CN109255430A - A kind of neuron coding circuit - Google Patents
A kind of neuron coding circuit Download PDFInfo
- Publication number
- CN109255430A CN109255430A CN201810762817.0A CN201810762817A CN109255430A CN 109255430 A CN109255430 A CN 109255430A CN 201810762817 A CN201810762817 A CN 201810762817A CN 109255430 A CN109255430 A CN 109255430A
- Authority
- CN
- China
- Prior art keywords
- signal
- pulse
- output
- frequency
- coding circuit
- 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
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
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- 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
-
- 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
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
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)
- General Engineering & Computer Science (AREA)
- Data Mining & Analysis (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Computational Linguistics (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Software Systems (AREA)
- Neurology (AREA)
- Amplifiers (AREA)
- Analogue/Digital Conversion (AREA)
Abstract
A kind of neuron coding circuit, belongs to technical field of integrated circuits.In the neuron coding circuit, the analog voltage signal of traditional neural network output exports stable control voltage signal after signal acquisition module is filtered；Control voltage signal and the current signal exported through adjustment of sensitivity unit are input to frequency coder, after the processing of frequency coder quantization encoding, the analog signal of output frequency and control voltage signal presentation sigmoid functional relation；Pulse-width modulator receives the analog signal and pulse-width control signal of frequency coder output, and output frequency is identical as analog signal, pulse width and the positively related pulse signal of pulse-width control signal；The pulse signal that module receives pulse-width modulator output is provided in pulse, and when film potential detection module granting signal Ready is enabled, which is passed to impulsive neural networks.Neuron coding circuit of the present invention has many advantages, such as that circuit structure is simple, frequency coverage is wide, low in energy consumption.
Description
Technical field
The invention belongs to technical field of integrated circuits, and in particular to a kind of neuron coding circuit.
Background technique
Neural network is the mathematical model or computation model of a kind of structure and function of mimic biology neural network, main root
It is coupled according to a large amount of artificial neuron and calculates.In general, inside artificial neural network can change on the basis of external information
Structure is a kind of Adaptable System.
Traditional neural network is a kind of operational model, is coupled to each other and is constituted by a large amount of node.Each node on behalf one
The specific output function of kind, referred to as excitation function；Connection between every two node all represents one for by the connection signal
Weighted value, referred to as weight is equivalent to the memory of artificial neural network.The output of network be then according to the connection type of network,
The difference of weighted value and excitation function and it is different.And network itself is approached certain algorithm of nature or function, either
A kind of expression to logic strategy.
And in impulsive neural networks, neuron transmitting is pulse, and each neuron has a film potential, works as neuron
When receiving input pulse, film potential can change.And when the film potential of neuron reaches a threshold value, arteries and veins will be sent
Punching, is transferred to next neuron.One pulse train of pulse shaping one by one of impulsive neural networks transmitting, and single arteries and veins
Time interval between punching be it is uncertain, temporal information is contained in pulse train, this is that traditional neural network is unable to table
It reaches.In addition, each neuron in impulsive neural networks only when receiving pulse signal, just will do it calculating, with biography
The troublesome calculation amount of system neural network is compared, and power consumption is lower, and calculating speed is faster.
However, impulsive neural networks do not find a preferable training algorithm temporarily at present, it is deep in particular for training
When layer network.And traditional artificial neural network has obtained abundant development by the mathematical tools such as statistics and optimization, what training obtained
Effect is more preferable than impulsive neural networks.It is thus typically necessary to trained artificial neural network is mapped as impulsive neural networks,
To avoid direct training pulse neural network.But at present when artificial neural network is mapped as impulsive neural networks, usually
Sampled using analog-digital converter, digital processing unit is encoded, the analog input signal of artificial neural network is converted to
Pulse is to realize the identification to impulsive neural networks；The circuit structure is complicated, and power consumption is big.
Summary of the invention
It is an object of the present invention to which the problem difficult for the impulsive neural networks training mentioned in background technique, proposes
A kind of neuron coding circuit, the circuit to form pulse nerve by encoding the value of a neuron in traditional neural network
The identifiable pulse signal of network, circuit structure is simple, low in energy consumption.
To achieve the above object, The technical solution adopted by the invention is as follows:
A kind of neuron coding circuit, which is characterized in that including signal acquisition module, pulse code module, film potential inspection
It surveys module and module is provided in pulse, the pulse code module includes frequency coder, pulse-width modulator and adjustment of sensitivity list
Member；
The input terminal of the signal acquisition module is connected with traditional neural network output end, and output end is coupled to frequency coding
Device；The sensitivity control terminal of frequency coder is connected with adjustment of sensitivity unit, and output end is connected with pulse-width modulator；The arteries and veins
The output end of module couples to pulse code module is provided in punching, and control terminal is connected with film potential detection module；Module is provided in pulse
Output end be connected with impulsive neural networks；
The analog voltage signal of traditional neural network output exports stable control after signal acquisition module is filtered
Voltage signal Vin processed；Sensitivity control signal K is after adjustment of sensitivity unit, output current signal；Control voltage signal Vin
It is input to frequency coder with the current signal exported through adjustment of sensitivity unit, after the processing of frequency coder quantization encoding,
The analog signal f (K, Vin) of sigmoid functional relation is presented in output frequency and control voltage signal Vin, and wherein K is sensitivity
Control signal (analog signal is the sinusoidal signal with frequency information)；Pulse-width modulator receives the simulation of frequency coder output
Signal f (K, Vin) and pulse-width control signal Vref, output frequency identical, pulse width and pulsewidth with analog signal f (K, Vin)
Control the positively related pulse signal of signal Vref；Value (the pulse code mould of the film potential detection module real-time detection film potential
The pulse signal of block output is to capacitor charging), when film potential is greater than or equal to threshold voltage, signal Ready will be provided and enabled；
The pulse signal that module receives pulse-width modulator output is provided in the pulse, and when granting signal Ready is enabled, by the pulse
Signal passes to impulsive neural networks.
Further, the sensitivity control signal K is an analog voltage signal, and range of voltage values is 0V~VCC,
Middle VCCIt is supply voltage value.
Further, the current value and sensitivity control signal K for the current signal that the adjustment of sensitivity unit exports are just
It is related.
Further, the range of voltage values of the received control voltage signal Vin of frequency coder is 0~VCC；Frequency coding
The current value positive of the current signal of the maximum frequency and adjustment of sensitivity unit output of the analog signal f (K, Vin) of device output
It closes.
Further, the pulse-width control signal Vref is an analog voltage signal, and range of voltage values is 0V~VCC,
Middle VCCIt is supply voltage value.
Further, in the pulse code module, different sensitivity control signals can be selected according to application scenarios
K, the different pulse signal of output maximum frequency.
Further, in the pulse code module, different pulse-width control signals can be selected according to application scenarios
Vref, the different pulse signal of output pulse width.
Further, the film potential detection module can set different threshold voltages according to application scenarios.
Further, the signal acquisition module receives the analog voltage signal of traditional neural network output, and extracts
The DC component of the analog voltage signal, the high fdrequency component for filtering out the analog voltage signal export stable control voltage signal
Vin。
Further, the adjustment of sensitivity unit is used to control the analog signal f's (K, Vin) of frequency coder output
Maximum frequency and minimum frequency.
Further, the pulse-width modulator is used to the analog signal f (K, Vin) that frequency coder exports becoming pulse
Signal and the duty ratio for changing pulse signal by the size of change pulse-width control signal Vref.
Further, the adjustment of sensitivity unit has a voltage input end and a current regulating signal output
End, the voltage signal received is converted into can control the current regulating signal of coding module sensitivity.
Further, the frequency coder has a voltage control signal input port, a current regulating signal
Input port and an analog signal output mouth.Voltage control signal input port and current regulating signal input port point
It Jie Shou not the voltage signal of signal acquisition module output and the current regulating signal of adjustment of sensitivity unit output and to receiving
Signal encoded, pass through the analog signal f (K, Vin) after analog signal output mouth exports coding.
Further, the pulse-width modulator have an input end of analog signal mouth, a pulse-width controlled port and
One pulse signal output end mouth, input end of analog signal mouth can receive frequency coder output analog signal f (K,
Vin the pulse signal of pulsewidth), and according to the size output phase of pulse-width control signal Vref is answered.
Further, it includes providing receiver port and switching circuit that module is provided in the pulse.The granting signal
Signal, response when simulation actual nerve member is stimulated are provided in the pulse that receiving port is used to receive pulse-width modulator output；
Whether the switching circuit, the pulse signal for being controlled and received according to the state for providing signal transmit backward.
Further, the switching circuit switchs closing when providing invalidating signal, and the pulse signal received is not backward
Transmitting.
Further, the switching circuit provide signal Ready it is enabled when switch open, the pulse signal received to
After pass to impulsive neural networks.
The invention has the benefit that
The invention proposes a kind of neuron coding circuit, the circuit is by by neuron in traditional neural network
Value coding forms the identifiable pulse signal of impulsive neural networks, compared to the method for conventionally employed digital processing, has circuit
The advantages that structure is simple, frequency coverage is wide, low in energy consumption.
Detailed description of the invention
Fig. 1 is a kind of circuit structure diagram of neuron coding circuit provided by the invention；
Fig. 2A is a kind of circuit structure diagram of neuron coding circuit provided in an embodiment of the present invention；
Fig. 2 B is the concrete structure schematic diagram of low-pass filter in Fig. 2A coding circuit；
Fig. 2 C is the schematic diagram of Fig. 2A coding circuit ring oscillator；
Fig. 2 D is a kind of electrical block diagram for realizing basic unit in Fig. 2 C；
Fig. 2 E is a kind of electrical block diagram for realizing speed access in Fig. 2 D；
Fig. 2 F is a kind of electrical block diagram for realizing gain control in Fig. 2 D；
Fig. 2 G is a kind of electrical block diagram for realizing pulse-width regulated in Fig. 2A coding circuit；
Fig. 2 H is the structural schematic diagram that module is provided in pulse in Fig. 2A coding circuit；
Fig. 3 is the output frequency of pulse signal and control in a kind of neuron coding circuit provided in an embodiment of the present invention
The tuning curve of voltage signal Vin and sensitivity control signal K.
Specific embodiment
In order to which the purposes, technical schemes and advantages of the embodiment of the present invention are more clearly understood, with reference to the accompanying drawing to this
Invention embodiment is further described in detail.Here, illustrative embodiments and their description of the invention are for explaining this hair
It is bright but not as a limitation of the invention.For those of ordinary skill in the art, before not making the creative labor
It puts, is also possible to obtain other drawings based on these drawings.
As shown in Figure 1, being a kind of circuit structure diagram of neuron coding circuit provided by the invention；Including signal acquisition mould
Module is provided in block, pulse code module, film potential detection module and pulse.
Signal acquisition module is connected with the output end of traditional neural network, for acquiring the simulation of traditional neural network output
Voltage signal, and the voltage of traditional neural network output is filtered, export stable control voltage signal Vin.
Pulse code module couples receive control voltage signal Vin, then to control to the output end of signal acquisition module
Voltage signal Vin carries out quantization encoding, and the pulse that sigmoid functional relation is presented in output frequency and control voltage signal Vin is believed
Number.
The value of film potential detection module real-time detection film potential will be provided when film potential is greater than or equal to threshold voltage
Signal Ready is enabled.
Pulse provides module couples to the output end of pulse code module, receives the pulse letter of pulse code module output
Number, and when granting signal Ready is enabled, pulse signal is passed into impulsive neural networks.
The pulse code module includes adjustment of sensitivity unit, frequency coder and pulse-width modulator.
Adjustment of sensitivity unit receiving sensitivity controls signal K, output current signal；Frequency coder receives control voltage
Sigmoid letter is presented in the current signal of signal Vin and the output of adjustment of sensitivity unit, output frequency and control voltage signal Vin
The analog signal f (K, Vin) of number relationship, wherein K is sensitivity control signal；Pulse-width modulator receives frequency coder output
Analog signal f (K, Vin) and pulse-width control signal Vref, output frequency and analog signal f (K, Vin) identical, pulse width with
The positively related pulse signal of pulse-width control signal Vref.
The embodiment of the invention provides a kind of neuron coding circuits, including signal acquisition module, pulse code module, film
Module is provided in potentiometric detection module and pulse；
The signal acquisition module has a voltage input end and a voltage output end, can be used in receiving input
Voltage, and input voltage is smoothed, extract stable DC component；
The pulse code module, is couple to the output end of the signal acquisition module, can be realized pulse output, is used in combination
In carrying out coded treatment to input voltage, the Function Fitting of output frequency and input voltage is realized；
Module is provided in the pulse, the output end of the pulse code module is couple to, for the pulse signal to input
It is controlled, mechanism is provided in the pulse of simulation actual nerve member.
In the present embodiment, the voltage input end of the signal acquisition module is couple to the output end of traditional neural network, uses
In the analog voltage signal of acquisition traditional neural network output.
In the present embodiment, the signal acquisition module can allow for the voltage signal of input to have certain fluctuation.
In the present embodiment, the signal acquisition module receives the analog voltage signal of traditional neural network output, and extracts
Out the DC component of the analog voltage signal, filter out the high fdrequency component of the analog voltage signal, export stable control voltage letter
Number Vin.
In the present embodiment, the pulse code module includes adjustment of sensitivity unit, frequency coder and pulse-width modulator；
The adjustment of sensitivity unit is used to control the maximum frequency of the analog signal f (K, Vin) of frequency coder output
And minimum frequency.
The frequency coder, the voltage signal for exporting to signal acquisition module carry out frequency coding, believe output
The frequency and control voltage signal Vin of number f (K, Vin) shows sigmoid functional relation.
In the present embodiment, the pulse-width modulator is used to the analog signal f (K, Vin) that frequency coder exports becoming arteries and veins
Signal and the size by changing pulse-width control signal Vref are rushed to change the duty ratio of pulse signal.
In the present embodiment, the adjustment of sensitivity unit has a voltage input end and a current regulating signal output
End, the voltage signal received is converted into can control the current regulating signal of coding module sensitivity.
In the present embodiment, the frequency coder has a voltage control signal input port, a current regulation letter
Number input port and an analog signal output mouth.Voltage control signal input port and current regulating signal input port
The voltage signal of signal acquisition module output and the current regulating signal of adjustment of sensitivity unit output are received respectively and to reception
To signal encoded, pass through the analog signal f (K, Vin) after analog signal output mouth exports coding.
In the present embodiment, the pulse-width modulator have an input end of analog signal mouth, a pulse-width controlled port with
And a pulse signal output end mouth, input end of analog signal mouth can receive frequency coder output analog signal f (K,
Vin the pulse signal of pulsewidth), and according to the size output phase of pulse-width control signal Vref is answered.
In the present embodiment, it includes providing receiver port and switching circuit that module is provided in the pulse.The granting letter
Signal, sound when simulation actual nerve member is stimulated are provided in the pulse that number receiving port is used to receive pulse-width modulator output
It answers；Whether the switching circuit, the pulse signal for being controlled and received according to the state for providing signal transmit backward.
In the present embodiment, the switching circuit switchs closing when providing invalidating signal, the pulse signal received not to
After transmit.
In the present embodiment, switching circuit switch when providing signal Ready and enabling is opened, the pulse signal received
Impulsive neural networks are passed to backward.
Fig. 2A is a kind of circuit structure diagram of neuron coding circuit provided in an embodiment of the present invention；Each mould of the circuit
Block can use different circuit structures according to actual needs.Wherein, signal acquisition module, pulse code module and pulse granting
Low-pass filter, annular voltage controlled oscillator and transmission gate is respectively adopted to realize in module.Specifically, traditional neural network exports
Analog voltage signal 1, filter out the high fdrequency component in voltage signal 1 by low-pass filter, the stable control electricity of output one
Press signal Vin to 2；From 6 one sensitivity control signal K of input, realizes that electric current is mutually added and subtracted by differential pair tube, generate one 7
A current signal；Control voltage signal Vin2 and the current signal 7 exported through differential pair tube are input to ring oscillator, annular vibration
Control of the electric current by signal at 2,7 liang for swinging device, in the frequency controlled signal f in control voltage signal Vin and K of 3 outputs one
(K, Vin)；Voltage comparator compares the size of 3 and 8 voltage values, output pulse width pulse signal relevant to 8 voltages to 4；Transmission
Door is opened or closed according to 9 state realizes that 4 to 5 on-off controls, i.e., whether information is transmitted backward.
Fig. 2 B is the concrete structure schematic diagram of low-pass filter in Fig. 2A coding circuit, and the present embodiment signal acquisition module is adopted
With three rank passive low ventilating filters as shown in Figure 2 B；Three rank passive low ventilating filter structures are simple, and can preferably filter out
High dither in input signal.1 receives the analog voltage signal of traditional neural network output, exported at 2 one it is stable
Control voltage signal Vin, transfer function are as follows:
In formula, s is the complex frequency of Laplace transform, according to theory analysis and practical engineering experience it is found that C1 is long-range
In C2, C3, R1 is greater than R3, can simplify based on formula (1) are as follows:
In formula (2),
The transfer function of whole system can be made to become rationally, to export more steady by adjusting the value of R1, R3, C1, C2, C3
Fixed control voltage signal Vin.
Fig. 2 C is the schematic diagram of Fig. 2A coding circuit ring oscillator；Wherein 2,3,6 be respectively low-pass filter output
Control voltage signal Vin, oscillator output signal f (K, Vin) and sensitivity control signal K.Basic unit includes in Fig. 2 C
Difference amplifier for the differential pair tube of adjustment of sensitivity and for constituting oscillator.Amplifier operation has in different frequency
Different phase shifts becomes entire negative-feedback just if the output of amplifier phase shift in high frequency itself is too big
, then oscillation will occur.According to " Barkhausen criterion ":
|H(jω0)|≥1 (3)
∠H(jω0)=180 ° (4)
In formula, j ω0For complex frequency.So in angular frequency0Place, closed loop gain approach are infinitely great.With this condition, circuit
By its own in ω0The noise component(s) at place infinitely amplifies.Since a pole at most contributes 90 ° of phase shift (in frequency infinity
Place), if to meet Barkhausen criterion, at least need three inverting amplifiers to be sequentially connected in series cyclization, this just constitutes most base
This ring oscillator.It is exactly the ring oscillator that three-level inverting amplifier is constituted used by the present embodiment, but wants circuit
Starting of oscillation must also meet loop gain greater than 1 simultaneously.The open-loop transmission function of three-level sign-changing amplifier:
In formula, A0Indicate the low-frequency gain of each amplifier, ωpIndicate the output pole of each amplifier.Only in frequency
Circuit just vibrates when relevant phase shift is equal to 180 °, that is, 60 ° every grade.The frequency vibrated can be given by:
In formula, ωoscIt indicates that the angular frequency in oscillation occurs, every grade of minimum voltage gain must make loop gain in frequency
Rate ωoscPlace is equal to 1:
It is obtained by formula (7) and (8)
A0=2 (9)
That is, the low-frequency gain of every grade of circuit of three-level annular oscillator requirement is 2, frequency of oscillation isWith
The increase of amplitude, circuit gradually work arriving signal state.Assuming that every grade of delay is TD, then the cycle of oscillation generated is
6TD.Frequency of oscillation is by original1/6T is moved on toD, and so on, the frequency f of N grades of oscillatorsosc=(2NTD)-1.Therefore,
Change TDJust change frequency of oscillation.
Fig. 2 D is a kind of electrical block diagram for realizing basic unit in Fig. 2 C；The present embodiment changes using interpolation method
Be delayed TD, each basic unit is made of a fast path and a slow path, and two-way output is added, their gain is by controlling
Voltage signal Vin is adjusted in the opposite direction respectively.A kind of extreme case of control voltage signal Vin is: only fast path conducting, slowly
Path shutdown, generates maximum frequency of oscillation.Opposite, in the case where another extreme, only slow path conducting, fast path
Shutdown, generates the smallest frequency of oscillation.If control voltage signal Vin fall in two it is extreme between, each path is partially led
Leading to and being always delayed is the weighted sum of the two delay, then generates intermediate frequency.Meanwhile the maximum gain in fast path and slow path is adjusted
Control of the range by K, therefore responding to control voltage signal Vin for output frequency can be changed by changing the value of K
Sensitivity, fit the sigmoid function of different center slopes.
Fig. 2 E is a kind of electrical block diagram for realizing speed access in Fig. 2 D；Every access can simply use difference
To realization, gain is controlled by tail current.Wherein MN1, MN4 constitute fast access, and gain is controlled by tail current ISS1.MN2,
R1, MN3, R2 constitute slow access, and gain is controlled by tail current ISS2.MN1 is identical as MN2 size, and MN3 is identical as MN4 size,
So influence of the parasitic parameter of differential pair to delay itself be it is identical, due to joined in slow access resistance R1 and R2 introduce
Additional capacitance-resistance delay, so speed is slower than fast access, and only related to the resistance value of resistance.
Fig. 2 F is a kind of implementation for realizing K and control voltage signal Vin control gain in Fig. 2A, and difference is flowed through in K control
Divide the electric current ratio to pipe MP1, MP2, the overdrive voltage of MP1 reduces (increase) when 6 voltage increases (reduction), flows through MP1's
Electric current reduces IMP1Reduce (increase), due to ISS=IMP1+IMP2For definite value, so flowing through the electric current I of MP2MP2Increase (reduction),
Further, the electric current for flowing into 7 increases (reduction), generates a current regulating signal, can also generate electricity using other structures
Flow adjustment signal.Flow into 7 electric current IMP2MN8 and MN9 are separately flowed into proportion, by the control for controlling voltage signal Vin at 2
The size for flowing into MN8 and MN9 electric current changes in the opposite direction.The electric current of MN8 and MN9 is mirrored in proportion by current mirror
ISS1 and ISS2 further controls the gain of amplifier, changes delay, to change frequency.
In the present embodiment, when K value is fixed, with the raising of control voltage signal Vin, the electric current for flowing through MN8 increases, stream
The electric current for crossing MN9 reduces, and further, the electric current for being mirrored to ISS1 increases, and the electric current for being mirrored to ISS2 reduces, and slow access is gradually
It is converted to fast access, output frequency gradually increases, finally by most reaching the maximum frequency determined by K value.Such as 2 institute of curve in Fig. 3
Show.When K value increases, 7 electric current increase is flowed into, the adjustable total current of control voltage signal Vin increases, and unit voltage is adjusted
Range increases, and output frequency curve becomes more precipitous, and total adjustable range increases, as shown in curve 1 in Fig. 3.When K value reduces,
7 electric current reduction is flowed into, the adjustable total current of control voltage signal Vin reduces, and unit voltage adjustable range reduces, and output is frequently
Rate curve becomes gentler, and total adjustable range reduces, as shown in curve 3 in Fig. 3.
Fig. 2 G is a kind of electrical block diagram for realizing pulse-width regulated in Fig. 2A coding circuit；By comparing 3 and 8 voltages
Size, control 3 to 8 voltages it is high (low) when output height (low) level, to realize pulse-width regulated.
Fig. 2 H is the structural schematic diagram that module is provided in pulse in Fig. 2A coding circuit；Switching circuit used transmission gate as
Switch, also can be used increasingly complex circuit to realize switching circuit, to achieve the effect that better control switch.
It should be noted that term " includes " or any other variant thereof is intended to cover non-exclusive inclusion, thus
So that including the commodity of a series of elements or system not only includes those elements, but also other including being not explicitly listed
Element, or further include for this commodity or the intrinsic element of system.In the absence of more restrictions, by sentence
The element that " including one ... " limits, it is not excluded that there is also other in the commodity or system for including the element
Identical element.
The foregoing is merely presently preferred embodiments of the present invention, is not intended to limit the invention, it is all in spirit of the invention and
Within principle, any modification, equivalent replacement, improvement and so on be should all be included in the protection scope of the present invention.
Claims (8)
1. a kind of neuron coding circuit, which is characterized in that including signal acquisition module, pulse code module, film potential detection
Module is provided in module and pulse, and the pulse code module includes frequency coder, pulse-width modulator and adjustment of sensitivity unit；
The analog voltage signal of traditional neural network output exports stable control electricity after signal acquisition module is filtered
Press signal Vin；Sensitivity control signal K is after adjustment of sensitivity unit, output current signal；Control voltage signal Vin and warp
The current signal of adjustment of sensitivity unit output is input to frequency coder, after the processing of frequency coder quantization encoding, output
The analog signal of sigmoid functional relation is presented in frequency and control voltage signal Vin；It is defeated that pulse-width modulator receives frequency coder
Analog signal and pulse-width control signal Vref out, output frequency is identical as analog signal, pulse width and pulse-width control signal
Positively related pulse signal；The value of the film potential detection module real-time detection film potential, when film potential is greater than or equal to threshold value
When voltage, signal Ready will be provided and enabled；The pulse signal of the pulse granting module reception pulse-width modulator output, and
When granting signal Ready is enabled, which is passed into impulsive neural networks.
2. neuron coding circuit according to claim 1, which is characterized in that the sensitivity control signal K is one
Analog voltage signal, range of voltage values are 0~VCC, wherein VCCIt is supply voltage value.
3. neuron coding circuit according to claim 1, which is characterized in that the electricity of the adjustment of sensitivity unit output
The current value and sensitivity control signal K for flowing signal are positively correlated.
4. neuron coding circuit according to claim 1, which is characterized in that the received control electricity of frequency coder
The range of voltage values for pressing signal Vin is 0~VCC；The maximum frequency and adjustment of sensitivity list of the analog signal of frequency coder output
The current value of the current signal of member output is positively correlated.
5. neuron coding circuit according to claim 1, which is characterized in that the pulse-width control signal Vref is one
Analog voltage signal, range of voltage values are 0~VCC, wherein VCCIt is supply voltage value.
6. neuron coding circuit according to claim 1, which is characterized in that in the pulse code module, Ke Yigen
According to application scenarios, different sensitivity control signal K, the different pulse signal of output maximum frequency are selected.
7. neuron coding circuit according to claim 1, which is characterized in that in the pulse code module, Ke Yigen
According to application scenarios, different pulse-width control signal Vref, the different pulse signal of output pulse width are selected.
8. neuron coding circuit according to claim 1, which is characterized in that the film potential detection module can basis
Application scenarios set different threshold voltages.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810762817.0A CN109255430B (en) | 2018-07-12 | 2018-07-12 | Neuron coding circuit |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201810762817.0A CN109255430B (en) | 2018-07-12 | 2018-07-12 | Neuron coding circuit |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109255430A true CN109255430A (en) | 2019-01-22 |
| CN109255430B CN109255430B (en) | 2022-03-15 |
Family
ID=65052018
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201810762817.0A Active CN109255430B (en) | 2018-07-12 | 2018-07-12 | Neuron coding circuit |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109255430B (en) |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110022299A (en) * | 2019-03-06 | 2019-07-16 | 浙江天脉领域科技有限公司 | A kind of method of ultra-large distributed network computing |
| CN112766480A (en) * | 2021-03-05 | 2021-05-07 | 电子科技大学 | Neuron circuit |
| CN113537449A (en) * | 2020-04-22 | 2021-10-22 | 北京灵汐科技有限公司 | Data processing method based on impulse neural network, computing core circuit and chip |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101971166A (en) * | 2008-03-14 | 2011-02-09 | 惠普开发有限公司 | Neuromorphic circuit |
| US20120310871A1 (en) * | 2011-06-02 | 2012-12-06 | Hrl Laboratories, Llc | High-order time encoder based neuron circuit |
| CN104899640A (en) * | 2014-07-21 | 2015-09-09 | 徐志强 | Simulator and method for neural network |
| CN107092959A (en) * | 2017-04-07 | 2017-08-25 | 武汉大学 | Hardware friendly impulsive neural networks model based on STDP unsupervised-learning algorithms |
| US20180107918A1 (en) * | 2016-10-14 | 2018-04-19 | International Business Machines Corporation | Core utilization optimization by dividing computational blocks across cores |
-
2018
- 2018-07-12 CN CN201810762817.0A patent/CN109255430B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN101971166A (en) * | 2008-03-14 | 2011-02-09 | 惠普开发有限公司 | Neuromorphic circuit |
| US20120310871A1 (en) * | 2011-06-02 | 2012-12-06 | Hrl Laboratories, Llc | High-order time encoder based neuron circuit |
| CN104899640A (en) * | 2014-07-21 | 2015-09-09 | 徐志强 | Simulator and method for neural network |
| US20180107918A1 (en) * | 2016-10-14 | 2018-04-19 | International Business Machines Corporation | Core utilization optimization by dividing computational blocks across cores |
| CN107092959A (en) * | 2017-04-07 | 2017-08-25 | 武汉大学 | Hardware friendly impulsive neural networks model based on STDP unsupervised-learning algorithms |
Non-Patent Citations (5)
| Title |
|---|
| ANDRE GRUNING 等: "Spiking Neural Networks:Principles and Challenges", 《ESANN 2014 PROCEEDINGS》 * |
| GUO-SHENG YI 等: "Input-output relation and energy efficiency in the neuron with different spike threshold dynamics", 《FRONTIERS IN COMPUTATIONAL NEUROSCIENCE》 * |
| MARINA IGNATOV 等: "A memristive spiking neuron with firing rate Coding", 《FRONTIERS IN NEUROSCIENCE》 * |
| 赵健 等: "基于灵敏度分析的海洋油气资源BP神经网络预测模型的优化", 《海洋科学》 * |
| 陈清: "Spiking神经网络的动态特性对液体状态机的影响研究", 《中国优秀博硕士学位论文全文数据库（硕士）信息科技辑》 * |
Cited By (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110022299A (en) * | 2019-03-06 | 2019-07-16 | 浙江天脉领域科技有限公司 | A kind of method of ultra-large distributed network computing |
| CN113537449A (en) * | 2020-04-22 | 2021-10-22 | 北京灵汐科技有限公司 | Data processing method based on impulse neural network, computing core circuit and chip |
| WO2021213471A1 (en) * | 2020-04-22 | 2021-10-28 | 北京灵汐科技有限公司 | Spiking neural network-based data processing method, computing core circuit, and chip |
| CN113537449B (en) * | 2020-04-22 | 2024-02-02 | 北京灵汐科技有限公司 | Data processing method based on impulse neural network, calculation core circuit and chip |
| US11900243B2 (en) | 2020-04-22 | 2024-02-13 | Lynxi Technologies Co., Ltd. | Spiking neural network-based data processing method, computing core circuit, and chip |
| CN112766480A (en) * | 2021-03-05 | 2021-05-07 | 电子科技大学 | Neuron circuit |
| CN112766480B (en) * | 2021-03-05 | 2023-10-27 | 电子科技大学 | Neuron circuit |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109255430B (en) | 2022-03-15 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN109255430A (en) | A kind of neuron coding circuit | |
| Thanou et al. | Distributed average consensus with quantization refinement | |
| Sun et al. | Admissibility analysis for interval type-2 fuzzy descriptor systems based on sliding mode control | |
| CN107977037A (en) | A kind of low-dropout regulator and its control method | |
| You et al. | Attainability of the minimum data rate for stabilization of linear systems via logarithmic quantization | |
| Zheng et al. | Fractional stochastic resonance multi-parameter adaptive optimization algorithm based on genetic algorithm | |
| Garanayak et al. | A high-speed master-slave ADALINE for accurate power system harmonic and inter-harmonic estimation | |
| Wu et al. | Exponential consensus analysis for multiagent networks based on time-delay impulsive systems | |
| Li et al. | Distributed frequency estimation over sensor network | |
| Ma et al. | Global cooperative output regulation of linear multiagent systems with limited bandwidth | |
| CN109546660B (en) | Active power filter circuit based on neural sliding mode control strategy and control method | |
| Chen et al. | Asynchronous biphasic pulse signal coding and its CMOS realization | |
| Dam et al. | The dual parameterization approach to optimal least square FIR filter design subject to maximum error constraints | |
| EP3193448A1 (en) | Band-pass filter | |
| CN106374491B (en) | A kind of DC capacitor voltage control method of Active Power Filter-APF | |
| CN109245752B (en) | Adjusting circuit and analog switch | |
| Wang et al. | Nonlinear dynamic system identification using least squares support vector machine regression | |
| Reza Karimi et al. | Analysis and synthesis of control systems over wireless digital channels | |
| CN111726179A (en) | Received signal strength detection circuit and method | |
| CN109982422A (en) | A kind of adaptive wireless sensor network transmission power control method | |
| CN109540176A (en) | A kind of silicon micro-gyroscope Sigma Delta detection closed-loop control system structure and parameter tuning method | |
| Kabziński | Adaptive control of Duffing oscillator with unknown input gain | |
| JPH03163937A (en) | Decrease in noise of signal coded in a plurality of preconsideration levels and device therefor | |
| CN103066601B (en) | Hybrid active direct current filter control method based on self-adaptive linear neurons | |
| JPH10300801A (en) | Phase detecting filter 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 |