Neurocomputing Home

Welcome to the new updated version of Neurocomputing.org as it prepares to become a resource center for those interested in simulations of simple animals in a 3D world with realistic brains. Figure 1 shows a prototype simulator with realistic neurons using action potentials. A fully functional simulator is nearing completion. What makes this possible is the recent development of Relativistic State Automata theory describing information processing in un-clocked systems such as the brain. In the absence of such a theory most previous neural network simulations have been based on unrealistic methods using numbers that are supposed to represent the average action potential rate. Alternately other simulations simulate neurons at a very detailed electrical level using such programs as NEURON with the result that the simulations are slow for any neural network consisting of more than a few neurons. Yet in both situations not knowing what parameters are important to information processing severely inhibits network design.

Relativistic State Automata theory is an extension of Finite Automata (Fiite State Machine Theory). Finite Automata theory describes the information processing for clocked systems like those found in computers and digital devices in which the states can be unambiguously identified by the processor itself. Its limitation is that it does not describe the actual state change process. One state magically transforms into another state at the next time interval. Relativistic State Automata overcomes this limitation by mathematically describing state changes in the most general unclocked case. Because of this the information processing in unclocked systems such as the brain can now be described.

Some basic computer simulations (such as shown in figure 1) have proven that Relativistic State Automata works. The next step for accomplishing larger and error free simulations is to create a simulation environment. The main users will probably be computational / theoretical neuroscience researchers although the ability of Relativistic State Automata Systems (RSAS) to handle uncertainty (like the brain) suggests it will also be used in sensor fusion. The RSAS Environment will allow brain researchers who are not programmers to create and test realistic neural circuits for a test animal existing in a 3-D virtual world. It will be released in the Spring of 2007. The paper describing Relativistic State Automata theory will be published here at that time as well. The theory needs to be made available to those researchers world wide who will be doing the simulations and not locked away in a few university libraries or only available online for an exhorbitant fee.

Figure 1 shows an early prototype simulation using a trilobyte in a 3-D world. It's "brain" also exists in a 3-D world and can be seen in the upper left hand corner. Some responses from selected neurons are seen in the upper right hand corner. The trilobyte is shown successfully moving around an obstical by using its antenna. The proper proceedure for simulating the brain is to follow the course of evolution and work towards complexity. Eventually the RSAS environment will be made to run on a set of networked computers.

Characteristics of the Brain Simulation Program

Create, edit, and simulate a network of realistic action potential using neurons with the size only limited by your computer speed. (initially 200 to 300 neurons might be as high as you would want to go). The creation is completely menu driven.
Specify the important neural parameters such as "synaptic efficiency" (wieghting), "synaptic distance from threshold" (affects the base shape and arrival time of the analog electrical signal), "conductance" (affects the persistence and thus the additive shape of the analog electrical signal over several impulses), "threshold value", "theshold depolarization signal level", "threshold depolarization signal conductance".
Make all the above neural parameters adaptive, that is changeable up or down within defined limits by action potentials. (very easy to do classical conditioning)
Each neuron can have up to 6 input connections with other neurons.
Use all the different neuron types as suggested by Relativistic State Automata theory which are the standard SUMMATION, new INCLUSIVE OR, new AND, and standard CONDITIONAL (the subtractive inhibitory neural inputs).
All network data to be saved as open standard XML files.
Compare the signals of up to 4 selected neurons in real time on screen (you can select different neurons while the simulation is running).
Examine and edit the neuronal network in a 3D space through which you can fly the camera.

The program runs on Windows XP and you will need .NET version 2 from Microsoft installed. Because of that it should run on the upcoming Vista without modification. Using two monitors is definately recommended! All windows are floatable meaning you can put them anywhere.

Analog Logic Neurons Found in the Frog Brain

Perhaps the most important prediction of Relativistic State Automata theory as applied to the brain is that the brain must accomplish analog like logic operations since logic operations are used to characterize states by testing for symmetries. Yet some evidence for this already exists in the literature only that its significance has not been realized.

Fig. 2

Figure 2 shows how auditory signals from each ear are combined in a single neuron in the Superior Olivary Nucleus of a frog. (Feng and Capranica - 1978). CL is the response from the contra-lateral ear while IL is the response from the ipsi-lateral ear (on the same side of the brain as the neuron). The CL neuron has a lower sound threshold (44dB) compared to the IL ear (49 dB) so it has a more vigorous response given the same sound intensity. Figure 1 gives the average response at 500 Hz (where they were most responsive) to 20 trials at one trial per second.

Notice that the signals are not summed at the high end away from the unreliable low end (very few action potentials). Instead the response from stimulating both ears is the same as if the CL ear was stimulated by itself. This is a classic multivalued (fuzzy) logic INCLUSIVE OR response.

In another paper (Feng, Hall, Gooler - 1990) the researchers found an AND like neuron. They say this on page 317:

"Unlike DMN (dorsal medullary nucleus) and SON (superior olivary nucleus) neurons however, some (14%) TS (torus semicircularis) cells respond poorly, if at all, to single tones regardless of frequency and level (Fuzessery and Feng, 1982). But these same cells respond vigorously to specific combination of tone and hence represent a neural analog of the logical AND function."

Notice the proposed AND neurons do not respond to individual inputs, no matter how intense the sound, so they cannot represent simple summation operations with thresholds. The neurons must accomplish this via a more complex design.

References

Feng, A.S. and Capranica, R.R. (1978). Sound Localization in Anurans. II. binaural Interaction in Superior Olivary Nucleus of the Green Tree Frog (Hyla cinerea). J. Comp. Physiol. 41:43-5

Feng, Albert S., Hall, Jim C., and Gooler, David M..(1990). Neural Basis of Sound Pattern Recognition in Anurans. Progress in Neurobiology 34:313-329

Fuzessery, Z.N., and Feng, Albert S. (1982). Frequency Selectivity in the Anuran Auditory Midbrain:Single Unit Responses to Single and Multiple Tones. Journal of Comparative Physiology 150:107-119

CHANGE LOG

Oct. 25, 2006 - Changed Continuous State Automata to Relativistic State Automata which is a better description of the theory and that it is an extension of finite automata instead of a compliment

Web site by David D. Olmsted. He can be contacted at brainsim1-contact at yahoo dot com (this is an anti-spam tactic. Type the address as normal). Original site established August 21, 1998 by David D. Olmsted. New home page published August 25, 2006

Information compiled by David D. Olmsted © 1998 to 2006 (Free to use for personal and educational use)