CN110097182A - Circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control - Google Patents
Circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control Download PDFInfo
- Publication number
- CN110097182A CN110097182A CN201910283951.7A CN201910283951A CN110097182A CN 110097182 A CN110097182 A CN 110097182A CN 201910283951 A CN201910283951 A CN 201910283951A CN 110097182 A CN110097182 A CN 110097182A
- Authority
- CN
- China
- Prior art keywords
- circuit
- operational amplifier
- neural network
- input terminal
- resistance
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 210000002569 neuron Anatomy 0.000 title claims abstract description 93
- 230000004913 activation Effects 0.000 title claims abstract description 83
- 238000003062 neural network model Methods 0.000 title claims abstract description 17
- 238000013528 artificial neural network Methods 0.000 claims abstract description 66
- 210000005036 nerve Anatomy 0.000 claims description 9
- 239000003990 capacitor Substances 0.000 claims description 8
- 230000004044 response Effects 0.000 claims description 8
- 230000000694 effects Effects 0.000 claims description 7
- 230000005611 electricity Effects 0.000 claims description 7
- 210000000225 synapse Anatomy 0.000 claims description 7
- 230000005674 electromagnetic induction Effects 0.000 claims description 5
- 230000003213 activating effect Effects 0.000 claims description 3
- 238000004088 simulation Methods 0.000 abstract description 10
- 238000000034 method Methods 0.000 abstract description 4
- 210000004556 brain Anatomy 0.000 abstract description 3
- 238000011160 research Methods 0.000 abstract description 2
- 238000013473 artificial intelligence Methods 0.000 abstract 1
- 230000006399 behavior Effects 0.000 description 16
- 238000002474 experimental method Methods 0.000 description 10
- 238000012795 verification Methods 0.000 description 7
- 238000004458 analytical method Methods 0.000 description 4
- 230000001537 neural effect Effects 0.000 description 4
- 230000008859 change Effects 0.000 description 3
- 238000013461 design Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 230000004927 fusion Effects 0.000 description 3
- 230000000638 stimulation Effects 0.000 description 3
- 230000003993 interaction Effects 0.000 description 2
- 230000008555 neuronal activation Effects 0.000 description 2
- 238000005424 photoluminescence Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
- 238000013178 mathematical model Methods 0.000 description 1
- 210000004218 nerve net Anatomy 0.000 description 1
- 230000003071 parasitic effect Effects 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/044—Recurrent networks, e.g. Hopfield networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D30/00—Reducing energy consumption in communication networks
- Y02D30/70—Reducing energy consumption in communication networks in wireless communication networks
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Health & Medical Sciences (AREA)
- Life Sciences & Earth Sciences (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Feedback Control In General (AREA)
Abstract
The present invention provides a kind of three-dimensional Hopfield neural network model realization circuit with neuron activation gradient λ control, realizes circuit and three-dimensional Hopfield neural network main circuit including the negative output hyperbolic tangent function based on activation gradient.A kind of three-dimensional Hopfield neural network model with neuron activation gradient λ control provided by the invention realizes circuit, based on conventional three-dimensional Hopfield neural network, control amount λ of the neuron activation gradient as model is introduced in hyperbolic tangent function, constitute the activation primitive tanh (λ x) based on neuron activation gradient, module of the activation primitive as neural network, the dynamics Controlling to three-dimensional Hopfield neural network may be implemented by neuron activation gradient λ, this control method is simple to operation, the reactiveness of fabulous simulation human brain, to Neuromorphic circuit, the development and research of electronic nerve cell and artificial intelligence have preferable application.
Description
Technical field
The present invention relates to neural network control technique fields, more particularly to a kind of with the three of neuron activation gradient λ control
It ties up Hopfield neural network model and realizes circuit.
Background technique
The Hopfield neural network (HNN) being made of neuron is that one in artificial neural network is extremely important
Model.In in the past few years, realize that control dynamic (dynamical) to HNN is existing by changing the synapse weight between different neurons
A large amount of reports.Hyperbolic functions are a kind of nonlinear functions, it can be used as the activation primitive imictron electrical activity of neuron
Behavior can change the slope of hyperbolic functions, the i.e. response speed of electrical activity of neurons by changing activation gradient λ, with nerve
Gradient λ is activated to realize that the control to Hopfield neural network is but never studied.Therefore, this neural network spy is further studied
It is not the method for using the Hopfield neural network dynamics of neuron activation gradient λ control, is it is necessary to and meaningful.
Summary of the invention
The technical problems to be solved by the present invention are: in order to overcome the shortcomings in the prior art, the present invention provides a kind of use
The three-dimensional Hopfield neural network model of neuron activation gradient λ control realizes circuit.
The present invention solves its technical problem technical solution to be taken: a kind of with the three of neuron activation gradient λ control
Tie up Hopfield neural network model realize circuit, including based on activation gradient negative output hyperbolic tangent function realize circuit and
Three-dimensional Hopfield neural network main circuit.The present invention is based on conventional three-dimensional Hopfield neural networks, in hyperbolic tangent function
The middle control amount λ for introducing neuron activation gradient as model, constitutes the activation primitive-of the negative output based on neuron activation gradient
Tanh (λ x), module of the activation primitive as neural network, it has an impact the dynamics of entire neural network, so passing through
The dynamics Controlling to three-dimensional Hopfield neural network may be implemented in neuron activation gradient λ.
The three-dimensional Hopfield neural network model of one routine can indicate are as follows:
In conventional neural network model (1), the representative value of a and b, i.e. a=0.7, b=-2 are assigned.In activation primitive
Control amount λ of the neuron activation gradient as model is introduced in tanh (x), it can indicate that the response after neuron is stimulated is fast
Degree.Thus neuron activation gradient is constituted based on the activation primitive tanh (λ x) under neuron activation gradient control amount, activation primitive
As the module of neural network, it affects the dynamic behavior of entire neural network.
On the basis of a conventional three-dimensional Hopfield neural network model, a neuron activation gradient function λ is introduced
As control variable, which is indicated with first-order ordinary differential equation system are as follows:
In formula, x1、x2And x3Respectively three state variables of neuron；Parameter a is connection third neuron and first
The synapse weight of a neuron, parameter b are the synapse weights for connecting second neuron and first neuron, are usually set respectively
It is set to a=0.7, b=-2；Control variable λ is the activation gradient of neuron, indicates the sound of the electrical activity of neurons under electromagnetic induction
Answer speed；It is worth noting that, model (1) has a zero balancing in the zone of reasonableness of control variable λ (0.7 < λ < 1.5)
Point and two non-zero equalization points, they show different states with the increase of λ, and complicated dynamic behavior is presented.
Three equations in formula (2), are respectively adopted integrating channel one, integrating channel two and integrating channel three to realize, and three
Main circuit of a integrating channel as three-dimensional Hopfield neural network assigns the representative value of a and b, i.e. a=0.7, b=-2, root
According to the electrology characteristic of Kirchhoff's laws of electric circuit and circuit components, then the corresponding circuit equation of formula (2) can indicate are as follows:
In formula (3), v1、v2And v3It is three Circuit variables, it corresponds in three-dimensional Hopfield nerve network system
x1、x2And x3, indicate the film potential of three neurons in neural network；Nonlinear function can be used as the activation letter of neural network
Number, can with a kind of dullness can micro- bounded function representation, hyperbolic tangent function is exactly the activation primitive in this system, it indicates refreshing
Through member by the state change after environmental stimuli；Control variable λ is the activation gradient of neuron, indicates neural under electromagnetic induction
The response speed of first electrical activity.As λ > 1, indicate neuron by being swift in response after outside stimulus；As λ=1, mind is indicated
Through member by the normal reaction after outside stimulus；As λ < 1, indicate that neuron is slow by the reaction after outside stimulus.
va、vb、vc、–vaWith-vbIt is the input terminal inside neural network, they are connected from different integrating channels, also conduct
The feedback port of neural network；v1、v2And v3It is the internal output terminal of neural network and the external output end under outside stimulus,
They can be attached the different channels of oscillograph to observe.If there is some neuron in neural network to receive outside
Stimulation, i.e., by faradic interference, then will increase a resistance R before the integrating channel that some neuron circuit is realized
Electric current is transmitted, some function/value of the input terminal connection outside stimulus of resistance R, this is the external input terminals of neural network.Such as
First neuron receives the stimulus of direct current that amplitude is 1V, then a R will be met before the integrating channel of first neuron0
The resistance of=10k Ω, the DC power supply v that one input of connection is 1V before resistanceDC, output is v1It is constant；For another example, second
It is 2V that neuron, which receives amplitude, and frequency is the exchange electro photoluminescence of 60Hz, then by before the integrating channel of second neuron
Meet a R0The resistance of=10k Ω, one input of connection is 2V before resistance, and frequency is the alternating-current voltage source v of 60HzAC, output
For v2It is constant.When neural network is not by outside stimulus, they are by each neuron neuron interaction, i.e., the three-dimensional studied now
Hopfield neural network can control this network by neuron activation gradient λ, so that various forms of dynamic behaviors are generated,
Simulate the state response of human brain.Dynamic behavior after outside stimulus is added is next emphasis that we study.
The realization circuit of integrating channel one includes the input terminal v inside neural networka、vb、vc, operational amplifier U1And U2,
And activation gradient tanh (λ x) function realizes circuit Tg1, input terminal v inside neural networka、vb、vcSeries resistance R respectively1、
R2And R3After be connected to operational amplifier U1Inverting input terminal；Operational amplifier U1Inverting input terminal and output end between simultaneously
The capacitor C of one, connection " 10nF " and the resistance R of one " 10k Ω ".Because the tanh module in nerve network circuit should be negative
The hyperbolic tangent function of output, so devising the hyperbolic tangent function with negative output in circuit design realizes circuit.Fortune
Calculate amplifier U1The voltage v of output end output1Circuit Tg is realized by activation gradient tanh (λ x) function1After obtain output voltage
va；Output voltage vaOperational amplifier U is connected to after concatenating the resistance R of one " 10k Ω "2Inverting input terminal；Operational amplifier
U2Inverting input terminal and output end between it is in parallel one " 10k Ω " resistance R, operational amplifier U2Output end output-va,
Output-vaAn input signal of circuit is realized as integrating channel two；Operational amplifier U1And U2Non-inverting input terminal connect
Ground；
The realization circuit of integrating channel two includes the input terminal-v inside neural networka、vb, operational amplifier U3And U4, with
And activation gradient tanh (λ x) function realizes circuit Tg2, input terminal-vaAnd vbSeries resistance R respectively4And R5It is followed by operation amplifier
Device U3Inverting input terminal；Operational amplifier U3Inverting input terminal and output end between it is in parallel one " 10nF " capacitor and one
The resistance R of a " 10k Ω "；Operational amplifier U3The voltage v of output end output2Electricity is realized by activation gradient tanh (λ x) function
Road Tg2After obtain output voltage vb；Output voltage vbThe resistance R of one " 10k Ω " is concatenated in operational amplifier U4Anti-phase input
End；Operational amplifier U4Inverting input terminal and output end between it is in parallel one " 10k Ω " resistance R, operational amplifier U4It is defeated
Outlet obtains output-vb, output-vbAn input signal of circuit is realized as integrating channel three；Operational amplifier U3And U4's
Non-inverting input terminal is grounded；
The realization circuit of integrating channel three includes the input terminal v inside neural networka、-vb、vc, operational amplifier U5, with
And activation gradient tanh (λ x) function realizes circuit Tg3, input terminal va、–vbAnd vcSeries resistance R respectively6、R7And R8It is followed by fortune
Calculate amplifier U5Inverting input terminal；Operational amplifier U5Inverting input terminal and output end between it is in parallel one " 10nF " electricity
Hold the resistance R of C and one " 10k Ω "；Operational amplifier U5The voltage v of output end output3By activating gradient tanh (λ x) function
Realize circuit Tg3After obtain output voltage vc；Output voltage vcIntegrating channel one, which is fed back to, as an input signal realizes electricity
Road；Operational amplifier U5Non-inverting input terminal connect " ground ".
Gradient tanh (λ x) function is activated to realize that circuit includes: integrator, triode T1And T2, DC voltage source E and direct current
Current source I0, wherein integrator includes operational amplifier UoAnd Ui, specific connection type are as follows: input terminal viConnect " a 10k
The resistance R of Ω " is followed by operational amplifier UiInverting input terminal；Operational amplifier UiInverting input terminal and output end between
One changeable resistance R of parallel connectionF；Operational amplifier UiOutput end access triode T1Base stage；Triode T1Collector
It is divided into two-way, the resistance R all the way through one " 10k Ω " is connected to operational amplifier UoNoninverting input, another way is through one
The resistance R of " 10k Ω "CIt is connected to DC voltage source E；Triode T2Emitter and triode T1Emitter be connected, and it is same
When meet DC current source I0, the current value of DC current source is " 1.19mA " in the present embodiment.Triode T2Collector also connect one
A resistance RCIt is connected on DC voltage source E, the other end laterally meets one " 10k Ω " resistance R to amplifier UoReversed input
End；Triode T2Base earth, then laterally meet one " 10k Ω " resistance R to amplifier UoNoninverting input.Operation amplifier
Device UoAnti-phase input terminate one " 10k Ω " resistance R to output end vo。
The DC voltage source E is fixed 15V DC power supply, the resistance R of control neural networkFIt, can for adjustable resistance
Tune range is 500 Ω -900 Ω, can directly control the dynamic behavior of neural network.
Neuron activation gradient λ can reflect the response speed of the electrical activity of neurons under electromagnetic induction stimulation.Base of the present invention
One is introduced in a three-dimensional Hopfield neural network model in order to preferably realize the control to Hopfield neural network
A neuron activation gradient function λ is as control variable.Therefore the invention proposes it is a kind of with neuron activation gradient λ control
Hopfield neural network.Using neuronal activation gradient λ as adjustable control parameter, by stability analysis based on equalization point,
Numerical analysis and hardware experiments verifying to mathematical model, have studied Hopfield neural network and are controlled by neuronal activation gradient λ
The dynamic behavior of system.The result shows that occurring complicated dynamic behavior in HNN model.To the neural network system of proposition
System is designed and experimental verification, and discovery test result and Numerical Simulation Results coincide preferably.
The beneficial effects of the present invention are: a kind of three-dimensional Hopfield controlled with neuron activation gradient λ provided by the invention
Neural network model realizes circuit, realizes a kind of Hopfield neural network equivalent circuit with advanced dynamic behavior.
The realization circuit structure is clear, and component used can simply be sought, and is easy to theory analysis and circuit integration.The circuit neuron activation
Gradient λ control, only need to change a variable resistance, that is, can produce complicated dynamic behavior, to artificial in engineer application
Neural network research has biggish value.
Detailed description of the invention
Present invention will be further explained below with reference to the attached drawings and examples.
Fig. 1 is the three-dimensional Hopfield nerve network system model realization circuit with neuron activation gradient λ control；
Fig. 2 is that the hyperbolic functions tanh (λ x) based on neuron activation gradient λ realizes circuit；
Fig. 3 is x when adjusting neuron activation gradient λ=0.931–x3MATLAB numerical simulation phase rail figure and experiment in plane
Verification result；
Fig. 4 is x when adjusting neuron activation gradient λ=11–x3MATLAB numerical simulation phase rail figure and experiment in plane are tested
Demonstrate,prove result；
Fig. 5 is x when adjusting neuron activation gradient λ=1.11–x3MATLAB numerical simulation phase rail figure and experiment in plane
Verification result；
Fig. 6 is x when adjusting neuron activation gradient λ=1.21–x3MATLAB numerical simulation phase rail figure and experiment in plane
Verification result；
Fig. 7 is x when adjusting neuron activation gradient λ=1.31–x3MATLAB numerical simulation phase rail figure and experiment in plane
Verification result；
Fig. 8 is x when adjusting neuron activation gradient λ=1.51–x3MATLAB numerical simulation phase rail figure and experiment in plane
Verification result；
Specific embodiment
Presently in connection with attached drawing, the present invention is described in detail.This figure is simplified schematic diagram, is only illustrated in a schematic way
Basic structure of the invention, therefore it only shows the composition relevant to the invention.
Such as Fig. 1 and Fig. 2, the circuit includes: that the hyperbolic functions tanh (λ x) based on activation gradient λ realizes 2 He of circuit diagram
Three-dimensional Hopfield neural fusion circuit diagram 1；It is three-dimensional that Fig. 2 hyperbolic functions circuit tanh (λ x) is realized that circuit introduces
In Hopfield neural fusion circuit, the novel Hopfield nerve net controlled with neuron activation gradient λ is constituted
Network verifies circuit, and after being sequentially connected such as identical port each in Fig. 1 and Fig. 2, the non-inverting input terminal of operational amplifier connects " ground ", by
This can show advanced dynamic behavior.
Hyperbolic functions circuit based on activation gradient includes: integrator, triode, DC voltage source and DC current source
Deng.Specific connection type are as follows: input terminal " vi" series connection one " 10k Ω " resistance be followed by operational amplifier UiAnti-phase input
End；Operational amplifier UiInverting input terminal and output end between one changeable R of parallel connectionFResistance；UiOutput end access three
Pole pipe T1Base stage；Triode T1Collector laterally connect the resistance of one " 10k Ω ", then meet the resistance R of one " 10k Ω "C；
RCThe DC voltage source E of one small magnitude of another termination of resistance, preferred amplitude is " 15V " in the present embodiment.Triode T2Hair
Emitter-base bandgap grading and triode T1Emitter be connected, and at the same time meeting DC current source I0, its value is " 1.19mA ".Triode T2's
Collector also meets a RCResistance is connected on DC voltage source, and the other end laterally meets one " 10k Ω " resistance R to amplifier U0's
Reverse input end；Triode T2Base earth, then laterally meet one " 10k Ω " resistance R to amplifier U0Noninverting input.
Operational amplifier U0Anti-phase input terminate one " 10k Ω " resistance R to output end vo。
Three-dimensional Hopfield nerve network system model realization circuit includes that integrating channel one, integrating channel two and integral are logical
Road three.
va、vb、vc、–vaWith-vbIt is the input terminal inside neural network, they are connected from different integrating channels, also conduct
The feedback port of neural network；v1、v2And v3It is the internal output terminal of neural network and the external output end under outside stimulus,
They can be attached the different channels of oscillograph to observe.If there is some neuron in neural network to receive outside
Stimulation, i.e., by faradic interference, then will increase a resistance R before the integrating channel that some neuron circuit is realized
Electric current is transmitted, some function/value of the input terminal connection outside stimulus of resistance R, this is the external input terminals of neural network.Such as
First neuron receives the stimulus of direct current that amplitude is 1V, then a R will be met before the integrating channel of first neuron0
The resistance of=10k Ω, the DC power supply v that one input of connection is 1V before resistanceDC, output is v1It is constant；For another example, second
It is 2V that neuron, which receives amplitude, and frequency is the exchange electro photoluminescence of 60Hz, then by before the integrating channel of second neuron
Meet a R0The resistance of=10k Ω, one input of connection is 2V before resistance, and frequency is the alternating-current voltage source v of 60HzAC, output
For v2It is constant.When neural network is not by outside stimulus, they are by each neuron neuron interaction, i.e., the three-dimensional studied now
Hopfield neural network can control this network by neuron activation gradient λ, so that various forms of dynamic behaviors are generated,
Simulate the state response of human brain.Dynamic behavior after outside stimulus is added is next emphasis that we study.
The realization circuit of integrating channel one includes the input terminal v inside neural networka、vb、vc, operational amplifier U1And U2,
And activation gradient tanh (λ x) function realizes circuit Tg1, input terminal v inside neural networka、vb、vcSeries resistance R respectively1、
R2And R3After be connected to operational amplifier U1Inverting input terminal；Operational amplifier U1Inverting input terminal and output end between simultaneously
The capacitor C of one, connection " 10nF " and the resistance R of one " 10k Ω ".Because the tanh module in nerve network circuit should be negative
The hyperbolic tangent function of output, so devising the hyperbolic tangent function with negative output in circuit design realizes circuit.Fortune
Calculate amplifier U1The voltage v of output end output1Circuit Tg is realized by activation gradient tanh (λ x) function1After obtain output voltage
va；Output voltage vaOperational amplifier U is connected to after concatenating the resistance R of one " 10k Ω "2Inverting input terminal；Operational amplifier
U2Inverting input terminal and output end between it is in parallel one " 10k Ω " resistance R, operational amplifier U2Output end output-va,
Output-vaAn input signal of circuit is realized as integrating channel two；Operational amplifier U1And U2Non-inverting input terminal connect
Ground；
The realization circuit of integrating channel two includes the input terminal-v inside neural networka、vb, operational amplifier U3And U4, with
And activation gradient tanh (λ x) function realizes circuit Tg2, input terminal-vaAnd vbSeries resistance R respectively4And R5It is followed by operation amplifier
Device U3Inverting input terminal；Operational amplifier U3Inverting input terminal and output end between it is in parallel one " 10nF " capacitor and one
The resistance R of a " 10k Ω "；Operational amplifier U3The voltage v of output end output2Electricity is realized by activation gradient tanh (λ x) function
Road Tg2After obtain output voltage vb；Output voltage vbThe resistance R of one " 10k Ω " is concatenated in operational amplifier U4Anti-phase input
End；Operational amplifier U4Inverting input terminal and output end between it is in parallel one " 10k Ω " resistance R, operational amplifier U4It is defeated
Outlet obtains output-vb, output-vbAn input signal of circuit is realized as integrating channel three；Operational amplifier U3And U4's
Non-inverting input terminal is grounded；
The realization circuit of integrating channel three includes the input terminal v inside neural networka、-vb、vc, operational amplifier U5, with
And activation gradient tanh (λ x) function realizes circuit Tg3, input terminal va、–vbAnd vcSeries resistance R respectively6、R7And R8It is followed by fortune
Calculate amplifier U5Inverting input terminal；Operational amplifier U5Inverting input terminal and output end between it is in parallel one " 10nF " electricity
Hold the resistance R of C and one " 10k Ω "；Operational amplifier U5The voltage v of output end output3By activating gradient tanh (λ x) function
Realize circuit Tg3After obtain output voltage vc；Output voltage vcIntegrating channel one, which is fed back to, as an input signal realizes electricity
Road；Operational amplifier U5Non-inverting input terminal connect " ground ".
A kind of Hopfield nerve network circuit with neuron activation gradient λ control is as shown in Figure 1, its system side
Journey is containing there are three state variable x1、x2And x3；Three variable v of corresponding circuits state equation1、v2And v3。
A kind of mathematical modeling: Hopfield neural fusion circuit controlled with neuron activation gradient λ of the present embodiment
Building is as shown in Figure 1.The present invention is based on a three-dimensional Hopfield neural networks, in order to preferably realize to Hopfield nerve
Network-based control introduces a neuron activation gradient function λ as control variable.It is realized in order to facilitate analysis and circuit, the mould
Type can be described with first-order ordinary differential equation system are as follows:
Wherein, x1、x2And x3Respectively three state variables of neuron.Parameter a is connection third neuron and first
The synapse weight of a neuron, parameter b are the synapse weights for connecting second neuron and first neuron.Usually set respectively
It is set to a=0.7, b=-2.Control variable λ is the activation gradient of neuron, indicates the sound of the electrical activity of neurons under electromagnetic induction
Answer speed.It is worth noting that, model (1) has a zero balancing in the zone of reasonableness of control variable λ (0.7 < λ < 1.5)
Point and two non-zero equalization points, they show different states with the increase of λ, and complicated dynamic behavior is presented.
Numerical simulation: when parameter of the neuron activation gradient λ as system, in two groups of initial values (0,1,0) and (0, -1,0)
Under, the dynamic behavior exhibition using MATLAB ODE23 algorithm to the Hopfield neural network controlled with neuron activation gradient λ
Open numerically modeling.When controlling variable λ=0.93, λ=1, Fig. 3 (a) and Fig. 4 (a), which depict two classes symmetrical attractor coexist, to exist
x1-x3Phase rail figure in plane, in initial value (0,1,0) and (0, -1,0), they correspond respectively to the upper and lower attractor coexisted,
Wherein Fig. 3 (a) presents 4 state of period coexisted, and Fig. 4 (a) illustrates the chaos spiral attractor coexisted.When control variable λ
=1.1, when λ=1.2, in initial value (0,1,0) and (0, -1,0), upper and lower attractor is respectively corresponded, Fig. 5 (a) and Fig. 6 (a) are retouched
It has drawn two classes and symmetrical attractor coexists in x1-x3Phase rail figure in plane.When variable λ=1.1, Fig. 5 (a), which is presented, to be coexisted
Spiral chaos attractor, when controlling variable λ=1.2, Fig. 6 (a) presents the multicycle state coexisted.Initial value be (0,1,
0) under conditions of, when controlling variable λ=1.3, Fig. 7 (a) presents the double scrollwork states of chaos；When controlling variable λ=1.5,
In Fig. 8 (a), system evolution is at 1 limit cycle of period.
In Fig. 2, the main circuit of three-dimensional Hopfield neural network is there are three integrating channel, for realizing the of formula (1)
One, second and third equation.According to the electrology characteristic of Kirchhoff's laws of electric circuit and circuit components, circuit shown in FIG. 1
Equation can be write as
Wherein, v1、v2And v3It is three Circuit variables, va、vbAnd vcIt is Circuit variable by the tanh based on gradient
Output voltage after function tanh (λ x) ,-va、–vbWith-vcIt is voltage value of this output voltage after inverting amplifier device.
Taking time precision is 0.1ms, i.e. R=10k Ω, C=10nF, in this way, can be obtained by contrast (1) and formula (2)
It arrives
So far, the present invention constructs a kind of Hopfield neural network and implementation with neuron activation gradient λ control.
Experimental verification: the design discrete device uses supply voltage for the TL082CP operation amplifier of ± 15V operating voltage
Device, discrete component select MPS2222 triode, metalfilmresistor, accurate adjustable resistance and monolithic capacitor.In experimentation, make
With Tyke TDS 3054C digital fluorescence oscilloscope test experiments result.When adjusting neuron activation gradient control variable λ=0.93,
That is adjustable resistance RF=450 Ω, capture in v1–v3Shown in phase rail figure such as Fig. 3 (b) in plane.Increase the activation ladder of neuron
λ value is spent, i.e., as λ=1, λ=1.1, λ=1.2, λ=1.3 and λ=1.5, in actual operation, adjustable electric resistance value is right respectively
It should be RF=511 Ω, RF=567 Ω, RF=600 Ω, RF=720 Ω and RF=850 Ω.Capture in v1–v3Phase in plane
Shown in rail figure such as Fig. 4 (b), 5 (b), 6 (b), 7 (b) and 8 (b).Ignore the numerical value as caused by calculating error and parasitic circuit parameter
Some fine differences between emulation and hardware circuit experiment, experimental result is almost consistent with numerical simulation, this shows proposition
It is a kind of with neuron activation gradient λ control Hopfield neural network formed advanced dynamic behavior can be demonstrate,proved by experiment
It is real.Therefore, a kind of Hopfield neural network with neuron activation gradient λ control constructed by the present invention has the theory of science
Foundation and realizability physically can play positive promotion to the engineer application of neuron models, artificial neural network and make
With.
Taking the above-mentioned ideal embodiment according to the present invention as inspiration, through the above description, relevant staff
Various changes and amendments can be carried out without departing from the scope of the present invention completely.The technical scope of this invention is not
The content being confined on specification, it is necessary to which the technical scope thereof is determined according to the scope of the claim.
Claims (4)
1. a kind of three-dimensional Hopfield neural network model with neuron activation gradient λ control realizes circuit, it is characterised in that: packet
It includes the negative output hyperbolic tangent function based on activation gradient and realizes circuit tanh module and the main electricity of three-dimensional Hopfield neural network
Road；
On the basis of the three-dimensional Hopfield neural network model of a routine, introduces a neuron activation gradient function λ and make
To control variable, which is indicated with first-order ordinary differential equation system are as follows:
In formula, x1、x2And x3Respectively three state variables of neuron；Parameter a is connection third neuron and first mind
Synapse weight through member, parameter b are the synapse weights for connecting second neuron and first neuron；Control variable λ is mind
Activation gradient through member indicates the response speed of the electrical activity of neurons under electromagnetic induction；
Three equations in formula (2), are respectively adopted integrating channel one, integrating channel two and integrating channel three to realize, three products
Main circuit of the subchannel as three-dimensional Hopfield neural network, according to the electricity of Kirchhoff's laws of electric circuit and circuit components
Characteristic, then the corresponding circuit equation of formula (2) can indicate are as follows:
In formula (3), v1、v2And v3It is three Circuit variables, it corresponds to the x in three-dimensional Hopfield nerve network system1、x2With
x3, indicate the film potential of three neurons in neural network.
2. circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control as described in claim 1,
It is characterized by: the realization circuit of integrating channel one includes the input terminal v inside neural networka、vb、vc, operational amplifier U1With
U2, and activation gradient tanh (λ x) function realization circuit Tg1, input terminal v inside the neural networka、vb、vcIt connects respectively
Resistance R1、R2And R3After be connected to operational amplifier U1Inverting input terminal；Operational amplifier U1Inverting input terminal and output end
Between be connected in parallel with a capacitor C and resistance R；Operational amplifier U1The voltage v of output end output1By activating gradient tanh (λ
X) function realizes circuit Tg1After obtain output voltage va；Output voltage vaOperational amplifier U is connected to after concatenating a resistance R2
Inverting input terminal；Operational amplifier U2Inverting input terminal and output end between parallel connection one resistance R, operational amplifier U2's
Output end output-va, output-vaAn input signal of circuit is realized as integrating channel two；Operational amplifier U1And U2It is same
Phase input terminal is grounded；
The realization circuit of integrating channel two includes the input terminal-v inside neural networka、vb, operational amplifier U3And U4, and swash
Stepladder degree tanh (λ x) function realizes circuit Tg2, input terminal-vaAnd vbSeries resistance R respectively4And R5It is followed by operational amplifier U3
Inverting input terminal；Operational amplifier U3Inverting input terminal and output end between be connected in parallel with a capacitor C and resistance R；Fortune
Calculate amplifier U3The voltage v of output end output2, the voltage v of output2Circuit Tg is realized by activation gradient tanh (λ x) function2Afterwards
Obtain output voltage vb；Output voltage vbA resistance R is concatenated in operational amplifier U4Inverting input terminal；Operational amplifier U4
Inverting input terminal and output end between parallel connection one resistance R, operational amplifier U4Output end obtain output-vb, output-vb
An input signal of circuit is realized as integrating channel three；Operational amplifier U3And U4Non-inverting input terminal be grounded；
The realization circuit of integrating channel three includes the input terminal v inside neural networka、-vb、vc, operational amplifier U5, and swash
Stepladder degree tanh (λ x) function realizes circuit Tg3, input terminal va、–vbAnd vcSeries resistance R respectively6、R7And R8It is followed by putting in operation
Big device U5Inverting input terminal；Operational amplifier U5Inverting input terminal and output end between be connected in parallel with a capacitor C and resistance
R；Operational amplifier U5The voltage v of output end output3, voltage v3Circuit Tg is realized by activation gradient tanh (λ x) function3Afterwards
To output voltage vc；Output voltage vcIntegrating channel one, which is fed back to, as an input signal realizes circuit, operational amplifier U5Together
Phase input terminal connects " ground ".
3. circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control as claimed in claim 2,
It is characterized by: activation gradient tanh (λ x) function realizes that circuit includes: integrator, triode T1And T2, DC voltage source E and
DC current source I0, wherein integrator includes operational amplifier UiAnd Uo, specific connection type are as follows: input terminal viIt connects an electricity
Resistance R is followed by operational amplifier UiInverting input terminal；Operational amplifier UiInverting input terminal and output end between one in parallel
Changeable resistance RF；Operational amplifier UiOutput end access triode T1Base stage；Triode T1Collector be divided into two
Road is connected to operational amplifier U through a resistance R all the wayoNoninverting input, another way is through a resistance RCIt is connected to direct current
Voltage source E；Triode T2Emitter and triode T1Emitter be connected, and at the same time meeting DC current source I0；Triode T2
Collector also meet a resistance RCIt is connected on DC voltage source E, another terminating resistor R to amplifier UoReverse input end；Three
Pole pipe T2Base earth, triode T2Base stage concatenate resistance R to amplifier U againoNoninverting input；Operational amplifier
UoAnti-phase input terminate a resistance R to output end vo。
4. circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control as claimed in claim 3,
It is characterized by: the DC voltage source E is fixed 15V DC power supply, the resistance R of control neural networkFFor adjustable resistance,
Adjustable extent is 500 Ω -900 Ω, can directly control the dynamic behavior of neural network.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910283951.7A CN110097182B (en) | 2019-04-10 | 2019-04-10 | Three-dimensional Hopfield neural network model realization circuit controlled by nerve activation gradient lambda |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201910283951.7A CN110097182B (en) | 2019-04-10 | 2019-04-10 | Three-dimensional Hopfield neural network model realization circuit controlled by nerve activation gradient lambda |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN110097182A true CN110097182A (en) | 2019-08-06 |
| CN110097182B CN110097182B (en) | 2023-03-24 |
Family
ID=67444519
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201910283951.7A Active CN110097182B (en) | 2019-04-10 | 2019-04-10 | Three-dimensional Hopfield neural network model realization circuit controlled by nerve activation gradient lambda |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN110097182B (en) |
Cited By (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112906878A (en) * | 2021-03-11 | 2021-06-04 | 杭州电子科技大学 | Hopfield neural network model under simulated electromagnetic radiation |
| CN113112010A (en) * | 2021-04-29 | 2021-07-13 | 齐鲁工业大学 | Nerve fiber equivalent circuit supporting soliton wave conduction |
| CN113379044A (en) * | 2021-05-21 | 2021-09-10 | 长沙理工大学 | Hopfield neural network system based on electromagnetic radiation effect, processor chip and terminal |
| CN114881220A (en) * | 2022-05-17 | 2022-08-09 | 常州大学 | Cubic nonlinear function fitting circuit based on FHN neuron |
| CN115062583A (en) * | 2022-06-15 | 2022-09-16 | 华中科技大学 | Hopfield network hardware circuit for solving optimization problem and operation method |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106485317A (en) * | 2016-09-26 | 2017-03-08 | 上海新储集成电路有限公司 | A kind of neutral net accelerator and the implementation method of neural network model |
| CN106815636A (en) * | 2016-12-30 | 2017-06-09 | 华中科技大学 | A kind of neuron circuit based on memristor |
| CN107784359A (en) * | 2017-09-19 | 2018-03-09 | 常州大学 | A kind of more stable state oscillation circuits based on Hopfield neutral nets |
| CN108427843A (en) * | 2018-03-14 | 2018-08-21 | 常州大学 | It is a kind of that there is the three-dimensional memristor Hindmarsh-Rose precircuits hidden and asymmetric behavior coexists |
-
2019
- 2019-04-10 CN CN201910283951.7A patent/CN110097182B/en active Active
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN106485317A (en) * | 2016-09-26 | 2017-03-08 | 上海新储集成电路有限公司 | A kind of neutral net accelerator and the implementation method of neural network model |
| CN106815636A (en) * | 2016-12-30 | 2017-06-09 | 华中科技大学 | A kind of neuron circuit based on memristor |
| CN107784359A (en) * | 2017-09-19 | 2018-03-09 | 常州大学 | A kind of more stable state oscillation circuits based on Hopfield neutral nets |
| CN108427843A (en) * | 2018-03-14 | 2018-08-21 | 常州大学 | It is a kind of that there is the three-dimensional memristor Hindmarsh-Rose precircuits hidden and asymmetric behavior coexists |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN112906878A (en) * | 2021-03-11 | 2021-06-04 | 杭州电子科技大学 | Hopfield neural network model under simulated electromagnetic radiation |
| CN113112010A (en) * | 2021-04-29 | 2021-07-13 | 齐鲁工业大学 | Nerve fiber equivalent circuit supporting soliton wave conduction |
| CN113379044A (en) * | 2021-05-21 | 2021-09-10 | 长沙理工大学 | Hopfield neural network system based on electromagnetic radiation effect, processor chip and terminal |
| CN114881220A (en) * | 2022-05-17 | 2022-08-09 | 常州大学 | Cubic nonlinear function fitting circuit based on FHN neuron |
| CN114881220B (en) * | 2022-05-17 | 2023-11-14 | 常州大学 | FHN neuron-based cubic nonlinear function fitting circuit |
| CN115062583A (en) * | 2022-06-15 | 2022-09-16 | 华中科技大学 | Hopfield network hardware circuit for solving optimization problem and operation method |
| CN115062583B (en) * | 2022-06-15 | 2024-05-31 | 华中科技大学 | Hopfield network hardware circuit for solving optimization problem and operation method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN110097182B (en) | 2023-03-24 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN110097182A (en) | Circuit is realized with the three-dimensional Hopfield neural network model of neuron activation gradient λ control | |
| Ding et al. | Memristor synapse-coupled piecewise-linear simplified Hopfield neural network: Dynamics analysis and circuit implementation | |
| Arthur et al. | Silicon-neuron design: A dynamical systems approach | |
| Bao et al. | Memristive cyclic three-neuron-based neural network with chaos and global coexisting attractors | |
| Saighi et al. | A library of analog operators based on the Hodgkin-Huxley formalism for the design of tunable, real-time, silicon neurons | |
| CN104573238B (en) | A kind of circuit design method of memristor cell neural network | |
| Levi et al. | Digital implementation of Hodgkin–Huxley neuron model for neurological diseases studies | |
| Eisenberg et al. | Hardware architecture of a neural network model simulating pattern recognition by the olfactory bulb | |
| CN113344191B (en) | Continuous Rulkov electronic neuron circuit with super multi-stability | |
| CN110399648A (en) | A kind of three-dimensional HNN model realization circuit with ideal memristor stiffness of coupling control | |
| CN109978159A (en) | A kind of easy Fitzhugh-Nagumo neuron circuit | |
| Cao et al. | A discrete Chialvo–Rulkov neuron network coupled with a novel memristor model: Design, Dynamical analysis, DSP implementation and its application | |
| CN103729518A (en) | Simple memristor emulator | |
| CN108427843A (en) | It is a kind of that there is the three-dimensional memristor Hindmarsh-Rose precircuits hidden and asymmetric behavior coexists | |
| CN107784359A (en) | A kind of more stable state oscillation circuits based on Hopfield neutral nets | |
| CN112906878A (en) | Hopfield neural network model under simulated electromagnetic radiation | |
| Zhang et al. | A multiplier-free Rulkov neuron under memristive electromagnetic induction: Dynamics analysis, energy calculation, and circuit implementation | |
| Wang et al. | Firing activities analysis of a novel small heterogeneous coupled network through a memristive synapse | |
| Bao et al. | Bifurcations to bursting oscillations in memristor-based FitzHugh-Nagumo circuit | |
| Saighi et al. | A conductance-based silicon neuron with dynamically tunable model parameters | |
| Zhang et al. | Switchable memristor-based Hindmarsh-Rose neuron under electromagnetic radiation | |
| Long et al. | Memristive-synapse spiking neural networks based on single-electron transistors | |
| Mladenović | Contemporary electronics with LTSpice and mathematica | |
| Xylouris et al. | A three-dimensional mathematical model for the signal propagation on a neuron's membrane | |
| Borwankar et al. | An analog implementation of fitzhugh-nagumo neuron model for spiking neural networks |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| EE01 | Entry into force of recordation of patent licensing contract | ||
| EE01 | Entry into force of recordation of patent licensing contract |
Application publication date: 20190806 Assignee: Changzhou Ruixinteng Microelectronics Co.,Ltd. Assignor: CHANGZHOU University Contract record no.: X2023980054127 Denomination of invention: Using neural activation gradients l Implementation of a circuit using a three-dimensional Hopfield neural network model for control Granted publication date: 20230324 License type: Common License Record date: 20231227 |