of inputs (such as visual inputs).

Hofstader's vision of the mind would lead me to think of the pre-frontal region of the brain as structured in a similar way but "stacked" logically ("on the fly") rather than physically (as in the retina), allowing "virtual tissues" to be instantly formed and re-formed at a rate of 40 to 100 HZ (brain wave frequency). To speak intelligibly about how such a thing might eventually "think", it would be necessary to first model perception, the most elementary "memes", which is why you want to start with the visual cortex. But the overall strategy of fractal, recursive logic would still apply. 

It looks feasible to me. The whole strategy could be demonstrated with java in the cloud, which would leave open the actual decision about where logic would "run" (it could, for example, run on multiple distributed PC's or just one PC). First step would be to shut yourself up for six months and read what everybody already knows about this. You'd need to find a research environment to host the project. There are hundreds of strong programs calling themselves "cognitive this or that" to chose from.

In the cortex, columns in the retina and visual cortex are the most studied. Horizontally, the cortex seems to have regional "specializations", such as mapping the body onto regions at the top of your head.