Feedback

We should look at feedback. Most explanations use something like a thermostat or a steam governor to illustrate this. The output is measured and its value affects the input to keep the output within a particular range. Room is too warm so heat is turned down, then room is too cool so heat is turned up. The human body is filled with feedback loops – blood sugar is too high so insulin is excreted then blood sugar is too low so insulin is cut off. With simple loops like this it is easy to follow the logic with your finger around the loop.

 

But as soon as you add a bunch of other feedback loops that share some of their inputs and outputs, overlapping loops, then the logic cannot be easily seen by following the loops with a finger. Overlapping loops can maintain a more or less stable state, but it is usually not the state that any of the individual loops would settle on.

 

These sorts of systems act a lot like iterative equations. Take a simple equation, x=2+squareroot(x). If we start with x=1 and put it in the equation then we get a new x, 2+squareroot(1)=3. If we take our new x=3 and put it in the equation then we get a new x, 2+squareroot(3)=3.73. If we take our new x=3.73 etc. we get a never ending series that homes in on 4. (1, 3 ,3.73 , 3.93 ,3.98 ,3.99) We can imagine this happening very quickly in a physical system. Equations like this generate series that either find a stable end point, or they oscillate between two values, or they run away to zero or infinity. Again this is a very simple example with one value, x, but imagine how difficult it would be to try to follow a large set of such equations with several overlapping unknowns.

 

So now look at the thalamus-cortex-thalamus loops. Every small place on the cortex has axons running to a particular small place in the thalamus, and thalamus axons run back to the same place on the cortex that sent axons to it. We have a billion or so loops. In the cortex each small place has loops with all its neighbouring small places. The same is true in the thalamus. Both maps also have some loops with places more distant then their immediate neighbours. This is what I have called MPOFBL, massively parallel over-lapping feedback loops.

 

How would a MPOFBL system act? It would be hard to say. One interesting type of behaviour is possible depending on the architecture of the loops. The system may oscillate and quickly stabilize on one particular state of activity pattern for the neurons. This state would be the best compromise of all the constraints of the architecture of the loops and the nature of the input (sensory and other). It might be thought of as the best-fit scenario for a model of the world. 

 

Here is the start of the PDP Primer by George Hollick. It is from about the time that the power of parallel processing started to be investigated as an important option. He is talking about an architecture that is similar but not identical to the thalamus-cortex-thalamus loops described above.

“In 1986, James L. McClelland, David E. Rumelhart and the PDP Research Group published a volume which was to become a seminal work in the field of cognitive psychology.

So just what is Parallel Distributed Processing (hereafter referred to as PDP) all about? Quite simply, proponents of PDP assert that the brain is NOT a computer, not a serial one anyway. In essence, thought is a parallel process, a network of multiple, graded constraints being considered simultaneously. Thought is not a single path of constraints being considered one at a time, as in conventional cognitive models. Moreover, structural differences in the network of constraints are important to the implementation of the thought process and can lead to qualitative differences in the final result. This is in direct opposition to previous cognitive theories which assert that structure has no effect on the outcome of the thought process.”

Leave a Reply

Your email address will not be published. Required fields are marked *