Welcome
Hello and welcome to our neocortex simulator website. Our site is new and was created in order to obtain funding for a research venture without the project constraints dictated by the typical Brain Research Institue or Government funding agency. The typical Brain research institute supports projects that gather small amounts of limited data on brains and thus do not have as their main purpose to seek out how the brain works. And the typical government agency employs decision makers and project reviewers who are biased toward similar limited research approaches, again which don't seek the "big" solution.
The Brain Research Corporation is convinced that the only approach to seeking out quickly how the brain works, is via trying out and testing different brain models of complexity and size approaching that of the human. As such, the corporation has designed a general purpose hardware simulator capable of reaching the performance of the human, but still missing many details in order to reach the performance of the human. The following section details the simulator and its initial constraints with the intent of showing the actual simplicity of this new approach and that its assumed generalities may already be sufficient to achieve significant intellectual behavior.
Press Release July 4, 2009
Brain Research Corporation is developing a hardware neocortex simulator capable of modeling a full human brain, consisting of 30 billion neurons. Completion of the simulator is set for two years from now and is expected to exhibit the full intellectual behavior of the human. Prediction of success of the simulator is predicated on the use of neuron nets as observed in the human cortex with the added (the unmeasured) constraint that the connections between excitatory and inhibitory neurons is large enough to guarantee that the inhibiting neuron will pulse if the connected excitatory neuron pulses, resulting in groups of neurons being mutually inhibiting. Adding a second (unmeasured) constraint that the inhibitory and excitatory connects vary by the same rule, results in the nets of mutually inhibiting neurons doing “normalized correlations” where the neuron that pulses is the one whose connection pattern best matches the input pattern. Adding a third constraint that the output of a neuron is its frequency of pulsing, not the pulsing waveform itself, results in a neuron net structure anticipated to exhibit human behavior at a fraction of the 30 billion number of neurons. These deductions are based on an electrical engineering principle that correlations are of little use unless they are normalized (the sum of squares of the excitatory connections is constant over all neurons, as the connections vary due to learning). The discovery that neurons are able to do normalized correlations (with only the connection values being constrained in order to cause normalization), makes it very likely that the cortex is using normalized correlations. Otherwise the nets of neurons would not perform useful pattern recognition behaviors. The modeling of a neuron with a frequency output, rather than a pulsing (at a frequency) output, has the advantage that many less neurons are needed for any useful behavior (it takes many pulsing neurons of a given frequency to produce a steady pattern input to a receiving neuron for pattern recognition purposes).
It is expected that the simulator will quickly evolve toward human behaviors as the trials are made to seek the major cortical to cortical connection patterns of the human brain, patterns which presently are only vaguely known or measured. The major effort of the project is to seek the cortical to cortical connection scheme and the initial values of the connections. Failure is not feasible, since the simulator can be reprogrammed to any connections, initial conditions, and neuron properties.
Using today parts and costs, the simulator will fit into a 2 foot cube and cost about 1 million dollars. This cost and volume will drop each year based on normal integrated circuit evolution of a factor of nearly 2 every year. Also, it is expected that 30 billion neurons is not need for useful intellectual behavior, so that the cost and size can be much lower for present practical applications.
One of the first uses of the simulator will be to serve as the brains for super robots for service in the military. The costs for these robotic soldiers will be significantly less than that of the human soldier, the numbers needed greatly reduced, and the performance greatly increased (these robots will be difficult to stop or kill.)
Although the intellectual performance of the neocortex simulator will be able to exceed that of the human, the units produced will be constrained to keep the machines in certain areas of endeavor so that the human is not displaced unnecessarily. The market size for the simulators will exceed 100 million units (world wide) in the near future for commercial purposes. Obviously the greater intellectual performance will be useful in research areas as well. Certain laws will be needed to keep these machines from replacing all humans.
Although funding for the project is from the two inventors of the simulator, in order to speed up the developments, funding is being sought from NSF, DARPA, corporations, and private investors.
Inquiry to this project can be made through email address, ronswallow@ptd.net. Doctors Ronald J. Swallow and Richard L. Swallow are the principle investigators.
BRAIN RESEARCH CORPORATION
Box 304, Trexlertown, Pa 18087
To the investor February 15, 2010
ASSERTION
I assert that the human neo-rcortex simply consists of layers of mutually inhibiting groups of neurons employing conditioned reflex leaning between stimulus sub patterns. All learning can be explained by area to area connections including area feed back forming short term memory. To prove this assertion, a neo-cortex simulator must be built. Success of demonstrating full human intellect has profound benefits to mankind and enormous profit potential. A four million dollar project is all that stands in our way.
EXPLANATION
It has been know for many years that the neo-cortex consists of many layers of neurons, connected one layer to another, relaying signals from the sensors through the layers eventually to the motor neurons. This connectivity structure can be identified by observing slides shaved from the cortex tissue and the gross connectivity determined by stimulating each area of the cortex and observing which areas are excited.
But because the neurons are so small, the individual pulsing behaviors of all the neurons of a region cannot be probed, leaving the detailed interconnection scheme basically unknown.
What we are basically missing are the detailed axonal connection strengths in the layers of neurons. Fortunately, we can deduce what matters with respect to these unknown connection strengths.
A neuron performs the cross correlation between its input axonal excitation pattern and its axonal connection pattern to that neuron forming the neuron’s psp within the neuron It is well known that correlations that are not normalized are generally worthless. Also, these correlations must be compared among a group of neurons so that the neuron with the largest correlation will fire alone and the others will not. The only way this can be accomplished by a group of neurons is to make them mutually inhibiting by sharing an inhibitory neuron that is excited when any of a group of excitatory neurons fires, so that the first to fire (the one with the largest psp), will prevent the others with a smaller psp, from firing. In order that the maximum psp be in the neuron whose input best matches the axonal pattern to that neuron, the axonal patterns must be normalized, i.e. their square root of the sum of squares of those connections to a neuron must be the same for all neurons of the group.
Now this may appear to be impossible to attain by simple neurons. However, the inhibitory signals to the neurons of a group pulse at the frequency of the excitatory neurons and if one studies (using calculus) how those inhibitory connections vary when the excitatory connections vary (using simple conditioned reflex axonal growth rules), one discovers that they remain proportional to the square root of the sum of the squares of the excitatory connections, effectively normalizing the excitatory psps.
This property of groups of mutually inhibiting neurons was discovered by Dr. Ronald Swallow over 40 years ago while he taught Electrical Engineering at Purdue University. This led him to believe that the neo-cortex simply consisted of layers of mutually inhibiting neurons. However, the state of art of the electronics industry was immature at that time and only today would it be feasible to build a full cortex simulator for 30 billion neurons. Thus, six years ago he began to design an electronic simulator for a full brain. Today, he now seeks the funding to build the first prototype which can be studied and modified to approach the behavior of the human brain. That is, what remains to be determined is the area to area connection scheme of the cortex and the initial connection values
In order for connection patterns to become proportional to the stimulus patterns, simple conditioned reflex learning rules are assumed. In order for the brain to exhibit stable, long term memory, connections needed to be enabled at times to change. Thus, an axon would grow toward a neuron only under certain conditions. Glial cells were assumed to shutter this growth. The simplest rule was to enable axonal growth when the excitatory psp fell below a threshold. (The glial cells can detect the excitatory excitation to a neuron and can shrink to allow growth or not.). The result of this rule is that learning occurs only if the stimulus patterns are “new”. This favors a curiosity behavior. With axons only able to grow or not, permanent memory is guaranteed and results in old age reduction of memory formation observed in humans. Other rules for axonal may be feasible and requires further studies.
The only way that we will know if the cortex consists of mutually inhibiting neurons is to build such a cortex by simulating it, electronically. That is the purpose of such a proposal.
CONCLUSION
I am seeking the funds to build and test a full brain simulator. I have located potential funding sources and only need a small amount of seed money to complete the acquisition of research funds. But I also will entertain full funding support from your organization if that be your requirement. I have fallen short of funds to acquire this funding and am coming to you for help. Let me know if you can and will help me.
I do not have your phone number of email address. If you are interested, please that data
so that I can further communicate with you. I have considerable documentation for the proposed project. Email may be best, since my phone may be terminated due to lack of operating funds until I am paid again by unemployment and social security.
Sincerely
Dr. Ronald J. Swallow
610 704 0914
DETAILED RESEARCH PROJECT
PROJECT SUMMARY
The objective of the project shall be to build a programmable general purpose hardware simulator, capable of modeling in real-time the complete cortex of the human, equivalent to a complexity of 30 billion neurons and a cost of under one million dollars. This two year project will be limited to a two billion neuron sized simulator which should be of sufficient size to verify the validity of the neuron model and neuron net connectivity and initial condition properties. The simulator will be programmed to model nets of neurons, employing a variety of connection shemes and initial connection strengths. Via the use of a personal computer, the neurons, their connections, and their initial connection strengths will be experimentally varied until the simulator eventually begins to approximate the behavior of the human. This cortex simulator shall be connected to a robotic device with visual, auditory, and propeioceptive inputs and motor outputs for the purpose of studying its learning behaviors in a human context learning environment. Initial major connections shall be drived from studies of embryo growth in mammals. Derivation of the detailed initial connections and connection strengths are the main topic of this study.
As a means to reduce our search for the human brain model, it is not generally known by neurologists that mutually inhibiting neurons will exhibit normalized cross correlations when both excitatory and inhibitory neuron synapses use the same Pavloian learning rule. Dr. Swallow, 40 uyears ago, discovered this property of a group of mutually inhibiting neurons, but has had to wait until now to be able to afford the cost of constructing nets of neurons large enough to model the human cortex. It is his belief that nets of neurons that perform correlations (which is the most basic function for a neuron), must be able to do normalized correlations in order to perform useful behavior. Although previous studies of the human cortex had shown that it consists of 20 percent inhibitory neurons which suggests the possibility of mutually inhibiting neurons, no study has proven that these neurons operate as groups of mutually inhibiting neurons. None the less, we shall further limit our search range for our neuron net properties by making the assumption that the cortex consists of mutually inhibiting neurons. This reduces our trials to that of seeking general connections between homogenous groups of mutually inhibiting neurons and the initial conditions between those neurons. If our studies find that mutually inhibitng groups do not behave properly, our hardware simulator will be capable of modeling any other neuron behaviors.
The human cortex varies which areas connect to which areas, but the areas themselves are very similar if not identical throughout the entire cortex. It is our belief that making trials of cortical to cortical connections and their initial connection strengths, should yield quick and interesting results when studied with the simulator.
The simulator shall be able to consist of up to 1000 identrial three inch square circuit boards plugged into each other, each containing a single field programmable gate array chip (FPGA) and about 19 memory chips, together costing under 500 dollars a board. Each board will be able to be programmed to model any neuronal net of up to 512K neurons.
The cortex simulator will be interfaced to sensory input hardware and motor output hardware. In particular , a camera with a neuronal interface, audio input with za neuronal interface, and a motor and proprioceptive neuronal interface will be employed. An electric motor system to control the direcrtion of the camera and to model a subset of the human body arm and leg elements will be employed in this simulation project.
AN UNKNOWN FACT
Very few of us know that we are born without our real brain, the cortex. At birth, our cortex consists of a small number of cells which during the next week duplicate and spread into a large convoluted shell of over 30 billion neurons, yet unconneted. During the following months, axons extend between these cell bodies and those of the thalamus, below. Except for reptiles, which have a more rudiment cortex, no species below the reptile have a cortex. It is the cortex that gives the human (the animal) its intellect. It is an add-on with a rather homeogeneous collection of neurons. Yet, the cortical areas, by connecting to the thalamus and the many other ortical areas, exhibits a unique behavior, characteristic of the animal kjnhdom.
Its neurons and connections develop on their own, exhibiting learning and other behaviors only exhibited by the animal. The neurons of the cortex appear quite homeogenious. Yet, variations in the cortical to cortical connections and the following learning, result in its unique intellectual behavior. The following brain model and project attempts to tie down the connection rules including their initial conditions that result in human behavior. We know that this add-on set of neurons had to evolve from but a few added neuron net properties, because it has simply eolved by adding a large area of what appears to be neurons essentially idential to most other neurons of the lower brain. Buy working with a simulator for the cortex, we hope to evolve the missing properties of the cortex. We hope our belief of its simplicity is correct and that we shall quickly observe the human behaviors expected of the cortex in our studies.
POTENTIAL APPLICATIONS
Success of this project will depend upon it deriving a brain model which produces behaviors similar to the learning behaviors of the human. Having reached this goal, it will be possible to apply this new knowledge in order to amplify human intelligence, or to reduce the cost of human behaviors, or to buld machines able to perform dangerous tasks such as required of the armed services, to name just a few purposes. Just to know how the human brain works would contribute significantly to our understanding of us and the world.
Further, it is likely that by varying cortical interconnections and initial conditions of the neuron models, that the intelligence of such models can be altered and increased. By providing other sensory inputs as microwave, laser, ultraviolet, etc., the simulators performance also will be able to be increased. Also, its computation speed can be increased by orders of magnitude making certain tasks possible that require that speed. Attaching the artificial brain to robotic bodies will permit performance in environments outside the range of the human body.
The neurological research community has failed to begin projects as this one becaurse they are not sufficiently inter-disciplinary. Even when there is an attempt to do so, the projects fall far short of what is possible - becaause their staff is not fully interdisciplinary. This proposed project requires the combined knowledge of neurology, biology, physical sciences, electronics, and math. Also, employment of special purpose hardware for simulation rather than the employment of maxi-computers makes possible a simulator able to reach the performance of the human at a cost several orders of magnitude less than otherwise possible and under one million dollars. Until now, the typical neurological research effort is generally unable to build special purpose hardware simulators and been constrained to use computers at costs one thousand times higher.
PROJECT APPROACH
It is not difficult to understand how knowing how to buld a human brain will permit us to build machines that can equal and surpass the capabilities of the human. We will have no difficulty in applying this new knowledge. However, what is needed is a new approach to find out how to acquire this new knowledge. Biologist and neurologists have poked around in brains, now for many years. Although some measurements have been useful, it is quite obvious that one will never be able to measure enough to buld a human brain because of the brains size and complexity.
What we are proposing here is an approach more likened to how nature achieved the creation of the brain - via trial making. Evolution, via a few trials, did arrive upon our brain and it did so without any intelligence. Certainly, we should do as well by employing our acquired knowledge and reasoning abilities in the search for human brain behavior.
Employing a simulator able to be varied in struture and performance, we will be able to duplicate the trial approach of nature and achieve human behaviors. Also, the costs to make these trials will be small since a full brain simulator right now would cost less than one million dollars and would not need to be rebuilt again and again for additional trials. Via further research efforts, such simulators' cost will fall an additional factor of 10 in less than 5 years, and will continue to drop 10 times every additional 5 years making these simulators quite cost effective devices in the near future. Another advantage of simulators and hardware machines in general, is that they can be duplicated both physically and behaviorally without the requirement for each new machine to be taught individually as is the case with the human. These devices will be able to be cloned including what they have learned.
The project will begin with the design of the small circuirt board. Software will be written to load the boards with initial connections and connection values. Software will be written to dump and reload rthe memory states to and from mass storage, so that snapshots of the simulator can be taken in order to begin new experiments while old experiments are still in progress by other researchers. The size of the simulator will be able to be increased in board count to at least 1024 boards, capable of reaching human cortex size.
A robot will be built to interface with the simulator. It will have vision, auditory input, and proprioceptive input. And will have motor outputs, modeling the major muscles of the human body, to move legs, arm, head, and eyeballs. It will model the voice output system of the human body art a later date, after the cortex model has been validated by this project.
Experiments with the devices will seek self-organizing behaviors likened to that exhibited by the human baby which eventually results in crawling, walking, talking, etc.. This behavior success will depend upon our finding the required cortical-cortical interconnections and initial conditions. These initrial conditions must be simple due to the limited number of genes involved in the human. We shall need to deduce these initial parameters knowing the genetic limitations of how they are achieved. Research literature hopefully will yield additional information, but probably additional experiments on mammals will be needed, in order to gather sufficient data.
There is the strong possibility that very simple cortical cortical connection schemes with very simple initial connection strengths (non zero connections between the corresponding neurons of each group of mutually inhibiting neurons) will produce human-like self organizing behaviors. The simulator hopefully will then learn from varying stimuli, to determine the correlations between the various stimuli and as a result learn to recognize patterns of stimuli or objects such as a car, a person, a house. That is, it must be able to recognize an obgject invariant of its size, position, orientation, and color. It must recognize spoken workds, invariant of the speaker. These capabilities are learned by the human on his or her own.
Thus, it will be most useful to know the sequence of embryo cortical cortical interconnection growth in order to properly understand the cortex connection sheme. Furthermore, those same studies could help to tie down the initial synaptic connection values that need to be found. Of great importance will be to make all sensory inputs able to reach the same groups of neurons, so that correlations between these inputs can be learned. As indicated above, a promising approach is to initialize each sensory input modality to excite a differnet neuron of the mutually inhibiting groups, and to have these groups initally connected with large connections beteen the same member of the groups (all neuron ones conected to only neruron ones of all groups, all neuron twos mutually connected, etc.)
Thus, our proposed project will need to extract as much as is kinown of the embyronic developments to aid in making the simulation trials. This task requires personnel searching the literature as a major component of the project.
Of greatest importance tro this project is its reliance upon the behavior of neurons to exhibit normalized cross-correlations. This behavior was discovered by Dr. Swallow over 40 years ago while teaching electrical engineering at Purdue. Unfortunately, after two publications of this normalized neuron net study, not one researcher has ever communicated to Dr. Swallow either the knowledge or an understanding of this model. We can assume either that they never read the papers or at least never undersrood the papers. This project owes irts potential success, to this potential for neurons to do normalized cross-correlations. Without mutually inhibitng neurons, nets of neurons cannot function in useful ways as demonstrated by all of the failures from research in artificial intelligence that has taken place over the last 50 years.
COMPETITORS
The only comparable effort by the research community in studying the human cortex is occurring by a European consortium of around six compnies and IBM, using a maxi-computer by IBM. This effort was begun around 6 years ago, but it is too ignorant in its approach in that it has not applied intelligent assumptions to achieve the proper initial model for the small neuron nets that are common throughout its entire cortex models. In particular, the neuron net models consist of random nets of pulsing neurons without any sense to the interconnections and inital connections between these small groups. The importance of correlations by neurons is not proposed or even mentioned. Certainly, the understanding of correlations and that they must be normalized does not occur in the reports from this research consortium ( the concept of normalized correlations comes from the electrical engineering discipline and is not on the minds of the typical neurologist.)
The consortium mentions FPGAs, but is biased to putting memory in the FPGAs rather than in additional chips outside the FPGA. This indicates a poor understanding of the costs of memory in memory chips versus the cost in an FPGA where memory costs are orders of magnitude greater. Thirdly, experimentations of simulating essentially random nets of neurons via the IBM computer, continues to emphasize pulsing behaviors of the neurons and their frequencies, rather than their partrern recognition and other behaviors. There is no mention of mutually inhibiting neurons because no one is aware of the importance of normalizing correlations ( as observed in the electrical engineering discipline). As a researcher of brains, we have deduce this need for normalized correlations as a basic cause for intelligent behavior to result. If as an alternative, we must first measure mutually inhibiting behavior before we use it in our models, we may never arrive upon correct neuron net models. The poinrt of all of this is that we must make intelligent deductions as to the structure of the nets of neurons in the cortex when we cannot measure these details. Each correct constraint in our trials gets us closer to an understanding of the cortex and eliminates a lot of trial making.
The brain has been a subject for study, now, for many decades. Hundreds of neurologists have measured and studied it, but witout any singificant or coherent model of how it works. The cortex is a mystery, yet to be explained in its amazing behavior - human behavior. Dr. Swallow, now proposes a simulator intended to be able to model the human. Certainly any reviewer of this proposal will likely give it little chance of success because of the lack of many details. However, Dr. Swallow has taken the approach that we have been acquiring bits and pieces of knowledge from diverse fields of knowledge from other disciplines that permit one to deduce the basics of how the brain may work. He presents his thought process in the following sections with the hope the reviewer will believe that his deductions are not as wild as theories or models by those before him and that the proposed project has a decent chanced of succeeding.
IMPORTANCE OF NORMALIZED CROSS-CORRELATIONS
IMPORTANCE OF NORMALIZED CROSS-CORRELATIONS
The use of a neuron net model which exhibits normalized cross-correlations is useful beyond the fact that it reduces the number of trials in the search for the human cortical parameters.
Neurons perform cross-correlations between the axonal input patterns and the associated axonal synaptic strength patterns. It is very important that nets of neurons can differentiate between these stimulus patterns, otherwise the human could not recognize different faces, cars, sounds, objects, etc. The mechanism for such recognition would require neurons, each with a connection pattern representing an object (face, car, etc.). The stimulus requiring recognition would need to cross-correlate with these different stored patterns, with the maximum normalized correlation determining the recognized pattern.
In order to determine the maximum correlation, these correlations would need to be compared, permitting the selection of the maximum one. By placing these neurons in a group of mutually inhibiting neurons sharing an inhibitory neuron, the excitatory neuron that fires in the group will be the one with the maximum correlation (the inhibitory neuron is triggered to fire if any of the excitatory neurons of the group fire, and as its inhibitory effect on the excitatory neurons decays, one of the neurons will fire again, repulsing the inhibitory neuron and preventing the other neurons from also firing). The shared inhibitory neuron of a group does more than comparing the correlations. Its connections to the group of neurons, learn (grow) along with the excitatory connections to the neurons of the group and surprisingly these inhibitory connections each grow and remain equal to the square root of the sum of the squares of the corresponding excitatory neurons of the group as their synaptic strengths learn, so that the post synaptic potentials in the excitatory neurons are proportional to normalized correlations.
The fact that simple neurons (both excitatory and inhibitory), whose axons’ connection strengths grow at a rate proportional to the axon pulsing frequencies, results in the inhibitory neuron connections to the excitatory neurons remaining proportional to the square-root of the sum of the squares of the corresponding excitatory connection strengths during learning, is not obvious. Only a study of such a net of neurons can show this result. Appendix A shows the proof of this behavior.
But the fact that it is true for such a simple net of neurons, makes it likely that this feature is used by the cortex. If this fact was not true, one would question whether nets of neurons could ever exhibit correct normalized pattern recognition behaviors.
Appendix A describes a net of mutually inhibiting neurons. A pulsing neuron does not present a steady signal to a receiving neuron, its pulses produce an intermittent signal to a receiving neuron. A better situation would result if the frequency of pulsing was represented by a single input, rather than by the group of pulsing train themselves.
.
Thus the only way frequency can be detected is to have more than one neuron at the same frequency. Each group operating at a particular frequency together will act as a single neuron producing a single input of value frequency. (These neurons would pulse at different phases relative to each other, with their average voltage level representing their frequency). Thus, it might well be that our boards, each modeling 512K frequency neurons is really modeling say around 64 times that number or 32 million pulsing neurons.
OTHER CONSTRAINTS
In order to achieve normalized cross-correlations, connection strengths increase, never decrease. Detailed observations of the synapse connections will show that the axon tip expands during learning, increasing its area of contact with the receiving neuron. The area expansion coincides with the arrival of a pulse to the axon tip from the cell body. This expansion appears to occur due to vesicles at the tip breaking through the synapse cell wall which is in contact with the receiving neuron. As the vesicle breaks through, sort of allowing the area of contact to increase when the cell wall repairs itself, the vesicle emits acetylcholine into the gap between the neurons. The acetylcholine causes the receiving neuron to increase its internal voltage (psp) proportionally to the amount of acetylcholine received (proportional to the area of synapse contact.
If synapses always grow when pulsed, the brain would expand forever. Thus there must be a mechanism to enable and disable this process. The simplest method is to assume the glial cells are involved and that they fill the unused volume (number of glial cells is around ten times that of the neurons) and shrink to allow the synapse expansion. This enabling would be under chemical control and this chemical must be released due to the pulsing axon tip. That chemical must be acetylcholine or a related chemical (such as the esterase which breads down the acetylcholine). Thus the glial cell will allow growth or not, depending upon the amount of these chemicals.
The amount or density of these chemicals will be proportional to the excitatory cross-correlation (no inhibitory neuron effects) which in turn is proportional to the summation of the effects of the excited synapses. Thus, we deduce a good rule for switching learning on and off is “excitatory cross-correlation less than a fixed limit”. We shall use this rule initially in our simulations.
An interesting consequence of using this rule is that the brain is rewarded by “newness”. That is, as the cross-correlations drop and learning is permitted if the sensory pattern fails to match the synaptic pattern by a critical amount. Thus, the human might be said to tend to remember (learn) whatever triggered trials that produce newer inputs. This simply causes the human to act curious.
INITIAL CONNECTIONS
Thus far, we have been able to deduce important neuron net constraints, such as normalized correlations and the newness condition for learning. It gets more difficult now to deduce how areas of the cortex connect and what their initial connection strengths are.
We shall make an assumption here. Each cortical area, say the first area fed from the thalamus, (via a chemical trigger) starts to pulse its young neurons which have yet to produce axons. These pulsing neurons will produce axons that grow out from the inner side of the cortex layer into the volume below. Assume there is an electrical field bias to start them to grow toward the adjacent cortex area. Assume the axons at the edge of the area (next to the adjacent area not yet producing axons,) immediately turn into that inactive region. Adjacent axons will follow the adjacent axons and will turn into the cortical area further away until the last axons from the first area will come out, pass over, and then into the farthest section of the adjacent area (second area). This process repeats with the second area outputting and projecting it axons into a third area, etc.. This process is programmed via the cells RNA and DNA which emit chemicals that excite neurons.
Other areas of the cortex can be sequenced in a similar fashion, with the last areas projected into, often being common to the last areas of other projection sequences. This process permits all areas to be layer to layer sequences which eventually meet with each other along the projection process. As the projections reach the frontal area, all senses are found to eventually project to common areas. Without these common areas the human brain would not be able to mentally associate all its senses together in a thought process or motor action.
Eventually this projection process maps all senses to all motor areas where these final areas output down the spinal cord to corresponding muscles.
This brings us to the problem of initial connection strengths. Memory is finite. Axons and synapses only grow for the most part. The “newness” rule of growth will over the years permit less and less growth, which is the normal consequence of aging. Thus, the initial conditions must leave most of the growth potential for future learning.
The initial conditions that would do so, are to begin with non-zero connections between one neuron of a group to only one neuron of the next group, etc. This would require that these neurons, one by one of each group, pulse together long enough to connect with non-zero synapse values. Since the number of neurons in a group will probably be around 16, these initial conditions would have used up less than 6 percent of the cortex memory.
We must assume that the brain has this ability to trigger the pulsing of different groups of neurons in synchrony. Thus, each sensory neuron will initially connect with non-zero synapse values through its layer to layer projection to a subset of motor neurons so that at first all muscles can be conditioned to respond to some stimulus and later can be conditioned to respond to any stimulus.
Since the newness calculation also depends upon the input frequencies which in turn are proportional to their detected normalized cross-correlations, each neuron as it learn different patterns, correlates less perfectly with an incoming pattern and thus produces on the average lower frequencies of output. Thus, the threshold for learning effectively drops and effectively permits more learning as the consequence of its earlier learning.
OBJECT RECOGNITION LEARNING
One of the remarkable features of the human brain is that it learns to recognize objects or sounds invariant to common variations such as a car invariant to its size, location, orientation, or a sound or word invariant to the speaker. This process must happen automatically in the cortex. A simple explanation of how this happens is based upon the tendency for subnets of neurons to feed back on themselves from their feed forward areas and be conditioned to hold their state when the remaining inputs vary. Thus a subnet being triggered by an image of a car at some size and location on the retina, will hold that state as the image of the car and its surround, continues to vary during the short time after. This allows these adjacent (in time) images of an object to be conditioned to trigger a single sub-image response. Thus, an inner layer response (further relays in from the sensor) can be conditioned to a large number of outer layer responses which are classified as minor variations of the inner layer response. This result is a natural response of inner (toward frontal) areas to variations of outer layers where all areas receiving feedback from their output effects will tend to lock to certain states until the input snaps them to new states.
If one thinks about this process, there is no alternative but to have the cortex experience and remember all these variations of an object that are all classified as the object. There is no magical process here. Fortunately the cortex is very large in neuron counts and should be able to solve the recognition problem this way. The amount of memory required to store all the variations is greatly reduced if the object, say a car, is actually recognized as made up of sub objects as wheels, fenders, hood, etc. This process would make the learning of an Oldsmobile a quick result of learning other car makes and their parts.
Since the neocortex pretty much learns on its own, what it learns simply depends upon pattern statistics and which cortical areas connect to each other. For instance, in the visual cortex, the connection patterns to the 16 mutually inhibiting will look like edge detectors, since the most common image from a small area of the retina tends to be edges. Similarly, conditioned reflexes between different cortical areas will seek out correlations between their synaptic excitation patterns, producing the observed behaviors of the human brain. What matters is that correlations between excitation patterns of the cortical areas are automatically found by the conditioned reflex behavior of groups of neurons and that the result of the conditioned reflex behavior is human behavior.
FRONTAL AREAS RESERVED FOR LATER LEARNING
The effects of feedback will make a stimulus pattern vary less and less if its effects are traced inward, following its projection areas. We can deduce that subnet lockings will greatly reduce pattern variation reaching the more inner frontal areas. Thus, higher level functions or functions learned later in life will occur further into the frontal areas. Memory in these areas will be higher level types since they will employ recognized patterns and not the elements of the pattern such as the sensory pixels. Generally it can be concluded that areas further (layer wise) into the cortex, will learn later in life.
SOFTWARE REQUIREMENTS
Each board (of up to 1024 boards proposed) will be addressable from a PC interface bus. The synaptic memory (32 bits per synapse) will be able to be read and written. The board to board connection memory (not yet discussed) also will be able to be read and written.
Programs will be written to aid the user to define the cortical areas, their general interconnections and the initial connection (synapse) values. Learning modifies only synapse values. Thus, a PC program will be written to dump or load the synapse memory to or from the PC disc. The researcher will be able to run the simulator and dump its state when done. Another researcher will then be able to load the simulator from his earlier dump and continue to run until he dumps the simulator again. Each board can simulate 512K neurons (initially organized as 32K groups of 16 neurons), each neuron receiving with 4096 input synapses, requiring 64 gigabits of DDR memory per board. A full brain simulator would require 64 terabits (64 trillion bits) of DDR memory or two trillion 32 bit words or 8 trillion bytes for synapse memory. If 4 bytes can be written in 10 nanoseconds, a dump would take 5000 billion nanoseconds or 5000 seconds or about 2 hours.
NEURON NET BOARD REQUIREMENTS
The circuit board is optimized to simply perform as many cross-correlations as possible. The Xilinx FPGA will be able to run with a 5 or 6 nanosecond clock cycle. The latest DDR memories will be able to read a 32 bit synapse connection value every 6 nanoseconds. The largest DDR memory holds 4 gigabits of data, or 512 billion bytes (memory words), or128 billion 32 bit datawords (synapses, each using 4 bytes/words of DDR memory). The FPGA input/output pin count and board size suggests limiting the board to 19 DDR memory chips utilizing 128 data i/o pins and internally running 16 multipliers and accumulators performing 16 correlations in parallel. After 256 cycles, the 16 correlations (each taking 16 cycles) will complete, permitting the determination of which of 16 neurons has the largest correlation and is the one to pulse (assuming these 16 neurons are in the same group and groups consist of 16 mutually inhibiting neurons).
Assuming 16 neurons per group, and a net update time of 48 milliseconds, there are 8 million available clock cycles (6 nano seconds each). If each group receives 256 inputs (from 256 groups each containing one of 16 neurons firing), then 8,000,000 divided by 256 or 32K groups can be simulated by this FPGA. The total amount of memory in the 19 DDR memory chips is 32K (groups) times 16 (neurons per group) times 256 (input groups to each neuron) times 16 (neurons per input group) for a total of 2 billion 32 bit words or 8 billion 8 bit words (which is 4 Gegabits per each of 16 DDR chips). 32K times 16 times 256 cycles of work per frame (doing 16 computations per clock cycle) or 8 million cycles of work results. With a 6 nanoseconds clock rate, 48 millisecond frame rate results.
The board consists of four major circuits. First there is a DDR memory containing pointers to the groups that feed each of the 32K groups whose states are solved by the board. Second, the groups’ pulsing states are stored (and read from) in a second DDR memory (one of two DDR memories so that one can be read, while the other is written (to be read during the next frame solution)). Third, 16 DDR memories contain the synapse variables used by the 32K groups being simulated. 16 memory chips are required so that 16 synaptic inputs can be read in a single clock cycle of 6 nanoseconds, and fed to a sum of products pipe which is capable of solving for a psp (post synaptic potential) every 16 cycles. After this process repeats for the other neurons of a group, (another 240 clock cycles), the maximum psp among the 16 solutions can be determined. This defines the state of that group for use during the next frame solution. This processs repeats for 32K neurons where the 32K state solutions are fed out of the board to a data buss that reaches all of up to 1024 boards where each of the boards stores these computed states (alternately per frame solution) in one of the two ping-ponged DDR memories described above. This buss is capable of transmitting 32,000,000 group states during a frame solution and is the factor limiting the number of boards to 1024. If the groups contain 16 neurons, the maximum number of neurons that can be simulated is 16 times 32,000,000 or 480 million neurons. Since each neuron models 64 pulsing neurons (its output is a frequency, not a pulse), this simulator is limited to modeling the frequency behavior, not pulsing behavior, of 32 billion pulsing neurons via the simulation of only 480 million “frequency” neurons.
Assuming 16 neurons per group and a layer of 32K groups, an array of 180 by 180 groups would be involved per board. If each neuron receives from 256 groups (a limitation that we are placing upon the simulator), it is receiving from a 16 by 16 group array. Thus, if each layer is the same size (180 by 180 groups), each neuron in a sending layer can reach a group up to 16 groups away in the receiving layer. Thus at best, 12 layers of neurons is required so that any neuron of layer 1 can be affected by any neuron of layer 12.
If the retina projects through layers of 1 million groups of neurons, then a 1000 by 1000 groups layer would need to relay through 1000/16 layers i.e. 60 layers in order that an input from the edge of the retina can reach the opposite edge of an inner relay layer (assuming each neuron receives from 256 groups). This is way too large a number of relays. It is obvious that the inner relay layers must be smaller than 1000 by 1000 groups.
Assuming that each following layer is only .7 smaller in the x and y direction, every two layers of projection would result in a layer with one forth the number of groups. After six layers of projection the final size of the inner layer would be 1/8th by 1/8th the size of the input layer or around 120 by 120 groups. Only a few more layers would be needed for a leftmost input to reach the right most neuron of an inner layer. Thus, we have deduced that the inner layers must be smaller than their source layers. As an alternative, we can actually predict the inner layers must be smaller than outer layers.
That is, one can predict that each receiving relay layer from the sensory input must smaller than its input layer by the fact that, as the axons from one layer turn and pass into the next layer, they will enter the receiving layer at a higher density that they leave a layer. That is, the output density will be related to the average neuron separation in the cortical layer. The entering density would be only dependant on the axon thichness which would tend to be less than the average neuron body separation. A 4 G bits memory chip permits around 4,000 synapses per neuron in groups of 16 mutually inhibiting neurons, 4 G bits memory chip permits around 4,096 synapses per neuron (probably a realist limit for a simulator)) .
SYNAPSE PLASTICITY
Our rule proposed for varying synaptic strengths, assumes that axons only grow and do not shrink back. This goes for both the excitatory and inhibitory synapses. Use of simple Pavlovian learning rule where a synapse increases if the axon and receiving neuron pulse together, results in conditioned reflex behavior by the neuron nets where the correlations computed by the neurons are normalized simply by assuming that the connection strengths from the excitatory neurons to the shared inhibitory neuron are large enough to guarantee that the inhibitory neuron always fires when any of the excitatory neurons fire. The mechanism for the connection strength increase assumes that a pulsing axon tip (containing vesicles) ejects the contents of the vesicle when the axon tip pulses. The vesicle emits its contents through the axon membrane so that the closure of the broken region (penetrated by the vesicle) allows the axon membrane to expand for each vesicle that breaks through, assuming that the volume surrounding the axon tip allows for the expanded tip size. In turn the surrounding glial cells are assumed to allow or not allow this expansion room.
It is possible to relax the axonal modification rule by assuming that a pulsing axon synapse, as vesicles are bursting out, will expand or shrink, not expand or freeze. It is possible to show that normalization can be maintained under this modified rule. The problem with this altered rule is that the resulting nets of neurons will not exhibit any permanent memory properties because connection strengths will be able to go up and down indefinitely.
Unfortunately, no one has ever measured or verified that an axon grows, only if it is pulsing. Searching the internet yields not one article on pulsing neurons and axonal growth in the same article. Obviously, the nets are too small to measure and observe the necessary parameters. There are articles describing axon growth and branching, but never a mention that neuron pulsing was present. However, logic for a conditioned reflex to be possible, requires that the input axon must be pulsing, otherwise growth would occur without a stimulus and would produce memory decay or modification and rediculous learning.
Secondly, the target neuron must pulse, otherwise learning would occur to a response that was not evolked, certainly not correlated to conditioned reflex behavior.
There have been reports that a pulsing axon with the proper pulsing phase with the target neuron pulsing, causes learning, positive or negative. This fact if true, would be of no use since the phase of pulsing of a group of input neurons and a target neuron, would tend to have no phase relations for the most part, and could not be a prime mechanism of synapse learning.
The problem here is that the typical neurologist is not thinking, and is not realizing the inconsistencies of their synapse learning models.
SIMULATOR AND I/O BOARD SPECIFICATIONS
Dr. Swallow will design the basic FPGA and its board. The boards will be manufactured by a printed circuit house. At first 5 boards, then 32 boards will be built and interfaced with a PC. The 32 boards should be available in 6 months. Because a programmable FPGA is employed, the boards will be able to be modified to nets of neurons with any properties at a later data if such constraints as normalized correlations, learning rules, etc. need to be modified. 32 boards will plug into each other, with the board to board connector acting as a wide buss with the data clocking to the next board on a 6 nsec clock cycle. The PC will send data to the first card, which relays the data to the second card, etc. until the last card sends its data back to the PC. The solved neuron states are computed by each board and placed on the buss and received by all the other cards at 48 millisecond frame rate. This neuron state data consists of the frequency of firing (8 bits), the name of the group it resided in (25 bits), and which neuron of the group (7 bits), it is. (We have used the 4 bit value of 16 in our examples of neuron nets, but shall build the simulator to handle groups of 128 mutually inhibiting neurons).
Dr. Swallow also will develop a board to translate a 256 by 256 image into 256K by 256 rod-like sensors with axonal outputs fed into the buss of the multi-board simulator. Similarly, a circuit will be developed to extract frequencies from sound, simulating a linear array of sensor neurons distributed along a frequency sensitive membrane whose output axonal frequencies will be each sensitive to a different frequency of sound. Thirdly, a simple robot will be constructed with electric motors to model a simple animal with 4 legs, 1 arm with 3 fingers, and a rotatable camera. These tasks will require another 3 months, but can be partially done in parallel with the FPGA and board developments.
LITERATURE SEARCH
`
In parallel with the hardware development and its production, a software staff member will write the programs to load and debug the hardware. Because most of the software developments end after 6 months, this staff member will need expertises in the biological and neurological disciplines. His/her next duties then will be to search the literature for neuron to neuron connectivities and to learn more of the embryo developmental interconnect sequences and initial connections. Dr. Swallow will also contribute to these literature search activities and will conduct early experiments to gain familiarity with the operation of the 32 board system.
A third neurologist will search the literature with the intent to gather as much cortical cortical interconnect data from different mammals. He/she will also extract from the literature results of the first 17 weeks of animal postnatal embryo development. .
With the few months of experimentation during the last months of the first year, we should know if we are on the right track with the assumed neuronal net properties.
If needed at the beginning of the second year, the neuron nets will be programmed to model portions of the thalamus and spinal cord, so that the simulator can be extended in performance, if needed, to at least reach the point where the robot will learn to walk and track visual and auditory stimuli on its own. That is, we don’t know how much head start an animal needs such as reflexes before it can begin to perform certain learning tasks as learning to walk. One thing is certain. our simulator will learn only if it is place in an environment with varying stimuli. The human baby upon being born is immediately stimulated by its digestive track. We may need to provide these stimuli, rather than relying totally upon normal varying stimuli provide by its vision and auditory inputs alone.
A 32 board system may be found not enough to reach the performances that we expect. Thus, extending the size of the simulator may be the obvious next step if the learning behaviors of the simulator appear unacceptable. The budget for this project, therefore, shall include optional funds for another set of 32 boards for use during the first part of the second year.
Beginning in the middle of the second year, a plan to evolve the simulator toward more and more human capabilities will be formulated. Only at this point will the project weaknesses and strengths be sufficiently known in order to derive the best next steps. With sufficient success at the beginning, a full 1024 board system will likely be needed to extend the performances to attempt to more closely model human behavior. Also, additional simulators and staff would be needed in order to further speed up the research activities which involve extensive and lengthy research and simulator interactions over long periods of time. That is, the search for human cortical cortical interconnection structures could require considerable efforts in that the evaluation of success or failure often would require long stretches of learning activities by the simulator.
BUDGET (in dollars)
Salary year 1 year2
Dr. Ronald Swallow 120,000 120,000
Dr. Richard Swallow 90,000 90,000
Neurologist 80,000 80,000
Overhead xxxxxx xxxxxx
Facilities (use of home, full basement) 0 0
Utilities (use of home) 0 0
Equipment
Computers (3 Dell desktop units and printer) 6,000 0
Integrated circuits (64 boards, interfaces) 20,000 15,000
Printed circuit board fabrications 3,500 3,500
Robot components 10,000
Office supplies 2,000 2,000
Travel and Conferences 5,000 5,000
Internet (3 lines) 1,000 1,000
_____ _____
337,500 316,500
Schedule
Month 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Design Simulator -------------------------
Build Simulator -----------------------------
Interface Computer ------------------------------
Develope Software -------------------------------------------------------
Design Robot -------------------------------------------------
Build Robot ------------------------------
Study Cortical Connections --------------------------------------------------------------------------------------------
Study Embryo Development -------------------------------------------------------------------------------------------
Study Simulator Behavior -----------------------------------------------------------
Extend Simulator Size to 64 Boards ----------------------------------
Evaluate Progress -----------------
Propose and Seek Additional Funding for Future Projects --------------
SPECIFIC PROJECT ACTIVIITIES
Board Development
Figure 1 shows the necessary architecture of the circuit board. The FPGA on the board is designed to perform many multiplies in parallel at the maximum clock rate of around 6 nano-seconds. Most of the large DDR (Double Data Rate) memory chips containing the synapse connections for the pulsing neurons, are reread every 48 milliseconds. Those connections that change are written every 48 milliseconds. Limiting the memory to 19 chips constrains the boards to modeling 512K neurons each. Keeping the boards small allows maximum communication speeds between the memory chips, the FPGAs, and the boards. These boards model mutually inhibiting neurons. The number of neurons in a group initially will be programmable between 8 and 32. The boards will require four months to design and four months to build.
Board to Board Communication
The boards will plug into each other in groups of 32. Groups of 32 boards will communicate via ribbon cable to each other. All boards will latch all buss signals and therefore extendable to 1024 boards with no additional hardware interfaces.
Computer Interface
A PC interface card will be designed to communicate with the initial simulator 32 board grouping. A wide bit input/output buss will be utilized for maximum communication rate of data transfer.
Robotic component
A camera will be sequentially scanned with a resolution of 256 by 256 by a special board that will format the output as an array of 256 by 256 axons with an output format similar to the 32 boards. A second board will be built able to model the auditory membrane with 2000 frequency sensitive sensors, also with an output axon format similar to interface to the 32 boards. An electronic controller board will be built able to be controlled by axon frequency drivers from the 32 boards. Motor position sensors will be provided to the 32 board system.
Software
Programs will be written in C++, able to be compiled on the PC. An interface language will be developed to all user to define and control the cortical interconnections and initial values.
Literature embryonic and cortical mapping search
As much as possible the Internet will be used to search the literature. Where necessary trips to Universities will be made to locate the latest and most detailed information obtained via the neuro-logical research community.
Simulator Operation
The simulator will be begin to be operated in the 7th month. Connection schemes will be experimented with. Various initial condition structures will be studied.
MANAGEMENT
Dr. Ronald Swallow will be in charge of the project and will deal with any external reporting requirements. The team will work together, communicate on a daily basis, and will need no weekly, monthly, etc meetings. Dr. Richard Swallow, a biologist and computer scientist, will handle the software efforts and later the literature search efforts. A scientist will be hired to do the literature search full time.
Monthly progress reports will be written for documentation and communication purposes. Conference papers will be written when and if appropriate. A report will be produced at the end of the two year project, documenting the research effort of the simulation project and its findings.
The services of an accountant firm will be used to oversee the financial activities of the project. Hardware will be ordered, soldered, and assembled by Dr. Ronald Swallow and Dr. Richard Swallow.
BIOGRAPHICAL SKETCH
of
Dr. Ronald J. Swallow
Dr. Swallow specifically aimed his education and professional capabilities to work on brain models. After having sought and discovered the neuron net model capable of normalized cross-correlations as an assistant Professor while teaching at Purdue (1965 to 1969), and due to the infancy of hardware technologies, he has continued for the last 40 years to develop, on his own time, a cortex hardware simulator while waiting for hardware performance and costs to drop. During this time, during the later 40 years, he has built his expertises in hardware design, by specializing in the designing of super image generators for use in automated instruction, flight simulation and gaming (his designs exceed the performance the best game units such as WII and Box 360. Having specialized in high performance hardware design, he now is returning to brain modeling for full employment and the construction and operation of the hardware simulator which he has been developing on his own time and is the key ingredient of this project.
EDUCATION
University of Illinois, Urbana, Illinois Engineering Physics BS 1958
University of Illinois, Urbana, Illinois Electrical Engineering MS 1962
University of Illinois, Urbana, Illinois Biophysics PhD 1964
PROFESSIONAL WORK EXPERIENCE
July 2008 – Present
Full-time consultant to L-3 navagation and space systems, Senior Staff Engr, Budd Lake, NJ.
Translated schematic designs (for a solid state gyro) into VHDL code. Simulated schematic and VHDL code to verify VHDL code.
July 2000 – July 2008
HONEYWELL CORPORATION, Senior Staff, Teterboro, NJ
Designed and debugged Xilinx FPGA based display board for Apache Helicopter. Designed FPGA to do raster conversion and image scaling. Extensive study of mpeg2 systems. Significant contribution to the FPGA design of B1 and B52 bomber displays.
July 1998 - July 2000, (from July 2000 to present, operations during weekends and evenings)
MEGAPOLYGON CORPORATION, Sole Owner, Emmaus, PA
Obtained license to build all of Tellurian products (image generator products). Obtained all R&D results from Tellurian. Continued to search for a real-time image generator capable of at least two-million polygon throughput at 30 Hz frame rate. Found the architecture capable of 10-million 3D polygon throughput at 30 Hz frame rate employing Xilinx Virtex II pro (and above) family of FPGAs. Presently detailing the 10-million polygon design, which employs a Z-buffer visible surface algorithm fully anti-aliased. Completed the color sequential helmet with regard to its plastic enclosure design and the development of a six degree of freedom mechanical head tracker, which also produces a vertical lift to reduce the helmet's weight on the user's head. Developed game and other software for gaming market.
1988 - July 1998
TELLURIAN, INC., CEO and V.P. Engineering, Mahwah, NJ
Head of R&D department with six technical staff members. Completed the production phase for the AT-100 image generator product (capable of 4K polygons in real-time). Designed, routed, and prepared for production the AT-200 image generator product (capable of 8K polygons in real-time). Designed, prototyped, routed, and prepared for production the Eagle image generator product capable of 30,000 polygons in real-time at 1000 by 2000 lines resolution employing six ACTEL 8K-gate family of FPGAs. Searched and found image generator hardware architectures capable of texture, semi-translucency, smooth shading, and anti-aliasing employing ATT ORCA2 family of FPGAs with 200K polygon throughput. Searched for, but failed to find a 2-million polygon throughput design. Managed the development of a 1000 by 2000 line color sequential helmet. Redesigned the Eagle product to drive a color sequential helmet. Directed staff to perform software tasks to aid in the schematic capture and routing of the FPGA chips and boards.
1985 - 1988
POLYEDGE TECHNOLOGY CORPORATION, Founder and President, Fair Lawn, NJ
Designed an image generator using Fujitzu ECL-based ASICs. Designed the first wire wrap prototype of the AT-100 image generator product, capable of 4K polygons in real-time.
1982 - 1985
TRILLIUM CORPORATION, Founder, CEO and Director of R&D, Glen Rock, NJ
Designed and built the 1100 series ECL-based real-time image generator system consisting of over 2000 SSI chips. Directed operations and allocated resources to take the corporation from the R&D stage, through beta test, and into the production with five systems now in the field within a budget of five million dollars.
1978-1981
ADVANCED TECHNOLOGY SYSTEMS, Principal Scientist, Fair Lawn, NJ
Upon satisfactory completion of a 32,000-edge system, succeeded in winning the subcontract for visual simulation portion of the F-18 under contract from Hughes Aircraft Corporation. Redesigned the CIG system to derive the video in order to project inside of forty-foot diameter movie screen at 8,000-line resolution and three channels. Developed an excellent staff, proficient in this design technique and architecture. Developed a neuron model which explains how the human brain does normalized cross-correlations, exhibits a curiosity drive, and has reduced memory capabilities as it ages.
1969-1978
HUMAN RESOURCES RESEARCH ORGANIZATION, Senior Scientist, Alexandria, VA
Performed research seeking a fully automated instruction system solution. Studied speech recognition, telegraphic terminal designs and image generator designs for logically generated images. Designed computer-controlled slide and movie projectors and speech synthesis systems for logically-generated speech and sound. Gained extensive experience in SSI and MSI circuit designs and in special hardware architectures, which permit extremely high performance and low cost solutions to the computing system requirements of an automated instruction. Designed, produced, and delivered a 12-terminal, computer-aided instruction system to the United States Post Office incorporating the real-time simulation of envelopes with varying handwriting, printing, stamps, envelope orientations, and motions which occur in the operation of the post office. While under contract to Advanced Technology Systems, designed the first 32,000 edge real time computer image generator system. Before completion of the project, the Advances Technology Systems was joined to the project officially.
1965-1969
PURDUE UNIVERSITY, E.E. Department as Assistant Professor, West Lafayette, IN
Taught linear and nonlinear circuit theory, electrical magnetic theory, computer design, artificial intelligence, and control system theory. Conducted research on brain models, speech recognition, and automated instruction systems (both hardware and software designs). Found a neuron model that was capable of normalized cross-correlations.
1964-1965
UNIVERSITY OF ILLINOIS, Biological Computer Laboratory, Research Associate, Urbana, Ill
Conducted research on self-organizing systems and brain models.
During graduate work, conducted research on radar hardware and theory and on brain models as Research Assistant at the Coordinated Systems Laboratory (U of I).
Publications
Modeling and Simulation, Volume 11, Part 3, ”A Model for Normalized Cross-Correlations at the Neuronal Level”, May, 1980, Proceedings of the Eleventh Annual Pittsburgh Conference