Data processing: artificial intelligence – Neural network – Structure
Reexamination Certificate
1999-06-14
2003-01-14
Davis, George B. (Department: 2121)
Data processing: artificial intelligence
Neural network
Structure
C706S034000, C706S041000, C706S025000
Reexamination Certificate
active
06507828
ABSTRACT:
FIELD OF THE INVENTION
This invention relates generally to neural networks and more particularly to circuits which model the behavior of biological neurons.
BACKGROUND OF THE INVENTION
As is known in the art, there exists a class of networks referred to as neural networks which model the behavior of certain human functions. Electronic neural networks have been used to implement mathematical or engineering abstractions of biological neurons. Circuits emulating biological neurons are typically implemented using digital circuits that operate up to a million times faster than actual neurons or with software which simulates the behavior of a biological neuron. One problem with the digital circuit approach, however is that it does not utilize life-like principles of neural computation. Furthermore, a biological nervous system contains thousands or millions of interconnected neurons and thus the complexity of a biological nervous system results in a complex digital circuit.
Similarly, given the complexity of the biological systems, software simulations can take many hours or days even using presently available state-of the-art processing systems. Thus software systems are not appropriate for use in applications which require real time or close to real time performance from such systems
An electronic circuit that emulates the analog behavior of actual biological neurons, on the other hand, can perform simulations in real time. Thus, to overcome the above limitations with systems implemented using digital circuits or software, electronic circuit neural networks which use principles of neural computation which are more life-like than the digital circuit or software approaches have been developed.
This type of neural network interacts with real-world events in a manner which is the same as or similar to biological nervous systems and can be utilized in a variety of systems including but not limited to electronic and electromechanical systems, such as artificial vision devices and robotic arms. Such neural networks can also be used as research tools to better understand how biological neural networks communicate and learn.
Much of the effort directed toward producing electronic implementations of biological neurons have focused on emulating the input-output functional characteristics of the neuron, essentially treating the neuron as an abstracted black box. These implementations focus on circuits and techniques for generating an action potential in an attempt to simulate the actions neurons take to communicate with one another. One problem with past approaches, however, is that such approaches fail to properly take into account or model the means which actually produces the action potential in a biological neuron.
Some prior art techniques have produced analog integrated circuits that mimic the functional characteristics of real neuron cells, by isomorphically emulating the membrane conductances within an actual neuron cell body. Thus, one problem with prior art approaches is that they fail to include circuitry for the synapse through which neurons communicate and/or the prior art approaches fail to include circuitry for the dendrite which is the connection between the synapse and neuron cell body. Prior art systems also fail to include effective circuitry to implement the adaptation or learning functions of real neurons.
In one particular prior art technique, a model of one type of synapse referred to as a Hebbian Synapse was provided. In a Hebbian Synapse, stimulation of the pre-synaptic neuron causes the release of neurotransmitters from an axon terminal. These transmitters include amino acid glutamate and bind to corresponding receptors on the post-synaptic membrane causing ion channels to open up through these receptors. Glutamate binds to three types of receptors: N-methyl-D-aspartate (NMDA), quisqualate, and kainate. Long Term Potentiation (LTP) and Long Term Depression (LTD) are both mediated by the NMDA receptors, which carry primarily Ca
2+
currents. The other receptors, termed the non-NMDA receptors, carry the remainder of the synaptic current, consisting mainly of Na
+
, with negligible Ca
2+
content. These receptors are located on a spine head connected to the dendritic shaft.
The spine head can be represented or modeled as an electrical circuit which includes four parallel circuit legs. The total synaptic current consists of the sum of the NMDA and non-NMDA currents. There also exists a small leakage conductance and the capacitance that represents the membrane capacitance of the spine head. The current through the non-NMDA channels in response to a pre-synaptic stimulus is given by an alpha function:
I
non
=(
E
non
−V
head
)&kgr;
g
p
te
t/p1
Equation 1
in which K=e/t
p
, e is the base of the natural logarithm, t
p
=1.5 ms, g
p
=0.5 nS, and E′
non
=0. It should be noted that non-NMDA receptor conductance is purely ligand (neurotransmitter) dependent. The NMDA conductance, on the other hand, is both ligand dependent, due to the binding of neurotransmitters released from the pre-synaptic neuron, and dependent on the spine head membrane voltage. The current through the NMDA receptors is given by:
I
NMDA
(
t
)
=
(
E
NMDA
-
V
head
)
g
n
(
ⅇ
1
t1
-
ⅇ
1
t1
)
1
+
n
[
Mg
2
+
]
ⅇ
-
tVhead
Equation
2
Where &tgr;
1
=80 ms, &tgr;
2
=0.67 ms, &eegr;=0.33 nM
−1
, &ggr;=0.06 mV
−1
, E
NMDA
=0 and g
n
=0.2 nS. The voltage dependence of the NMDA receptor arises from the fact that the receptors are inhibited by magnesium ions Mg
2+
having a binding rate constant which is dependent upon the spine head membrane voltage. Near the resting membrane potential, the NMDA receptor channels are almost completely blocked by the Mg
2+
ions, and thus little current flows. As the spine head membrane becomes partially depolarized, the Mg
2+
ions become dislodged and more NMDA current flows.
The post-synaptic flow of Ca
2+
ions through the NMDA receptor channels is crucial for the induction of LTP and LTD. Upon entering the dendritic spine, the Ca
2+
ions trigger a series of events that lead to the induction and maintenance of LTP or LTD. The precise mechanisms, however, are not well understood. One theory is that a second messenger, such as nitric oxide, is activated by the Ca
2+
ions and certain calcium dependent proteins and then diffuses back to the pre-synaptic terminal, stimulating the release of more glutamate. Thus, this retrograde messenger operates as a positive feedback mechanism. This model is limited, however, in that it only explains LTP.
Another model suggests that rather than affecting pre-synaptic neurotransmitter release, the induction of LTP or LTD modulates the post-synaptic conductance of the non-NMDA receptor channels, which carry the bulk of synaptic current. This model also relies on the influx of Ca
2+
ions into the dendritic spines through the NMDA channel. During high frequency simulation, the Ca
2+
ions reach high concentrations in the compartmentalized spine head and preferentially activates a protein kinase. During low frequency simulation, lower concentrations are reached and a protein phosphatase is released. Both proteins act on a common phosphoprotein, which triggers LTP or LTD by modulating the non-NMDA receptor channel conductance.
It would, therefore, be desirable to provide an analog very large scale integrated (VLSI) circuit implementation of a biological neuron which includes circuitry for the synapse through which neurons communicate, and the dendrite which serves as the connection between the synapse and neuron cell body. It would also be desirable to provide a circuit which represents the means which actually produce the action potential in a biological neuron. It would also be desireable to include in an analog VLSI implementation of a biological neuron, circuitry for adaptation or learning.
SUMMARY OF THE INVENTION
In accordance with the presen
Leonard Jason
Poon Chi-Sang
Daly, Crowley & Mofford LLP
Davis George B.
LandOfFree
Neuron circuit and related techniques does not yet have a rating. At this time, there are no reviews or comments for this patent.
If you have personal experience with Neuron circuit and related techniques, we encourage you to share that experience with our LandOfFree.com community. Your opinion is very important and Neuron circuit and related techniques will most certainly appreciate the feedback.
Profile ID: LFUS-PAI-O-3071589