Data processing: artificial intelligence – Neural network – Learning task
Reexamination Certificate
2001-05-01
2004-03-16
Davis, George B. (Department: 2121)
Data processing: artificial intelligence
Neural network
Learning task
Reexamination Certificate
active
06708159
ABSTRACT:
BACKGROUND OF THE INVENTION
1. Field of the Invention
The present invention is directed to an automaton for use as a logical building block, and particularly to an automaton suitable for use in a neural architecture.
2. Description of the Prior Art
Modern computing owes much of its logical structure to John Von Neumann, who incidentally hypothesized that circuits as complex as the human nervous system would never be realized. The logic structure that does pervade the computing world relies on one of two models. The first model is a simple all or nothing state model. Beginning with vacuum tubes and extending to the present transistors, the element is either on or off. Ones and zeros may be assigned to either state as needed, being consistent within the circuit. Multiple elements may be assembled to form gates. The gate reacts to input depending on its nature and provides an output of either on or off. Typical gates include AND, OR, XOR, NOT, or the like. While helpful for circuit design, this sort of logic is not conducive to the creation of automata capable of inferential decision making.
The second model was derived from the study of nerve cells and relies on threshold logic. Threshold logic implies that there is a gate, and when inputs to that gate meet or exceed the threshold, the gate is triggered and an output is generated. While functional to model simply the action potential behavior of the nerve cell, these models are not capable of emulating completely biologic nerve cell behavior. Further, a single threshold logic gate cannot implement an XOR function. The lack of synthesis procedures limits this logic to adapt in modern logic design procedures. Attempts to use either of these models to create useful automata based on neuronal signaling have failed.
These failures extend past the creation of useful automata. Threshold logic, as applied to neural networks, led to neural net circuits whose dynamics are neither controllable nor observable due to the analog nature of the components. Specifically, because analog systems are flow through systems, an observer cannot tell exactly what it is doing something and when it would do such. Threshold gates further fail to model with any success inhibitory mechanisms in human nerve cells.
To date modern research in neurobiology has contributed to further understanding nerve cell signaling and uncovering vast complexities in synaptic organizations. To this effect, elaborate structures in chemical synaptic connections with extensive contacts between cells have also been observed. The perception that an axon terminating onto a single synapse has consequently been revised. To date, there has not been a logic circuit based on the electrochemical communication between cells.
SUMMARY OF THE INVENTION
An automaton may be constructed digitally, modeled on the ability of nerve cells to operate at a plurality of discrete electrochemical states. The automaton of the present invention may comprise a plurality of inputs. To make the automaton more adaptable to represent synaptic arrangements more complex than the ones modeled by threshold logic, these inputs may each be weighted independently of the others, or weights may be assigned to combinations of inputs, also reflecting the biological phenomena of coupling between presynaptic terminals. These weighted inputs are fed into a state computing unit. An output is generated by the state computing unit. The output is simultaneously fed back to a weight computing unit together with the inputs of the automaton. The weight computing unit in turn controls dynamically the weights assigned to each input or combination of inputs. A digital clock drives the inputs, the state computing unit, and the weight computing unit.
The output of the automaton is thus a digital value having more than two state levels that more closely reflects the electrochemical communication between biologic neurons. With this automaton as a building block, more advanced neural architectures maybe constructed.
REFERENCES:
patent: 5226092 (1993-07-01), Chen
patent: 5517597 (1996-05-01), Aparicio, IV et al.
patent: 2002/0184174 (2002-12-01), Kadri
patent: 0560595 (1993-09-01), None
patent: 0834817 (1998-04-01), None
Chen et al, “On Neural-Network Implementations of K-Nearest Neighbor Pattern Classifiers”, IEEE Transactions on Circuits and Systems, Jul. 1997.
Coats & Bennett P.L.L.C.
Davis George B.
LandOfFree
Finite-state automaton modeling biologic neuron does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Finite-state automaton modeling biologic neuron, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Finite-state automaton modeling biologic neuron will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3212632