CN106815636B - A kind of neuron circuit based on memristor - Google Patents
A kind of neuron circuit based on memristor Download PDFInfo
- Publication number
- CN106815636B CN106815636B CN201611258488.3A CN201611258488A CN106815636B CN 106815636 B CN106815636 B CN 106815636B CN 201611258488 A CN201611258488 A CN 201611258488A CN 106815636 B CN106815636 B CN 106815636B
- Authority
- CN
- China
- Prior art keywords
- circuit
- neuron
- oxide
- semiconductor
- metal
- 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.)
- Active
Links
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
- G06N3/06—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
- G06N3/065—Analogue means
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)
- General Health & Medical Sciences (AREA)
- Data Mining & Analysis (AREA)
- Evolutionary Computation (AREA)
- Computational Linguistics (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- General Engineering & Computer Science (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Artificial Intelligence (AREA)
- Neurology (AREA)
- Networks Using Active Elements (AREA)
Abstract
The invention discloses a kind of neuron circuit based on memristor, including Sudden-touch circuit, neuron activation functions circuit and synapse weight control circuit;In Sudden-touch circuit, a memristor changes memristor value under the control of four metal-oxide-semiconductors to simulate variation of the synapse weight in biological neural network.Designed neuronal synapse circuit can be connected directly with digital logic level, easy real-time synapse weight to be adjusted.This feature is limited by supply voltage using the output voltage of operational amplifier, realizing neuron circuit activation primitive is saturated linear function.The synapse weight of neuron, which changes circuit, can be used existing CMOS microcontroller, meanwhile, neural network algorithm can be loaded in the microcontroller and changes synapse weight, realize corresponding function.Multiple neuron circuits can be connected into large-scale neural network by the invention patent, realize complicated function.Such as pattern-recognition, signal processing, associative memory, non-linearity mapping etc..
Description
Technical field
The invention belongs to analog and digital circuit fields and emerging field of circuit technology, are based on recalling more particularly, to one kind
Hinder the neuron circuit of device.
Background technique
Memristor is taught in 1971 by the Cai Shaotang of University of California Berkeley (Leon.O.Chua) from theory at first
On be derived by.2008, HP Lab was manufactured that memristor in kind truly for the first time.Memristor is one
Dynamic element, when applying external voltage to it, memristor value can be increased or reduced with alive direction is applied.Therefore, memristor is used
The memristor value of device is come to indicate the strong and weak of the weight of Synaptic junction in neural network be a suitable selection.
Artificial neural network (Artificial Neural Network, ANN), artificial intelligence since being the 1980s
The research hotspot that energy field is risen.It is the spy according to biological neural network (Biological Neural Network, BNN)
What point was set up has powerful operational capability, and the operational model that can be realized by means of information technology.One large-scale people
The problem of artificial neural networks are able to solve pattern-recognition, signal processing, associative memory etc..
Single neuron in neural network is divided into two parts: cynapse part and neuron activation functions part.Cynapse
It is connected to the input terminal of neuron, a neuron there can be multiple Synaptic junctions as input.Cynapse acts through cynapse
Weight shows, and the bigger cynapse of weight is stronger to the effect of neuron.When signal is by cynapse, signal is done with synapse weight
Product, obtained result are transferred to neuron activation functions part.Neuron activation functions part is by each cynapse first
Product signal is summed, then obtained summing signal is transformed to the output needed according to activation primitive.A large amount of nerve
Member forms a large-scale neural network by Synaptic junction together, solves the problems, such as complexity with this.For example, with existing meter
The insoluble traveling salesman problem of calculation machine can be quickly obtained result using Hopfield neural network very much.Therefore, single mind
Design through first circuit is extremely important for realizing large-scale neural network.
Use memristor as cynapse, there are several different neuron circuit implementations.
Kurtis D.Cantley etc. devises a kind of Leaky Integrate-and-Fire neuron circuit.This mind
7 kinds of different supply voltage power supplies are needed through first circuit, and the precise requirements of power supply voltage are very high.For example, wherein
Needing a kind of supply voltage is 0.35V, and this supply voltage is realized relatively difficult in engineering, is unfavorable for circuit realization.Due to
Its neuron circuit designed, the description of not stringent mathematical relationship, neuron can not answer also without activation primitive
It uses and forms large-scale neural network in practice and solve practical problems.
South Korea KIM professor team devises a kind of neuron circuit.Four memristors that its Sudden-touch circuit is connected by bridge-type
Composition, Sudden-touch circuit are connected to subsequent circuit composition neuron circuit, and this neuron circuit only realizes input signal
Summation, is not neural meta design activation primitive circuit.Moreover, the change of the signal input part and synapse weight of this neuron
End is in the same port, and therefore, neuron is to the response of input signal and the change of synapse weight (i.e. in change bridge-type cynapse
Four memristors memristor value) need to be divided into two steps to complete, the speed of service of neuron is relatively low.
S.G.Hu etc. devises a kind of neuron circuit, and forms Hopfield neural network with it.But in nerve
First circuit run when, synapse weight (the memristor value of memristor) need to be adjusted offline by instrument (need memristor from
Individually taken out in circuit, changed and be placed into circuit after memristor value again) neuron that designs in this way can only use
In the effect of verifying design, can not be applied in actual circuit.
Disclose a kind of neuron circuit in Chinese invention patent CN104335224A, but this neuron circuit mould
Block is more, implements complexity, is not easy to composition catenet.
Summary of the invention
In view of the drawbacks of the prior art, the purpose of the present invention is to provide a kind of neuron circuit based on memristor, purports
The neuron based on memristor is being solved, Sudden-touch circuit cannot be compatible with existing digital logic level, in neuron circuit
Memristor value cannot synchronize on-line control, and not using saturated linear function as the circuit of activation primitive, neural network algorithm is not easy to
It is loaded into the problems in neuron circuit.
The present invention provides a kind of neuron circuits based on memristor, comprising: n Sudden-touch circuit, neuronal activation letter
Number circuit and synapse weight control circuit, each Sudden-touch circuit have an input terminal, a feedback end and an output end, dash forward
For the input terminal on electric shock road for receiving input voltage, the feedback end of Sudden-touch circuit is connected to the defeated of the synapse weight control circuit
Outlet;The neuron activation functions circuit have n input terminal and an output end, respectively with the output of the Sudden-touch circuit
End connects one to one, output end of the output end of the neuron activation functions circuit as neuron circuit, the cynapse
The input terminal of weight control circuit is connected to the output end of the neuron activation functions circuit;The Sudden-touch circuit is used for will be defeated
Enter signal and synapse weight do product after export, the neuron activation functions circuit be used for by after n product signal addition again
Neuron output is obtained according to the relationship of activation primitive;The synapse weight control circuit is used to be fed back according to the output of neuron
Control signal is generated, and adjusts the synapse weight.
Further, the structure of n Sudden-touch circuit is identical, and each Sudden-touch circuit includes: the first metal-oxide-semiconductor T1,
Two metal-oxide-semiconductor T2, third metal-oxide-semiconductor T3, the 4th metal-oxide-semiconductor T4, phase inverter M1, resistance R1, operational amplifier A1With memristor Rm1;It is described anti-
Phase device M1Input terminal for receive control signal VC1, the first metal-oxide-semiconductor T1Control terminal be connected to the phase inverter M1's
Output end, the first metal-oxide-semiconductor T1One end for receiving input voltage, the first metal-oxide-semiconductor T1The other end be connected to it is described
Memristor Rm1Anode;The second metal-oxide-semiconductor T2One end be connected to the first metal-oxide-semiconductor T1One end, second metal-oxide-semiconductor
T2The other end be connected to the memristor Rm1Cathode, the second metal-oxide-semiconductor T2Control terminal be connected to the phase inverter M1's
Input terminal;The third metal-oxide-semiconductor T3Control terminal be connected to the phase inverter M1Input terminal, the third metal-oxide-semiconductor T3One end
It is connected to the memristor Rm1Cathode, the third metal-oxide-semiconductor T3The other end be connected to the operational amplifier A1Same phase
Input terminal;The 4th metal-oxide-semiconductor T4One end be connected to the memristor Rm1Anode, the 4th metal-oxide-semiconductor T4The other end
It is connected to the operational amplifier A1Non-inverting input terminal, the 4th metal-oxide-semiconductor T4Control terminal be connected to the phase inverter M1's
Output end;The resistance R1One end be connected to the operational amplifier A1Non-inverting input terminal, the resistance R1Another termination
Ground;The operational amplifier A1Inverting input terminal be connected to its output end, the operational amplifier A1Output end be used as described in
The output end of Sudden-touch circuit.
Further, the neuron activation functions circuit includes: resistance R7, resistance R8, operational amplifier A4And voltage
Summing circuit;The voltage summing circuit has n input terminal and an output end, n input of the voltage summing circuit
The n input terminal respectively as the neuron activation functions circuit is held, the output end of the voltage summing circuit is connected to institute
State operational amplifier A4Non-inverting input terminal, the resistance R7With the resistance R8It is sequentially connected in series and is put on ground and the operation
Big device A4Output end between, and the series connection end of the resistance R7 and the resistance R8 and the operational amplifier A4It is anti-
The connection of phase input terminal, the operational amplifier A4Output end of the output end as the neuron activation functions circuit.
Further, the voltage summing circuit includes: resistance R9With n first resistor;One end of n first resistor
As n input terminal of the voltage summing circuit, the other end of n first resistor is all connected with and sums as the voltage electric
The output end on road;The resistance R9One end be connected to the other end of n first resistor, the resistance R9The other end ground connection.
Further, the synapse weight control circuit includes: microcontroller, the input terminal conduct of the microcontroller
The input terminal of the synapse weight control circuit, the output end of the microcontroller is as the defeated of the synapse weight control circuit
Outlet.
Further, it when work, sums, swashs after the voltage of n input signal is done product with corresponding synapse weight
Function living carries out the result that summation obtains to convert the output as neuron.The output feedback of neuron is controlled to synapse weight
Circuit generates control signal with this to change synapse weight.
Further, it in the microcontroller of synapse weight control circuit, is adjusted by load neural network algorithm
Synapse weight realizes corresponding function.
In general, through the invention it is contemplated above technical scheme is compared with the prior art, can obtain and control
When synapse weight and composition large size neural network, there is good effect:
(1) the memristor value in Sudden-touch circuit can be with on-line control.
(2) neuron circuit works under existing digital logic voltage, it is only necessary to general digital voltage such as 5V.
(3) neuron can be realized simultaneously response input signal and synapse weight is adjusted.
(4) the neuron activation functions circuit of saturated linear function performance is devised.
(5) neural network algorithm (such as Widrow-Hoff algorithm) can be loaded in digital control circuit to adjust weight,
The neuron is simultaneously convenient for the large-scale neural network of composition.
Detailed description of the invention
Fig. 1 is the structure chart of neuron circuit designed by the present invention.
Fig. 2 is the specific implementation circuit diagram of neuron designed by the present invention, and there are three inputs to dash forward for the neuron band in figure
Touching.
Fig. 3 is the circuit diagram of Sudden-touch circuit in neuron designed by the present invention.
Fig. 4 is the circuit diagram of activation primitive circuit in neuron designed by the present invention.
Fig. 5 is the math function figure that neuron circuit designed by the present invention can be realized.
Fig. 6 is the structure chart of synapse weight control circuit in neuron designed by the present invention.
Specific embodiment
In order to make the objectives, technical solutions, and advantages of the present invention clearer, with reference to the accompanying drawings and embodiments, right
The present invention is further elaborated.It should be appreciated that the specific embodiments described herein are merely illustrative of the present invention, and
It is not used in the restriction present invention.
The purpose of the present invention is to provide a kind of neuron circuits based on memristor.It can be directly with existing number electricity
What road generated control signal to control memristor in Sudden-touch circuit recalls change in resistance.In Sudden-touch circuit, the memristor value of memristor can
To carry out on-line control, that is, memristor value is adjusted in circuit operation.The input signal end of the neuron and synapse weight
Adjustment control terminal is different port, and therefore, it can be realized simultaneously response input signal and synapse weight adjustment.Utilize operation
The output voltage of amplifier is limited this feature by supply voltage, and devising neuron activation functions is saturated linear function.
Since this design can combine digital logic techniques with memristor, it can add in synapse weight control circuit
Neural network algorithm is carried to adjust synapse weight, realizes corresponding function.
Fig. 1 is the structure chart of neuron circuit designed by the present invention.As shown in Figure 1, neuron circuit is by Sudden-touch circuit,
Three parts of neuron activation functions circuit and synapse weight control circuit form.Sudden-touch circuit is by input signal and synapse weight
Product is done, next stage is output to.Obtained product signal is summed and then according to activation primitive by neuron activation functions circuit
Relationship obtain neuron output.Synapse weight control circuit is fed back according to the output of neuron generates control signal, and adjustment is prominent
Touch weight.As signal VI1-VInIt is input to n Sudden-touch circuit, the structure of this n Sudden-touch circuit is identical, VI1-VInRespectively
It is exported after doing product with n synapse weight.This n product signal is summed in neuron activation functions circuit, is obtained
Summing signal is mapped as certain voltage output according to the functional relation of activation primitive.This output voltage is fed back simultaneously arrives cynapse
Weight control circuit can be according to the synapse weight more new algorithm of neural network (such as in synapse weight control circuit
Widrow-Hoff algorithm) amount that weight needs to change is obtained, then circuit is weighed cynapse by generating the time of control signal
Change again to corresponding value.
Clearer in order to illustrate neuron circuit, here by taking n=3 as an example, describing band, there are three input cynapse
Neuron circuit.Specific circuit diagram is as shown in Figure 2.The neuron circuit that sub-module place of matchmakers is invented below.
Because the structure of three Sudden-touch circuits be it is identical, only take first Sudden-touch circuit to be illustrated here, first
The circuit diagram of Sudden-touch circuit is as shown in Figure 3.VI1It is input signal, VI1Connect the first metal-oxide-semiconductor T1With the second metal-oxide-semiconductor T2Source electrode.
First metal-oxide-semiconductor T1Drain electrode connect memristor Rm1"+" end, the second metal-oxide-semiconductor T2Drain electrode connect memristor Rm1"-" end.The
Three metal-oxide-semiconductor T3Drain electrode connect memristor Rm1"-" end, the 4th metal-oxide-semiconductor T4Drain electrode connect memristor Rm1"+" end, third
Metal-oxide-semiconductor T3With the 4th metal-oxide-semiconductor T4Source electrode connect resistance R1One end.Resistance R1This one end be connected to operational amplifier A1It is same
Xiang Duan, resistance R1The other end ground connection, A1Reverse side connect A1Output end.T1And T2It is p-type metal-oxide-semiconductor, T3And T4It is N-type
Metal-oxide-semiconductor.Control signal VC1Connect phase inverter M1Input terminal, the second metal-oxide-semiconductor T2Grid and third metal-oxide-semiconductor T3Grid.Instead
Phase device M1Output end connect the first metal-oxide-semiconductor T1Grid and the 4th metal-oxide-semiconductor T4Grid.
VC1And M1Co- controlling T1, T2, T3, T4Conducting and cut-off.T1, T2Grid level logic state on the contrary, protect
Card is in VC1Under effect, only one in two PMOS tube is connected.Same connection, guarantees VC1Under effect, T3, T4In only one
A conducting.Work as VI1When for high level, if control signal VC1For high level, then phase inverter M1Export low level.In this state
Under.T1And T3Conducting, T2And T4Cut-off.Electric current is from T1, memristor, T3And R1Flow into ground terminal.In this case, memristor
Resistance value increases.Opposite, if control signal VC1For low level, then phase inverter Mb1Export high level, T1And T3Cut-off, T2And T4
Conducting, the memristor value of memristor are in reduction state.Work as VI1When for low level, signal VC no matter is controlled1For high level or low electricity
It is flat, there is no electric current to pass through in memristor, the resistance value of memristor will not all change.Therefore, for Sudden-touch circuit, when there is input
When signal, i.e., when input signal is high level, it is only necessary to change the level of control signal, so that it may realize the online of synapse weight
Mode is adjusted.
T1-T4, memristor Rm1With resistance R1A divider is formed, when the memristor value of memristor increases, R1On get
Voltage reduces.Opposite, when the memristor value of memristor reduces, R1On the voltage got increase.R1On partial pressure be following formula:
Metal-oxide-semiconductor T1-T4It all works on or off state, because the equivalent resistance of the metal-oxide-semiconductor of on state is very small,
Relative to memristor Rm1With resistance R1Resistance value can ignore this equivalent resistance, it is therefore assumed here that in the conductive state
The equivalent resistance of transistor is 0.When memristor resistance value is constant, that is, when input signal is zero, resistance R1On partial pressure be 0V.
Resistance R1On partial pressure be connected to operational amplifier A1In-phase end, operational amplifier A1Connect into a voltage with
With device, by resistance R1On partial pressure copy to the output end of operational amplifier.The effect of voltage follower be by divider and after
The neuron circuit in face is isolated, and is avoided interfering with each other between the two, is also improved the carrying load ability of Sudden-touch circuit.
Neuron activation functions circuit is as shown in Figure 4.A, B, C respectively represent the output voltage of three Sudden-touch circuits, A connection
To first resistor R4One end, B is connected to second resistance R5One end, C is connected to 3rd resistor R6One end.R4, R5, R6It is another
One end is connected to the 4th resistance R9One end, R9This one end be connected to operational amplifier A simultaneously4In-phase end, R9The other end
Ground connection.5th resistance R7One end ground connection, the other end connect the 6th resistance R8One end and operational amplifier reverse side.R8's
The other end is connected to the output end of operational amplifier.
Resistance R4, R5, R6, R9A voltage summing circuit is formed, the voltage of A, B, C-terminal are added with certain relationship.This
In assume R4, R5, R6, R9Resistance be the equal R that is set as, then the addition relationship of voltage is as follows:
Operational amplifier A4, R7And R8A voltage amplifier circuit is formed, by voltage VR9Amplify certain multiple.Here it sets
Determine R8Resistance value be R7Twice, therefore, VR9It is amplified twice.
The output voltage of operational amplifier is limited by supply voltage, is greater than operation when passing through amplified output voltage
When the supply voltage of amplifier, output is limited in positive supply voltage.Similarly, when amplified output voltage be less than 0V,
0V can be limited in.Due to summing, obtained voltage is amplified twice, the voltage after summation arrives supply voltage in 0V
When between half, which can be linearly amplified.If the voltage after summation is greater than the half of supply voltage, exporting can be limited
System is in supply voltage.When voltage after summation is less than 0V, output is limited in 0V.It is hereby achieved that the activation of this neuron
Function is saturated linear function.The mathematic(al) representation of saturated linear function is as follows:
Therefore, for this band, there are three the neuron circuits of cynapse, the pass between its input voltage and output voltage
System can be indicated by following formula:
The supply voltage of operational amplifier is Vdd, here by synapse weight is defined as:
The math function that this band is realized there are three the neuron circuit of cynapse is as shown in Figure 5.VI1, VI2, VI3It is three
Input signal, respectively with three synapse weight W1, W2, W3Do product.Then, obtained product is summed, according still further to linear
The constraint of saturation function obtains the output voltage V of neuronout。
The synapse weight control circuit of neuron is as shown in fig. 6, the output signal of neuron feeds back the cynapse to neuron
The input terminal of weight control circuit carries out A/D conversion feeding to this voltage in the input terminal of microcontroller and is loaded with nerve net
The microcontroller of network algorithm (such as Widrow-Hoff algorithm).In the microcontroller, by the mesh of collected voltage and setting
Mark voltage compares.Then, according to error between the two, it is converted into control signal VC1, VC2, VC3Level duration, come
Control the variation of corresponding memristor value.
When input signal is applied to neuron, neuron exports corresponding voltage value and responds input signal.Meanwhile cynapse
Weight control circuit output control signal, controls increaseing or decreasing for corresponding synapse weight.In this way, designed neuron circuit
The response and synapse weight for being achieved that input signal change while carrying out.Input signal in this circuit, control signal are all
The high level of digital logic signal, logical signal is set as Vdd, equally the supply voltage of operational amplifier is set as Vdd, in this way
The function of neuron can be completed by only using a kind of voltage.Because input and control signal are all logical signal, synapse weight control
Circuit processed is made of microcontroller, therefore, designed neuron be convenient for existing circuit connection, and be easy to be extended to
Large-scale neural network.
The present invention provides a kind of neuron circuit based on memristor, designed neuron circuit has novel dash forward
Electric shock road has the synapse weight using saturated linear function as the neuron activation functions circuit of function, realized with digit chip
Control circuit.This neuron circuit has very strong scalability, and large-scale neural network is realized convenient for combination.
As it will be easily appreciated by one skilled in the art that the foregoing is merely illustrative of the preferred embodiments of the present invention, not to
The limitation present invention, any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should all include
Within protection scope of the present invention.
Claims (6)
1. a kind of neuron circuit based on memristor characterized by comprising n Sudden-touch circuit, neuron activation functions electricity
Road and synapse weight control circuit, each Sudden-touch circuit have an input terminal, a feedback end and an output end, cynapse electricity
For the input terminal on road for receiving input voltage, the feedback end of Sudden-touch circuit is connected to the output of the synapse weight control circuit
End;The neuron activation functions circuit have n input terminal and an output end, respectively with the output end of the Sudden-touch circuit
It connects one to one, output end of the output end of the neuron activation functions circuit as neuron circuit, the cynapse power
The input terminal of weight control circuit is connected to the output end of the neuron activation functions circuit;
The Sudden-touch circuit for exporting after input signal and synapse weight are done product, use by the neuron activation functions circuit
Relationship after being added n product signal further according to activation primitive obtains neuron output;The synapse weight control circuit
Control signal is generated for feeding back according to the output of neuron, and adjusts the synapse weight;
The structure of n Sudden-touch circuit is identical, and each Sudden-touch circuit includes: the first metal-oxide-semiconductor T1, the second metal-oxide-semiconductor T2, the 3rd MOS
Pipe T3, the 4th metal-oxide-semiconductor T4, phase inverter M1, resistance R1, operational amplifier A1With memristor Rm1;
The phase inverter M1Input terminal for receive control signal VC1, the first metal-oxide-semiconductor T1Control terminal be connected to it is described
Phase inverter M1Output end, the first metal-oxide-semiconductor T1One end for receiving input voltage, the first metal-oxide-semiconductor T1The other end
It is connected to the memristor Rm1Anode;
The second metal-oxide-semiconductor T2One end be connected to the first metal-oxide-semiconductor T1One end, the second metal-oxide-semiconductor T2The other end connect
It is connected to the memristor Rm1Cathode, the second metal-oxide-semiconductor T2Control terminal be connected to the phase inverter M1Input terminal;
The third metal-oxide-semiconductor T3Control terminal be connected to the phase inverter M1Input terminal, the third metal-oxide-semiconductor T3One end connect
It is connected to the memristor Rm1Cathode, the third metal-oxide-semiconductor T3The other end be connected to the operational amplifier A1It is same mutually defeated
Enter end;
The 4th metal-oxide-semiconductor T4One end be connected to the memristor Rm1Anode, the 4th metal-oxide-semiconductor T4The other end connection
To the operational amplifier A1Non-inverting input terminal, the 4th metal-oxide-semiconductor T4Control terminal be connected to the phase inverter M1Output
End;
The resistance R1One end be connected to the operational amplifier A1Non-inverting input terminal, the resistance R1The other end ground connection;
The operational amplifier A1Inverting input terminal be connected to its output end, the operational amplifier A1Output end as institute
State the output end of Sudden-touch circuit.
2. neuron circuit as described in claim 1, which is characterized in that the neuron activation functions circuit includes: resistance
R7, resistance R8, operational amplifier A4With voltage summing circuit;
The voltage summing circuit has n input terminal and an output end, and n input terminal of the voltage summing circuit is distinguished
As n input terminal of the neuron activation functions circuit, the output end of the voltage summing circuit is connected to the operation
Amplifier A4Non-inverting input terminal, the resistance R7With the resistance R8It is sequentially connected in series on ground and the operational amplifier A4
Output end between, and the resistance R7With the resistance R8Series connection end and the operational amplifier A4Anti-phase input
End connection, the operational amplifier A4Output end of the output end as the neuron activation functions circuit.
3. neuron circuit as claimed in claim 2, which is characterized in that the voltage summing circuit includes: resistance R9With n
First resistor;
N input terminal of the one end of n first resistor as the voltage summing circuit, the other end of n first resistor connect
It is connected together, and the output end as the voltage summing circuit;
One end of the resistance R9 is connected to the other end of n first resistor, the other end ground connection of the resistance R9.
4. neuron circuit as described in any one of claims 1-3, which is characterized in that the synapse weight control circuit packet
It includes: microcontroller, input terminal of the input terminal of the microcontroller as the synapse weight control circuit, the microcontroller
Output end of the output end as the synapse weight control circuit.
5. neuron circuit as claimed in claim 4, which is characterized in that when work, by the voltage of n input signal with it is corresponding
Synapse weight do product after sum, activation primitive carries out the obtained result of summation to convert the output as neuron;Nerve
The output feedback of member generates control signal to synapse weight control circuit with this to change synapse weight.
6. neuron circuit as described in any one of claims 1-3, which is characterized in that in the micro-control of synapse weight control circuit
In device processed, synapse weight is adjusted by load neural network algorithm, realizes corresponding function.
Priority Applications (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611258488.3A CN106815636B (en) | 2016-12-30 | 2016-12-30 | A kind of neuron circuit based on memristor |
Applications Claiming Priority (1)
Application Number | Priority Date | Filing Date | Title |
---|---|---|---|
CN201611258488.3A CN106815636B (en) | 2016-12-30 | 2016-12-30 | A kind of neuron circuit based on memristor |
Publications (2)
Publication Number | Publication Date |
---|---|
CN106815636A CN106815636A (en) | 2017-06-09 |
CN106815636B true CN106815636B (en) | 2019-03-05 |
Family
ID=59110552
Family Applications (1)
Application Number | Title | Priority Date | Filing Date |
---|---|---|---|
CN201611258488.3A Active CN106815636B (en) | 2016-12-30 | 2016-12-30 | A kind of neuron circuit based on memristor |
Country Status (1)
Country | Link |
---|---|
CN (1) | CN106815636B (en) |
Families Citing this family (44)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN106779059B (en) * | 2016-12-30 | 2019-03-05 | 华中科技大学 | A kind of Circuit of Artificial Neural Networks of the Pavlov associative memory based on memristor |
US11270194B2 (en) * | 2017-07-26 | 2022-03-08 | International Business Machines Corporation | System and method for constructing synaptic weights for artificial neural networks from signed analog conductance-pairs of varying significance |
CN107545305B (en) * | 2017-09-15 | 2020-01-03 | 中国科学技术大学 | CMOS (complementary metal oxide semiconductor) process-based digital-analog mixed charge domain neuron circuit |
CN107742153B (en) * | 2017-10-20 | 2020-02-21 | 华中科技大学 | Memristor-based neuron circuit with steady-state plasticity |
CN107909146B (en) * | 2017-11-13 | 2021-09-17 | 中国科学院微电子研究所 | Neuron circuit based on volatile threshold transition device |
CN109829539A (en) * | 2017-11-23 | 2019-05-31 | 旺宏电子股份有限公司 | Class nerve computing device |
CN107832846A (en) * | 2017-11-30 | 2018-03-23 | 电子科技大学 | A kind of neuron circuit for integrating a variety of activation primitives |
US11354562B2 (en) * | 2018-01-03 | 2022-06-07 | Silicon Storage Technology, Inc. | Programmable neuron for analog non-volatile memory in deep learning artificial neural network |
CN108416432B (en) * | 2018-01-19 | 2021-10-01 | 北京大学 | Circuit and method for operating a circuit |
CN108268938B (en) * | 2018-01-24 | 2020-04-21 | 清华大学 | Neural network, information processing method thereof and information processing system |
CN108804786B (en) * | 2018-05-26 | 2022-04-15 | 江西理工大学 | Memristive model circuit design method for realizing plasticity of synaptic weights of associative neural network |
US11847553B2 (en) * | 2018-06-14 | 2023-12-19 | International Business Machines Corporation | Parallel computational architecture with reconfigurable core-level and vector-level parallelism |
FR3084503B1 (en) * | 2018-07-26 | 2020-10-16 | Thales Sa | SYNAPTIC CHAIN INCLUDING SPINTRONIC RESONATORS BASED ON THE SPIN DIODE EFFECT AND NEURON NETWORK INCLUDING SUCH SYNAPTIC CHAIN |
JP6789576B2 (en) * | 2018-08-02 | 2020-11-25 | 株式会社フローディア | Multiply-accumulate arithmetic unit |
CN109102071B (en) * | 2018-08-07 | 2020-12-11 | 中国科学院微电子研究所 | Neuron circuit and neural network circuit |
CN109255437B (en) * | 2018-08-17 | 2019-06-14 | 郑州轻工业学院 | A kind of memristor nerve network circuit of flexibly configurable |
CN109948792B (en) * | 2018-08-30 | 2022-08-02 | 东北师范大学 | Simulation association learning circuit based on transistor and memristor and control method |
CN109447250B (en) * | 2018-09-14 | 2020-07-10 | 华中科技大学 | Artificial neuron based on battery effect in memristor |
CN109472348B (en) * | 2018-10-23 | 2022-02-18 | 华中科技大学 | LSTM neural network system based on memristor cross array |
CN109635941B (en) * | 2018-11-06 | 2022-02-15 | 华中科技大学 | Maximum pooling processor based on 1T1R memory device |
CN109711537B (en) * | 2018-11-30 | 2020-09-11 | 郑州轻工业学院 | Prediction circuit based on memristor neural network |
CN109978159B (en) * | 2019-03-19 | 2023-03-24 | 常州大学 | Simple Fitzhugh-Nagumo neuron circuit |
CN110097182B (en) * | 2019-04-10 | 2023-03-24 | 常州大学 | Three-dimensional Hopfield neural network model realization circuit controlled by nerve activation gradient lambda |
CN110135571B (en) * | 2019-04-11 | 2023-09-29 | 上海集成电路研发中心有限公司 | Single-layer linear neural network based on cell synaptic structure |
US11694070B2 (en) | 2019-05-07 | 2023-07-04 | Hrl Laboratories, Llc | Bipolar all-memristor circuit for in-memory computing |
TWI760746B (en) * | 2019-05-22 | 2022-04-11 | 力旺電子股份有限公司 | Control circuit for multiply accumulate circuit of neural network system |
CN110298435B (en) * | 2019-05-30 | 2022-10-18 | 长安大学 | Memristor type full-domain value BP neural network circuit |
CN110443356B (en) * | 2019-08-07 | 2022-03-25 | 南京邮电大学 | Current type neural network based on multi-resistance state memristor |
CN110991629B (en) * | 2019-11-02 | 2023-05-02 | 复旦大学 | Neuron circuit based on memristor |
WO2021092899A1 (en) | 2019-11-15 | 2021-05-20 | 江苏时代全芯存储科技股份有限公司 | Neural-like circuit and operation method |
CN111630528B (en) | 2019-11-15 | 2023-08-29 | 北京时代全芯存储技术股份有限公司 | Neural circuit and operation method |
CN111656369B (en) * | 2019-11-15 | 2023-09-01 | 北京时代全芯存储技术股份有限公司 | Neural circuit and operation method |
WO2021092901A1 (en) | 2019-11-15 | 2021-05-20 | 江苏时代全芯存储科技股份有限公司 | Neural-like circuit and operation method |
CN113646912B (en) * | 2020-02-27 | 2024-02-27 | Tdk株式会社 | Arithmetic circuit and neuromorphic device |
CN111582460B (en) * | 2020-05-21 | 2023-04-14 | 中国人民解放军国防科技大学 | Hidden layer neuron self-adaptive activation method and device and terminal equipment |
CN111652362B (en) * | 2020-06-04 | 2023-06-06 | 苏州大学 | Neuron activation function circuit and neuron circuit based on CMOS |
CN111898748B (en) * | 2020-07-31 | 2022-07-05 | 四川大学 | Fractional-order memristive bridge circuit, and synaptic circuit and neuron circuit formed by same |
CN112651495B (en) * | 2020-12-16 | 2021-10-12 | 郑州轻工业大学 | Neural network circuit for emotion homeostasis adjustment and associative memory |
CN113095497B (en) * | 2021-05-06 | 2023-04-07 | 安徽大学 | Finite time synchronization method and device for fractional order quaternary memristor neural network |
CN113344194B (en) * | 2021-05-31 | 2024-02-13 | 郑州轻工业大学 | Operational condition reflecting circuit based on memristor |
CN113469334B (en) * | 2021-06-29 | 2022-09-13 | 中国地质大学(武汉) | Memristor recurrent neural network circuit |
CN113408719B (en) * | 2021-06-30 | 2023-02-10 | 华中科技大学 | Multilevel associative memory circuit based on non-molten state phase change device |
CN114219082B (en) * | 2021-12-29 | 2024-05-14 | 杭州电子科技大学 | Improved memristor neuron circuit |
CN115456157B (en) * | 2022-11-11 | 2023-02-07 | 华中科技大学 | Multi-sense interconnection memory network circuit based on memristor |
Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105224986A (en) * | 2015-09-29 | 2016-01-06 | 清华大学 | Based on the deep neural network system of memory resistor |
CN105701541A (en) * | 2016-01-13 | 2016-06-22 | 哈尔滨工业大学深圳研究生院 | Circuit structure based on memristor pulse nerve network |
Family Cites Families (1)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
US8996431B2 (en) * | 2012-07-25 | 2015-03-31 | Hrl Laboratories, Llc | Spike domain neuron circuit with programmable kinetic dynamic, homeostatic plasticity and axonal delays |
-
2016
- 2016-12-30 CN CN201611258488.3A patent/CN106815636B/en active Active
Patent Citations (2)
Publication number | Priority date | Publication date | Assignee | Title |
---|---|---|---|---|
CN105224986A (en) * | 2015-09-29 | 2016-01-06 | 清华大学 | Based on the deep neural network system of memory resistor |
CN105701541A (en) * | 2016-01-13 | 2016-06-22 | 哈尔滨工业大学深圳研究生院 | Circuit structure based on memristor pulse nerve network |
Also Published As
Publication number | Publication date |
---|---|
CN106815636A (en) | 2017-06-09 |
Similar Documents
Publication | Publication Date | Title |
---|---|---|
CN106815636B (en) | A kind of neuron circuit based on memristor | |
Ertuğrul et al. | A detailed analysis on extreme learning machine and novel approaches based on ELM | |
Li et al. | Diagnosis and location of the open-circuit fault in modular multilevel converters: An improved machine learning method | |
CN110383282A (en) | The system and method calculated for mixed signal | |
CN110097182B (en) | Three-dimensional Hopfield neural network model realization circuit controlled by nerve activation gradient lambda | |
KR20190114212A (en) | Single Column Memristor Crossbar and CMOS Activation Function Circuit for Binarized Neural Networks | |
CN109460818A (en) | A kind of multilayer neural network design method based on memristor bridge and array | |
Bhaduri et al. | Spiking neural classifier with lumped dendritic nonlinearity and binary synapses: a current mode VLSI implementation and analysis | |
Konal et al. | Electronically controllable memcapacitor emulator employing VDCCs | |
Zhou et al. | Gradient-based neuromorphic learning on dynamical RRAM arrays | |
KR19980023768A (en) | Low Voltage MOSFET Controlled Multiplier | |
KR940004429B1 (en) | Variable resistive mosfet analog multipler | |
Song et al. | Energy-efficient high-accuracy spiking neural network inference using time-domain neurons | |
Oh et al. | 1.2-mw online learning mixed-mode intelligent inference engine for low-power real-time object recognition processor | |
CN105913120A (en) | Transconductance operation amplification circuit and cell nerve network | |
RU2547225C1 (en) | Multidigit logical element of cyclic shift | |
RU2547233C1 (en) | Logical element of loose comparison for inequality of two multivalued variables | |
RU2553071C1 (en) | Multi-valued logical gate of reverse end-around shift | |
RU2546085C1 (en) | LOGICAL COMPARISON ELEMENT OF k-DIGIT VARIABLE WITH THRESHOLD VALUE | |
RU2554557C1 (en) | Multiple-valued logical element of reverse cyclic shift | |
KR940004430B1 (en) | Mosfet resistive control type multiply operator | |
Kant et al. | 0.5 V Sinh-domain design of activation functions and neural networks | |
CN109255437B (en) | A kind of memristor nerve network circuit of flexibly configurable | |
Li et al. | A 0.7 v low-power fully programmable gaussian function generator for brain-inspired gaussian correlation associative memory | |
WO2018072066A1 (en) | Pulse-based neural circuit |
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 |