CN105095967B - A kind of multi-modal neuromorphic network core - Google Patents
A kind of multi-modal neuromorphic network core Download PDFInfo
- Publication number
- CN105095967B CN105095967B CN201510419465.5A CN201510419465A CN105095967B CN 105095967 B CN105095967 B CN 105095967B CN 201510419465 A CN201510419465 A CN 201510419465A CN 105095967 B CN105095967 B CN 105095967B
- Authority
- CN
- China
- Prior art keywords
- dendron
- unit
- modal
- network core
- aixs cylinder
- 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
Landscapes
- Feedback Control In General (AREA)
Abstract
The present invention provides a kind of multi-modal neuromorphic network core, including：Mode register, aixs cylinder input block, synapse weight memory cell, dendron unit and neuron computing unit.The multi-modal neuromorphic network core can both carry out artificial neural network computing, can also carry out impulsive neural networks computing, and can switch as required between artificial neural network operational mode and impulsive neural networks operational mode.
Description
Technical field
The present invention relates to a kind of neuromorphic network core.
Background technology
Neutral net is the computing system that a kind of mimic biology brain cynapse-neuronal structure carries out data processing, by dividing
The connection composition of calculate node and interlayer for multilayer.Each node simulates a neuron, performs some certain operations, such as
Activation primitive, the connection analog neuron cynapse between node, connects and is represented for the weighted value inputted from last layer node
Synapse weight.Neutral net has powerful non-linear, adaptive information disposal ability.
Neuron in artificial neural network is used as itself after the accumulated value from connection input is handled by the use of activation primitive
Output.Corresponding to different network topology structures, neuron models and learning rules, artificial neural network includes perceiving again
The tens of kinds of network models such as device, Hopfield networks, Boltzmann machine, it is possible to achieve diversified function, pattern-recognition,
Complex control, signal transacting and optimization etc. have application.Traditional artificial neural network data, which may be considered, to be passed through
The frequency information coding of neuron pulse, each layer neuron are serially run successively.The nerve of artificial Neural Network Simulation biology
System hlerarchy, but influence of the information processing architecture such as time serieses of matching cortex completely to study is the failure to, and make
For real biological cortex in processing information for, the study to information data is not independent static state, but over time
There is the contact of context.Impulsive neural networks are the new neural networks occurred in recent ten years, the third generation of being known as
Neutral net.Data in impulsive neural networks are encoded with the space time information of neuron pulse signal, the input and output of network with
And the information transmission between neuron shows as the pulse of neuron transmission and sends the temporal information of pulse, neuron needs same
Shi Binghang is run.Compared with traditional artificial neural network, impulsive neural networks are in information processing manner, neuron models, simultaneously
Row etc. has a relatively big difference, and the method for operation is closer to real biosystem.Impulsive neural networks application accurate timing
Pulse train nerve information is encoded and handled, this computation model comprising time calculating elements more biology explain
Property, it is the effective tool for carrying out complicated space time information processing, multi-modal information can be handled and information processing is more real-time.
But the discontinuity of the neuron models of impulsive neural networks, the complexity of space-time code, the uncertainty of network structure cause
It is difficult to mathematically complete to the overall description of network, therefore, it is difficult to build effective and general supervised learning algorithm, limitation
Its calculation scale and accuracy.
The content of the invention
In view of this, it is necessory to provide it is a kind of can carry out artificial neural networks or enter horizontal pulse nerve
The neuromorphic network core of network calculations.
A kind of multi-modal neuromorphic network core, including：Mode register, aixs cylinder input block, synapse weight storage are single
Member, dendron unit and neuron computing unit；
The mode register is connected with the aixs cylinder input block, dendron unit and neuron computing unit, control
Said units operate in artificial neural network pattern or impulsive neural networks pattern；
The aixs cylinder input block is connected with the dendron unit, is received and is stored aixs cylinder input；
The synapse weight memory cell is connected with the dendron unit, stores synapse weight matrix；
The dendron unit is connected with the neuron computing unit, including dendron multiplicaton addition unit and dendron summing elements,
When operating in artificial neural network pattern, the aixs cylinder input vector is sent into the dendron multiplicaton addition unit with synapse weight matrix and entered
Row multiply-add operation, when operating in impulsive neural networks pattern, the aixs cylinder input vector is sent into the tree with synapse weight matrix
Prominent summing elements carry out accumulating operation；
The neuron computing unit includes the first computing unit and the second computing unit, operates in artificial neural network mould
During formula, the multiply-add operation result that the dendron multiplicaton addition unit is sent is sent into the first computing unit progress artificial neural network fortune
Calculate, when operating in impulsive neural networks pattern, the accumulating operation result that the dendron summing elements are sent is sent into second meter
Calculate unit and carry out impulsive neural networks calculating.
Compared with prior art, multi-modal neuromorphic network core provided by the invention can both carry out artificial neural network
Computing, impulsive neural networks computing can also be carried out, and can be as required in artificial neural network operational mode and pulse god
Through switching between Network operation mode, real-time, multi-modal or complicated space-time signal can be carried out and calculate and can guarantee that calculating
Accuracy.
Brief description of the drawings
Fig. 1 is in the hybrid system of the artificial neural network that first embodiment of the invention provides and impulsive neural networks
Basic computational ele- ment structure chart.
Fig. 2 is the cascaded structure schematic diagram of the present invention
Fig. 3 is the parallel-connection structure schematic diagram of the present invention.
Fig. 4 is the parallel organization schematic diagram of the present invention.
Fig. 5 is the learning structure schematic diagram of the present invention.
Fig. 6 is the feedback arrangement schematic diagram of the present invention.
Fig. 7 is that elementary layer level structure schematic diagram is calculated in hybrid system provided by the invention.
Fig. 8 is the hybrid system of artificial neural network provided by the invention and impulsive neural networks.
Fig. 9 is the signal that the numerical quantities of artificial neural network output are converted to pulse train in second embodiment of the invention
Figure.
Figure 10 is that the frequency coding pulse train in second embodiment of the invention by impulsive neural networks output is converted to number
Value amount schematic diagram.
Figure 11 is that the Population Coding pulse train in second embodiment of the invention by impulsive neural networks output is converted to number
Value amount schematic diagram.
Figure 12 is that the time encoding pulse train in second embodiment of the invention by impulsive neural networks output is converted to number
Value amount schematic diagram.
Figure 13 is that the binary-coding pulse train in second embodiment of the invention by impulsive neural networks output is converted to number
Value amount schematic diagram.
Figure 14 is the multi-modal neuromorphic network nuclear structure block diagram that third embodiment of the invention provides.
When Figure 15 is that the multi-modal neuromorphic network core that third embodiment of the invention provides operates in artificial neural network
Structured flowchart.
Figure 16 is the operational flow diagram of one time step of multi-modal neuromorphic network core under artificial neural network pattern.
When Figure 17 is that the multi-modal neuromorphic network core that third embodiment of the invention provides operates in impulsive neural networks
Structured flowchart.
Figure 18 is the operational flow diagram of one time step of multi-modal neuromorphic network core under impulsive neural networks pattern.
Figure 19 is the hybrid system of the artificial neural network that fourth embodiment of the invention provides and impulsive neural networks.
Figure 20 is fourth embodiment of the invention routing nodes structured flowchart.
Figure 21 is fourth embodiment of the invention routing data inclusion composition.
Figure 22 is fourth embodiment of the invention routing nodes workflow diagram.
Main element symbol description
| Hybrid system | 100 | Mode register | 211 |
| Basic computational ele- ment | 110 | Aixs cylinder input block | 212 |
| First basic computational ele- ment | 110a | Synapse weight memory cell | 213 |
| Second basic computational ele- ment | 110b | Dendron unit | 214 |
| Unit | 111 | Dendron multiplicaton addition unit | 214a |
| Neuron | 115 | Dendron summing elements | 214b |
| Cynapse | 116 | Neuron computing unit | 215 |
| Composite computing unit | 120 | First computing unit | 215a |
| Series connection recombiner unit | 120a | Second computing unit | 215b |
| Recombiner unit in parallel | 120b | Dendron expands memory cell | 2151 |
| Parallel composition unit | 120c | Parameter storage unit | 2152 |
| Learn recombiner unit | 120d | Integrate leak calculation unit | 2153 |
| Feedback complex unit | 120e | Trigger signal counter | 216 |
| Hybrid system | 200 | Controller | 217 |
| Neuromorphic network core | 210 | Routing node | 220 |
| Multi-modal neuromorphic network core | 210a |
Following embodiment will combine above-mentioned accompanying drawing and further illustrate the present invention.
Embodiment
Multi-modal neuromorphic network core provided by the invention is made below in conjunction with the accompanying drawings and the specific embodiments further
Detailed description.
First embodiment of the invention provides a kind of hybrid system 100 of artificial neural network and impulsive neural networks,
Including at least two basic computational ele- ments 110, at least two basic computational ele- ment 110, at least one is ANN
Network computing unit, artificial neural networks are undertaken, at least one is impulsive neural networks computing unit, undertakes pulse nerve net
Network is calculated, and at least two basic computational ele- ment 110 is connected with each other according to topological structure, realizes neural computing work(jointly
Energy.
Fig. 1 is referred to, an at least artificial neural networks unit calculates with an at least impulsive neural networks
Unit is considered as an independent neutral net respectively, and the neutral net includes multiple neurons 115, the plurality of neuron
Connected between 115 by cynapse 116, form single or multiple lift structure.Synapse weight represents postsynaptic neuron and receives the presynaptic
The weighted value of neuron output.
An at least impulsive neural networks computing unit is used to perform impulsive neural networks calculating to the data received.
The data at least transmitted between the input data of an impulsive neural networks computing unit, output data and neuron 115 are
Spike sequence, the model of the neuron 115 at least described in an impulsive neural networks computing unit is based on spike arteries and veins
Rush calculate neuron models, can be but be not limited to leakage-integration-fire model, spike response model and
At least one of Hodgkin-Huxley models.
An at least artificial neural networks unit is used to perform artificial neural networks to the data received.
The data at least transmitted between the input data of an artificial neural networks unit, output data and neuron 115 are
Numerical quantities.At least artificial neural networks unit is further according to neuron models, network structure, learning algorithm
Difference, can be perceptron neural network computing unit, BP neural network computing unit, Hopfield neural computing lists
Member, adaptive resonance theory neural computing unit, depth conviction neural computing unit and convolutional neural networks calculate
At least one of unit.
An at least artificial neural networks unit and at least an impulsive neural networks computing unit Topology connection
To form a complex neural network computing unit.
The topological structure of the Topology connection includes cascaded structure, parallel-connection structure, parallel organization, learning structure and feedback
At least one of structure.
Fig. 2 is referred to, described two basic computational ele- ments 110 are connected in series to form a series connection composite computing unit
120a.Described two basic computational ele- ments 110 are respectively the first basic computational ele- ment 110a and the second basic computational ele- ment 110b,
The output end of the first basic computational ele- ment 110a connects the second basic computational ele- ment 110b input.Described first is basic
In computing unit 110a and the second basic computational ele- ment 110b, one is artificial neural network computing unit, and another is pulse
Neural computing unit.System input first passes around the first basic computational ele- ment 110a processing, and the result after processing is used as the
Two basic computational ele- ment 110b input, the result after the second basic computational ele- ment 110b processing export for system.
Fig. 3 is referred to, described two basic basic computational ele- ments 110 are connected in parallel to form a recombiner unit in parallel
120b.Described two basic computational ele- ments 110 are respectively the first basic computational ele- ment 110a and the second basic computational ele- ment 110b,
Input connection the second basic computational ele- ment 110b of first basic computational ele- ment 110a input, described first
Basic computational ele- ment 110a output end connects the output end of the second basic computational ele- ment 110b.First basic calculating
In unit 110a and the second basic computational ele- ment 110b, one is artificial neural network computing unit, and another is pulse nerve
Network calculations unit.System input is input to the first basic computational ele- ment 110a and second basic computational ele- ment simultaneously
110b carries out parallel processing, the place that the first basic computational ele- ment 110a and the second basic computational ele- ment 110b are each obtained
Reason result is collected, and is exported as system.
Fig. 4 is referred to, described two basic basic computational ele- ments 110 are connected in parallel to form a parallel recombiner unit
120c.Described two basic computational ele- ments 110 are respectively the first basic computational ele- ment 110a and the second basic computational ele- ment 110b,
The input of the first basic computational ele- ment 110a and the second basic computational ele- ment 110b input are each independent, the first base
This computing unit 110a output end connects the second basic computational ele- ment 110b output end.First basic computational ele- ment
In 110a and the second basic computational ele- ment 110b, one is artificial neural network computing unit, and another is impulsive neural networks
Computing unit.System input is divided into 2 two parts of input 1 and input, wherein the first basic computational ele- ment 110a is sent into input 1,
By the first basic computational ele- ment 110a processing, the second basic computational ele- ment 110b is sent into input 2, by the second basic calculating list
First 110b processing, make after the first basic computational ele- ment 110a and the second basic computational ele- ment 110b result is collected
Exported for system.
Fig. 5 is referred to, it is multiple that described two basic basic computational ele- ments 110 and a unit 111 form a study
Close unit 120d.Described two basic computational ele- ments 110 are respectively the first basic computational ele- ment 110a and the second basic calculating list
First 110b.In the first basic computational ele- ment 110a and the second basic computational ele- ment 110b, one is artificial neural network meter
Unit is calculated, another is impulsive neural networks computing unit.System input obtains after the first basic computational ele- ment 110a processing
Reality output, the difference of the reality output and target output is input to unit 111, the unit 111 is according to this
Difference adjusts the parameters such as the network structure of the second basic computational ele- ment 110b, synapse weight, in the unit
Practising algorithm can be for Delta rules, BP algorithm, simulated annealing, genetic algorithm etc., the learning algorithm used in the present embodiment
For BP algorithm.The second basic computational ele- ment 110b output can as the first basic computational ele- ment 110a network structure,
The parameters such as synapse weight, or according to the second basic computational ele- ment 110b the first basic computational ele- ment of output adjustment 110a's
The parameters such as network structure, synapse weight.
Refer to Fig. 6, in a certain embodiment of the present invention, described two basic computational ele- ments 110 form a feedback complex list
First 120e.Described two basic computational ele- ments 110 are respectively the first basic computational ele- ment 110a and the second basic computational ele- ment
110b.First basic computational ele- ment 110a output end is connected with the second basic computational ele- ment 110b input, and second is basic
Computing unit 110b result of calculation is output to the first basic computational ele- ment 110a as feedback.First basic computational ele- ment
In 110a and the second basic computational ele- ment 110b, one be artificial neural network computing unit another be impulsive neural networks meter
Calculate unit.System input exports after the first basic computational ele- ment 110a processing, and output result is as the second basic computational ele- ment
110b input, the second basic computational ele- ment 110b output are input to the first basic computational ele- ment 110a as value of feedback.
Each example is to be combined two basic computational ele- ments 110 under certain topological structure above, is formed various
Composite computing unit, further, greater number of basic computational ele- ment 110 can also be carried out under certain topological structure
Combination, forms various composite computing units, and various composite computing units are carried out into group again under certain topological structure
Close, form increasingly complex mixing and calculate structure, structure is calculated so as to produce rich and varied mixing.Refer to Fig. 7, first layer
Composite computing unit 120 is that a serial mixing calculates structure, and the first layer composite computing unit 120 is decomposed into two in the second layer
The series connection of individual composite computing unit 120, further, the second layer composite computing unit 120 are broken down into two in third layer
The parallel connection of composite computing unit 120, above-mentioned decomposable process can go on always, until being broken down into substantially in last layer
Computing unit 110, basic computational ele- ment 110 are minimum computing unit structure.Fig. 8 is referred to, the figure is by above-mentioned level
The mixing of specific an artificial neural network and impulsive neural networks that design method obtains calculates structure, including series connection
Structure, parallel-connection structure and feedback arrangement.
The hybrid system 100 of the artificial neural network and impulsive neural networks further comprises at least one form
Converting unit, it is arranged between the artificial neural networks unit and pulse nerve net computing unit, form conversion
Unit is used to realize the data transfer between different types of neural computing unit.The format conversion unit can be by people
The numerical quantities of artificial neural networks computing unit output are converted to pulse train, or impulsive neural networks computing unit is exported
Pulse train is converted to numerical quantities and realizes that form is changed, to ensure to carry out data between different types of basic computational ele- ment 110
Transmission.
Further, the hybrid system 100 of the artificial neural network and impulsive neural networks may include multiple institutes
State artificial neural network unit and multiple impulsive neural networks units, the plurality of artificial neural network unit and multiple pulses god
Topology connection is realized with above-described topological structure through NE.
The artificial neural network and the hybrid system 100 of impulsive neural networks that first embodiment of the invention provides combine
The computation schemas of two kinds of neutral nets of artificial neural network and impulsive neural networks, artificial neural network is used to needing accurate
Data processing either need complete mathematical describe computing unit, by impulsive neural networks be used for need snap information handle or
Person is complicated space-time signal processing or needs to handle the computing unit of multi-modal signal (such as audio visual signal) simultaneously, is formed
Real-time, multi-modal or complicated space-time signal and the system that can guarantee that counting accuracy can be carried out.Such as done by the system
Audio visual Integrated Real-time Processing, it can will include the multi-modal complicated space-time signal input pulse nerve of image (video) and sound
Network calculations unit is pre-processed, quick to reduce space-time characteristic needed for signal Space-time Complexity or extract real-time, pretreatment
Data afterwards continue to input artificial neural networks unit, can be by building Complete mathematic model or having supervision in this step
The mode of learning algorithm builds the artificial neural network that can be realized compared with precise information processing, ensure that the accuracy of output.
Second embodiment of the invention provides a kind of mixed communication method of artificial neural network and impulsive neural networks nerve,
Including：Whether the data type of sender and recipient are consistent in the neutral net basic computational ele- ment 110 that judgement is communicated,
Carry out data transmission if consistent, if inconsistent, perform Data Format Transform, be by the data type conversion that sender sends
With recipient's identical data type, and carry out data transmission.The data type of artificial neural network is numerical value, pulse nerve net
The data type of network is pulse train.
Specifically, during the Data Format Transform is performed, if sender is the artificial neural networks
Unit, recipient are impulsive neural networks computing unit, then the numeric format output of artificial neural network are converted into pulse sequence
Row input pulse neutral net；If sender is impulsive neural networks computing unit, recipient is artificial neural networks list
Member, then the pulse train form output of impulsive neural networks is converted into numerical value input artificial neural network.
The form of the different types of data input of basic computational ele- ment 110 output difference.Artificial neural network is based on
Numerical operation, input and output are all numerical value, and impulsive neural networks are based on pulse train computing, and input and output are all pulse train.
Different types of basic computational ele- ment 110 is integrated into same mixing to calculate in structure, it is necessary to solve the problems, such as to communicate each other.
Need to provide a kind of artificial neural network the mixed communication method neural with impulsive neural networks so that artificial neural network meter
Calculating the numeric format output of unit can be received by impulsive neural networks computing unit, and impulsive neural networks computing unit
The output of pulse train form can be received by artificial neural networks unit.
Refer to Fig. 9, in a certain embodiment of the present invention, the impulsive neural networks unit is data receiver, the people
Artificial neural networks unit is data sender, and the communication process is：The numerical quantities that artificial neural networks unit is exported
The pulse train of respective frequencies, and the input using the pulse train as impulsive neural networks computing unit are converted to, it is described right
Frequency is answered to refer to that the frequency of the pulse train after conversion is directly proportional to the size of numerical quantities.
In a certain embodiment of the present invention, the artificial neural network unit is data receiver, the impulsive neural networks
Unit is data sender, and the communication process is：The pulse train form output of impulsive neural networks is converted to corresponding big
Small numerical quantities, and using the numerical quantities as the input as artificial neural network.Compiled according to impulsive neural networks pulse train
The difference of code mode, can be divided into following 4 kinds of situations again：
Figure 10 is referred to, pulse train uses frequency coding, and network output effective information is only represented by pulse frequency.Work as arteries and veins
When rushing neutral net and using this coded system, the pulse train of frequency coding is converted to the method for numerical quantities is：By pulse
Sequence is converted to the numerical quantities of corresponding numerical value, and the corresponding numerical value refers to the size of numerical quantities and the frequency of pulse train into just
Than, that is, the inverse process of the above-mentioned communication means from artificial neural network to impulsive neural networks.
Figure 11 is referred to, pulse train is Population Coding, the characterized same information of output of multiple neurons, effectively letter
Cease the neuron number that spike is sent for synchronization.It is corresponding logical when impulsive neural networks use this coded system
Letter method is：The neuron number of the transmission spike of synchronization is converted into corresponding numerical value, the size of numerical value and granting
The neuron number of spike is directly proportional.
Figure 12 is referred to, pulse train is time encoding, and effective information is the time that neuron sends spike.Work as arteries and veins
When rushing neutral net and using this coded system, corresponding communication means is：The time that neuron is sent to spike changes
For corresponding numerical value, size and the spike Time Of Release exponentially functional relation of numerical value.
Figure 13 is referred to, pulse train is binary-coding, and whether effective information provides point for neuron within a certain period of time
Peak pulse.When impulsive neural networks use this coded system, corresponding communication means is：If point is provided in limiting time
Peak pulse, then numerical value is 1, and otherwise numerical value is 0.
The mixed communication method of the artificial neural network that second embodiment of the invention provides and impulsive neural networks nerve is real
The direct communication of artificial neural network and impulsive neural networks is showed, above two neutral net can be realized on this basis
Hybrid operation.
Figure 14 is referred to, third embodiment of the invention provides a kind of multi-modal neuromorphic network core 210a, including：Pattern
Register 211, aixs cylinder input block 212, synapse weight memory cell 213, dendron unit 214 and neuron computing unit 215.
In the present embodiment, the multi-modal neuromorphic network core 210a has a bars input aixs cylinder, b bars dendron and b nerve
First cell space, every aixs cylinder are connected with a bar dendrons respectively, are a cynapse at each tie point, share a × b cynapse, connection weight
It is synapse weight again.The corresponding pericaryon of every dendron.Therefore the multi-modal neuromorphic network core 210a is maximum
The network size that can be carried is a inputs × b neurons.Wherein a, b are the integer more than 0, and a, b value are typically according to reality
Depending on.A minimums can be taken as 1, and maximum can be realized to be limited to hardware resource, including storage resource, logical resource etc.
Single neuromorphic network core maximum aixs cylinder inputs number；B minimum desirable 1, maximum while hardware resource is limited to, also by
It is limited to the neural n ary operation total degree that each neuromorphic network core can perform in the single trigger signal cycle.A in the present embodiment
Value with b is 256.It is appreciated that in actual applications, the specific number of above-mentioned input aixs cylinder, dendron and pericaryon
Mesh can be adjusted according to concrete condition.
The mode register 211 controls the operational mode of the multi-modal neuromorphic network core 210a.The pattern
Register 211 is connected with the aixs cylinder input block 212, dendron unit 214 and neuron computing unit 215, and control is above-mentioned
Unit operates in artificial neural network pattern or impulsive neural networks pattern.The width of the mode register 211 is 1,
Its register value can pass through user configuration.
The aixs cylinder input block 212 is connected with the dendron unit 214.The aixs cylinder input block 212 receives a bar axles
It is prominent to input and store.In the present embodiment, every aixs cylinder has the memory cell of 16.When operating in impulsive neural networks pattern,
The input of every aixs cylinder is all the spike of 1, and the memory cell stores the aixs cylinder input of 16 time steps.Work as operation
In artificial neural network pattern, the input of every aixs cylinder is all the signed number of 8, and the memory cell stores a time
8 aixs cylinders input of step.
The synapse weight memory cell 213 is connected with the dendron unit 214, the synapse weight memory cell 213
A × b synapse weight is stored, in the present embodiment, the synapse weight is the signed number of 8.
The input of the dendron unit 214 connects the aixs cylinder input block 212 and synapse weight memory cell 213,
Output end connects the neuron computing unit 215.The dendron unit 214 performs a aixs cylinder input vector and a × b cynapses
The vector-matrix multiplication computing of weight matrix, a result of calculation that computing obtains are that the aixs cylinder of a neuron is inputted.Institute
Stating dendron unit 214 includes dendron multiplicaton addition unit 214a and dendron summing elements 214b.When operating in artificial neural network pattern
When, the aixs cylinder input vector is sent into the dendron multiplicaton addition unit 214a with synapse weight matrix and carries out multiply-add operation, moment of a vector
Battle array multiplication is realized by multiplier and accumulator.When operating in impulsive neural networks pattern, the aixs cylinder input vector is with dashing forward
Touch weight matrix and be sent into the dendron summing elements progress accumulating operation, now, aixs cylinder input is 1, and vector-matrix multiplication is led to
Cross data selector and accumulator is realized.
The neuron computing unit 215 is used to perform neuron calculating, including the meters of the first computing unit 215a and second
Calculate unit 215b.When operating in artificial neural network pattern, multiply-add operation result that the dendron multiplicaton addition unit 214a is sent
It is sent into the first computing unit 215a and carries out artificial neural network computing.Neuron passes through one 1024 × 8 in the present embodiment
The look-up tables'implementation Any Nonlinear Function function of position, is exported as 8 bit values.When operating in impulsive neural networks pattern, institute
State the accumulating operation result that dendron summing elements 214b is sent and be sent into the second computing unit 215b progress impulsive neural networks
Calculate.Neuron in the present embodiment is integration-leakage-igniting (LIF) model, is exported for spike signal and when cephacoria electricity
Position.
The multi-modal neuromorphic network core 210a further comprises trigger signal counter 216, the trigger signal
Counter 216 receives trigger signal and records trigger signal number, i.e., current time walks.The trigger signal for the fixed cycle when
Clock signal, the cycle is 1ms in the present embodiment.
The multi-modal neuromorphic network core 210a further comprises controller 217.The controller 217 and the axle
Prominent input block 212, synapse weight memory cell 213, dendron unit 214, neuron computing unit 215 connect, the control
Device 217 controls above-mentioned aixs cylinder input block 212, synapse weight memory cell 213, dendron unit 214, neuron computing unit
215 operation sequential.In addition the controller 217 be also responsible for controlling the multi-modal neuromorphic network core 210a startup and
Terminate, the controller 217 starts the multi-modal neuromorphic network core 210a computings simultaneously in the trigger signal rising edge
Control computing flow.
Impulsive neural networks operational mode and the people of the multi-modal neuromorphic network core 210a will be introduced respectively below
Artificial neural networks operational mode.
Figure 15 is referred to, the multi-modal neuromorphic network core 210a operates in artificial neural network pattern, the pattern
Under the aixs cylinder input that receives of the aixs cylinder input block 212 be the packet of 16, including 8 target aixs cylinder sequence numbers (0~
255) and 8 input (signed number -128~127).The now memory of the aixs cylinder input block 212 256 × 8, this is deposited
Reservoir records 8 inputs in 256 aixs cylinders.
The synapse weight memory cell 213 stores 256 × 256 synaptic weight values, and each synaptic weight value is 8
Signed number.
The dendron multiplicaton addition unit 214a includes multiplier and adder, is calculated per time step by neuron on its dendron
The product of 256 dimension synapse weight vectors and 256 dimension aixs cylinder input vectors, and using result of calculation as first computing unit
215a input.
The first computing unit 215a performs non-linear/linear activation primitive meter of neuron in artificial neural network
Calculate.Before use, first according to used in the neuromorphic network activation primitive, calculate for value be 0~1023 input
Corresponding 8 output, and above-mentioned result of calculation is stored in the look-up table storage of neuron computing unit.Operationally, in the future
Inputted from the multiply-add result of dendron of the dendron multiplicaton addition unit 214a as the address of look-up table, the look-up table address is stored
8 outputs are neuron output.
Refer to Figure 16, under artificial neural network pattern, described multi-modal mono- time step of neuromorphic network core 210a
Operational process comprise the following steps：
S11, after detecting trigger signal, from adding 1, the controller 217 starts for the circulation of trigger signal counter 216
Multi-modal neuromorphic network core 210a operations；
S12, the aixs cylinder input block 212 read current time step axle according to the value of the trigger signal counter 216
Prominent input vector, and the dendron multiplicaton addition unit 214a is sent to, the synapse weight memory cell 213 sequential reads out 1~256
Number neuron dendron synapse weight vector, and it is sent to the dendron multiplicaton addition unit 214a；
S13, the dendron multiplicaton addition unit 214a calculated successively according to input sequence the aixs cylinder input vector with described 1~
No. 256 neuron dendron synapse weight vector products, and it is sent to the first computing unit 215a；
S14, the first computing unit 215a obtain using the output of the dendron multiplicaton addition unit 214a as table address is searched
Go out neuron output；
S15, the controller 217 stop the multi-modal neuromorphic network core 210a operations, and return to step S11.
Figure 17 is referred to, the multi-modal neuromorphic network core 210a operates in impulsive neural networks pattern, the pattern
Under the aixs cylinder input that receives of the aixs cylinder input block 212 be the packet of 12, including 8 target aixs cylinder sequence numbers (0~
And 4 delayed datas (0~15) 255).The delayed data represents the difference that the time step that the input comes into force walks with current time
Value.Now the memory of the aixs cylinder input block 212 is 256 × 16, the memory record it is current in 256 aixs cylinders and it
The input of totally 16 time steps afterwards.If certain position is 1, corresponding aixs cylinder activates on corresponding time step, i.e. input is 1；If certain position
For 0, then correspond to aixs cylinder and do not activated on corresponding time step, be i.e. input is 0.
The synapse weight memory cell 213 stores 256 × 256 synaptic weight values, and each synaptic weight value is 8
Signed number.
The dendron summing elements 214b includes data selector and adder, calculates current time on every dendron successively
Walk all activated cynapse weight and, and using result of calculation as the second computing unit in the neuron computing unit 215
215b input.
The second computing unit 215b is used to carry out impulsive neural networks calculating, and the neuron in the present embodiment is product
Point-leakage-igniting (LIF) model.The second computing unit 215b further comprises：Dendron expands memory cell 2151, ginseng
Number memory cell 2152 and integration leak calculation unit 2153.In each time step, the second computing unit 215b is transported successively
Row 256 times, the computing of 256 neurons is realized by time-multiplexed mode.Wherein described dendron expands memory cell 2151
Comprising 256 memory cell, input data bag is expanded for receiving the extraneous dendron sent, the dendron expands input data bag
Including transmitting terminal membrane potential of neurons value and target nerve member sequence number, and according to neuron sequence number by transmitting terminal membrane potential of neurons
Corresponding memory cell is arrived in value storage.The parameter storage unit 2152 includes 256 memory cell, stores 256 neurons
Film potential, threshold value and leakage value.The integration leak calculation unit 2153 each neuron is performed successively integration-leakage-
Ignition operation, when film potential exceedes positive threshold value, output spike pulse signal and current film potential value.It is wherein described to integrate-let out
Leakage-ignition operation is as follows：
Film potential=former film potential+dendron input+dendron expands input-leakage value
If film potential is more than positive threshold value, spike signal and film potential value are sent, and film potential is reset.If
Film potential is less than negative threshold value, then does not send signal and reset film potential.
Figure 18 is referred to, one time step of impulsive neural networks pattern of the multi-modal neuromorphic network core 210a
Operational process comprises the following steps：
S21, after detecting trigger signal, from adding 1, the controller 217 starts for the circulation of trigger signal counter 216
Multi-modal neuromorphic network core 210a operations；
S22, the aixs cylinder input block 212 read current time step axle according to the value of the trigger signal counter 216
Prominent input vector, and the dendron summing elements 214b is sent to, the synapse weight memory cell 213 sequential reads out 1~256
The dendron synapse weight vector of number neuron, and it is sent to the dendron summing elements 214b；
S23, the dendron summing elements 214b calculated successively according to input sequence the aixs cylinder input vector with described 1~
No. 256 neuron dendron synapse weight vector products, and it is sent to the second computing unit 215b；
S24, the second computing unit 215b read 1~No. No. 256 god from the parameter storage unit 2152 successively
Calculated through first parameter, and with the input value from the dendron summing elements 214b, output spike pulse signal and current
Film potential；
S25, after neuron calculating terminates, the controller 217 stops multi-modal neuromorphic network core 210a operations,
And return to step S21.
It is possible to further obtain two kinds on the basis of the multi-modal neuromorphic network core 210a that provides the present embodiment
Single mode neuromorphic network core.A kind of single mode neuromorphic network core operated under artificial neural network pattern, including：
Aixs cylinder input block 212, synapse weight memory cell 213, dendron multiplicaton addition unit 214a and the first computing unit 215a.It is a kind of
The single mode neuromorphic network core operated under impulsive neural networks pattern, including：Aixs cylinder input block 212, synapse weight
Memory cell 213, dendron summing elements 214b and the second computing unit 215b.Above two single mode neuromorphic network core
It is that corresponding simplification has been done on the basis of 3rd embodiment, the specific attachment structure and function of internal each unit are referred to
Third embodiment of the invention.
It is appreciated that multi-modal neuromorphic network core 210a and two kinds of single modes that third embodiment of the invention is provided
State neuromorphic network core can calculate as the mixing of first embodiment of the invention artificial neural network and impulsive neural networks
Basic computational ele- ment 110 in system 100.
The multi-modal neuromorphic network core 210a that third embodiment of the invention provides can both carry out artificial neural network
Computing, impulsive neural networks computing can also be carried out, and can be as required in artificial neural network operational mode and pulse god
Through switching between Network operation mode, real-time, multi-modal or complicated space-time signal can be carried out and calculate and can guarantee that calculating
Accuracy.
Figure 19 is referred to, fourth embodiment of the invention provides a kind of mixed stocker of artificial neural network and impulsive neural networks
System 200, including：Multiple neuromorphic network cores 210 and with the plurality of neuromorphic network core multiple routes correspondingly
Node 220, the multiple routing node 220 form the route network of m × n network structures, wherein, m, n are whole more than 0
Number.Line direction in the m n array is defined as X-direction, column direction is defined as Y-direction, each pair neuromorphic network core 210 with
Routing node 220 has an only local XY coordinate.
The neuromorphic network core 210 corresponds with routing node 220, refers to each neuromorphic network core 210
Corresponding to a routing node 220, each routing node 220 corresponds to a neuromorphic network core 210.One corresponding to mutually
To in neuromorphic network core 210 and routing node 220, neuromorphic network core 210 is referred to as the local nerve of routing node 220
Form network core, routing node 220 are referred to as the local routing node of neuromorphic network core 210.The neuromorphic network core
210 input both is from its local routing node, and neuron result of calculation can also be sent to the local routing node, and then pass through
Route network exports or is sent to target nerve form network core.
In the present embodiment, the quantity of the neuromorphic network core 210 and routing node 220 is 9,9 routes
Node 220 forms the route network of 3 × 3 network structures.
The neuromorphic network core 210 can be single mode neuromorphic network core, or multi-modal neural shape
State network core, such as the multi-modal neuromorphic network core 210a that third embodiment of the invention provides.So-called single mode nerve shape
State network core refers to that the neuromorphic network core can only operate in artificial neural network pattern or impulsive neural networks pattern；It is so-called
Multi-modal neuromorphic network core refers to that the neuromorphic network core has two kinds of operational modes：Artificial neural network pattern and pulse
Network mode, it can be switched by configuring inherent parameters and realizing between above two operational mode, each multi-modal god
Can single configuration operation pattern through form network core.
The neuromorphic network core 210 has multiple neuron elements and the input of multiple aixs cylinders.In the present embodiment, the god
There are 256 neuron elements and 256 aixs cylinder inputs through form network core 210, maximum can undertake the god for including 256 neurons
Through network operations.If the scale of neural network to be realized is more than 256 neurons, need to realize will be more by route network
Individual neuromorphic network core 210 connects together on the topology, and each neuromorphic network core 210 undertakes a part of nerve
Network operations, collectively constitute a big neutral net.The hybrid system 200 that the present embodiment is provided has 9 neuromorphic nets
Network core 210, maximum can undertake the neural computing for including 2304 neurons, and the neural computing can be artificial god
Hybrid network computing through network operations, impulsive neural networks computing or artificial neural network and impulsive neural networks.
When the neuromorphic network core 210 is operated under artificial neural network pattern, the neuromorphic network core 210
In the result of calculation of 256 neuron elements be all the numerical value of 8；When neuromorphic network core 210 operates in pulse nerve
When under network mode, the result of calculation of the unit of 256 neurons in the neuromorphic network core 210 is all the spike of 1
Pulse.Which kind of no matter operate under pattern, the result of calculation of the unit of neuron can all be sent directly to the neuromorphic network
The local routing node of core 210.
The multiple routing node 220 forms the route network of network structure, and the route network undertakes data-transformation facility,
The data transfer includes：Transmission between system input, system output and neuromorphic network core 210, wherein neuromorphic
Between network core 210 transmission again be divided into the internuclear data transfer of model identical neuromorphic network and different mode neuromorphic
The internuclear data transfer of network.Any routing node 220 in the route network can use an XY that is unique and determining
Plane coordinates represents.In the present embodiment, 9 routing nodes 220 form 3 × 3 array, and line direction in array is defined as into X side
To column direction is defined as Y-direction, and each routing node 220 can be born with it in X positive directions, X negative directions, Y positive directions, Y
The adjacent routing node in direction directly carries out data transmission, and forms mesh topology.It is appreciated that the mesh topology
In addition to above-mentioned network structure, or other common structures such as hub-and-spoke configuration, bus type structure.
The data of the route network transmission include：System input data, system output data and neuromorphic network
The data transmitted between core 210.Above-mentioned data are transmitted according to routing rule set in advance in route network.The present embodiment
Middle routing rule is：Data are transmitted in X direction first, are exported again along Y-direction after reaching target X-coordinate routing node, directly
To the routing node for reaching target XY coordinates.If representing the coordinate of starting routing node with (x0, y0), (x1, y1) represents mesh
The coordinate of routing node is marked, above-mentioned routing rule is：(x0,y0)→(x1,y0)→(x1,y1).It is appreciated that actually should
In, different routing rules can also be set according to specific circumstances.
The implementation process transmitted individually below between introducing system input, system output and neuromorphic network core 210.
The transmitting procedure of the system input data is：System input data is first inputted to appointing for route network outermost
One routing node, target nerve form network core can be sent to according to above-mentioned routing rule by route network afterwards.
The transmitting procedure of the system output data is：The result of calculation of the neuromorphic network core 210 is sent first
To local routing node, any route of route network outermost is sent to according to above-mentioned routing rule by route network afterwards
Node, then be sent to by the routing node outside system, complete system output.
The process transmitted between the neuromorphic network core 210 is：The result of calculation of the neuromorphic network core 210 is first
Local routing node is first sent to, target routing node is sent to according to above-mentioned routing rule by route network afterwards, then by
The target routing node is sent to its local neuromorphic network core, completes the data transfer between neuromorphic network core 210.
Refer to Figure 20, the routing node 220 includes the routing table with multiple memory cell, the routing table it is each
Memory cell corresponds to a neuron elements of local neuromorphic network core.The memory cell stores corresponding neuron list
XY coordinate address, target aixs cylinder input sequence number and the delayed data of the purpose neuromorphic network core of member output.In the present embodiment
Neuromorphic network core 210 includes 256 neuron elements, and the number of memory cell is also 256 in the routing table.
In the present embodiment, communicated between the system input data, system output data and neuromorphic network core 210
Data are transmitted in the form of 32 route data bags between routing node 220.Figure 21 show the lattice of above-mentioned route data bag
Formula, 6 target nerve form network core X-direction addresses, 6 target nerve form network core Y sides are included in the route data bag
Sequence numbers and 8 data are inputted to address, 4 aixs cylinder delays, 8 target aixs cylinders, altogether 32 data.Wherein, the aixs cylinder of 4 is prolonged
When operate under impulsive neural networks pattern effectively when target nerve form network core, 8 data are artificial neural network pattern
The output of lower neuron.
The packet that the routing node 220 is dealt into local neuromorphic network core is from above-mentioned 32 route data bags
4 aixs cylinders delay of interception, 8 target aixs cylinders input sequence numbers and 8 data, altogether 20 data.For operating in artificial god
Through the neuromorphic network core 210 under network mode, after receiving the packet from local routing node, 8 data are made
For the input of corresponding sequence number aixs cylinder.For operating in the neuromorphic network core 210 under impulsive neural networks pattern, receive and
From after the packet of local routing node, input of the corresponding sequence number aixs cylinder after corresponding be delayed is put 1.
Figure 22 is referred to, the workflow of the routing node 220 includes：
S1, the routing node 220 are received come the neuron result of calculation for local neuromorphic network core of hanging oneself；
S2, the routing node 220 read the routing iinformation of corresponding neuron from routing table, and by the routing iinformation with
The neuron result of calculation is combined as route data bag；
S3, the routing node 220 judge the sending direction of the route data bag, and the route is sent according to judged result
Packet.
In step S1, when operating under artificial neural network pattern, the neuron result of calculation is output data；Work as fortune
For row under artificial neural network pattern, the neuron result of calculation is spike.
In step S3, the route data bag include target nerve form network core in the relative address x of X-direction and
Relative address y in the Y direction.The routing node 220 judges the sending direction of the route data bag according to x and y numerical value,
Specially：Work as x>When 0, the route data bag is sent to the adjacent routing node of X positive directions, works as x<When 0, by the route
Packet is sent to the adjacent routing node of X negative directions, works as y>When 0, it is adjacent that the route data bag is sent to Y positive directions
Routing node, work as y<When 0, the route data bag is sent to the adjacent routing node of Y negative directions, as x=y=0, by institute
State the local neuromorphic network core that route data bag directly transmits back the routing node 220.Next routing node receives
After the route data bag that one routing node is sent, the relative address in the route data bag is modified, is specially：If x<0,
Revised relative address x '=x+1 is then made, if x>0, then revised relative address x '=x-1 is made, if y<0, then order amendment
Relative address y '=y+1 afterwards, if y>0, then make revised relative address y '=y-1.The relative address often passes through one
Routing node 220 is once corrected, until x=y=0.
In the present embodiment, the judgement and transmission of X-direction are carried out first, after X-direction, which is sent, to be terminated, i.e. x '=0, then carry out
The judgement and transmission of other direction, sent until by route data bag to target nerve form network core.
The artificial neural network and the hybrid system 200 of impulsive neural networks that fourth embodiment of the invention provides combine people
The computation schema of two kinds of neutral nets of artificial neural networks and impulsive neural networks, both with impulsive neural networks complicated space-time letter
Number disposal ability, and can makes full use of artificial neural network abundant and powerful computing capability, can carry out real-time, multi-modal
Or complicated space-time signal calculates and can guarantee that the accuracy of calculating.
In addition, those skilled in the art can also do other changes in spirit of the invention, certainly, these are according to present invention essence
The change that god is done, it should all be included within scope of the present invention.
Claims (11)
- A kind of 1. multi-modal neuromorphic network core, it is characterised in that including：Mode register, aixs cylinder input block, cynapse power Weight memory cell, dendron unit and neuron computing unit；The mode register is connected with the aixs cylinder input block, dendron unit and neuron computing unit, and control is above-mentioned Unit operates in artificial neural network pattern or impulsive neural networks pattern；The aixs cylinder input block is connected with the dendron unit, is received and is stored aixs cylinder input；The synapse weight memory cell is connected with the dendron unit, stores synapse weight matrix；The dendron unit is connected with the neuron computing unit, including dendron multiplicaton addition unit and dendron summing elements, operation In artificial neural network pattern, the aixs cylinder input vector is sent into the dendron multiplicaton addition unit with synapse weight matrix and multiplied Add computing, when operating in impulsive neural networks pattern, the aixs cylinder input vector is sent into the dendron with synapse weight matrix and tired out Unit is added to carry out accumulating operation；The neuron computing unit includes the first computing unit and the second computing unit, operates in artificial neural network pattern When, the multiply-add operation result that the dendron multiplicaton addition unit is sent is sent into the first computing unit progress artificial neural network fortune Calculate, when operating in impulsive neural networks pattern, the accumulating operation result that the dendron summing elements are sent is sent into second meter Calculate unit and carry out impulsive neural networks calculating.
- 2. multi-modal neuromorphic network core as claimed in claim 1, it is characterised in that the multi-modal neuromorphic network Core further comprises trigger signal counter, and the trigger signal counter is connected with the aixs cylinder input block, receives triggering Signal simultaneously records trigger signal number, i.e., current time is walked, and the trigger signal number is sent into the aixs cylinder input block.
- 3. multi-modal neuromorphic network core as claimed in claim 1, it is characterised in that the multi-modal neuromorphic network Core further comprises controller, the controller and the aixs cylinder input block, synapse weight memory cell, dendron unit, god Connected through first computing unit, control the operation sequential of connected said units, the controller controls the multi-modal god Startup and termination through form network core.
- 4. multi-modal neuromorphic network core as claimed in claim 1, it is characterised in that the dendron multiplicaton addition unit includes multiplying Musical instruments used in a Buddhist or Taoist mass and adder, for calculating the product of synapse weight vector and aixs cylinder input vector on its dendron.
- 5. multi-modal neuromorphic network core as claimed in claim 3, it is characterised in that first computing unit includes one Individual look-up table, look-up table is inputted using the multiply-add result as the address of look-up table, the output that the address is stored is nerve Member output.
- 6. multi-modal neuromorphic network core as claimed in claim 5, it is characterised in that described under artificial neural network pattern The operational process of one time step of multi-modal neuromorphic network core comprises the following steps：S11, after detecting trigger signal, for the trigger signal counter cycle from adding 1, the controller starts multi-modal nerve Form network core is run；S12, the aixs cylinder input block read current time step aixs cylinder input vector according to the value of the trigger signal counter, And the dendron multiplicaton addition unit is sent to, the synapse weight memory cell sequential reads out 1~n neuron dendron synapse weights Vector, and it is sent to the dendron multiplicaton addition unit；S13, the dendron multiplicaton addition unit calculate the aixs cylinder input vector and 1~n nerves according to input sequence successively First dendron synapse weight vector product, and it is sent to first computing unit；S14, first computing unit show that neuron is defeated using the output of the dendron multiplicaton addition unit as table address is searched Go out；S15, the controller stop the multi-modal neuromorphic network core operation, and return to step S11.
- 7. multi-modal neuromorphic network core as claimed in claim 1, it is characterised in that the dendron summing elements include number According to selector and adder, calculate successively on every dendron the weight of all activated cynapse of current time step and.
- 8. multi-modal neuromorphic network core as claimed in claim 3, it is characterised in that neural in second computing unit Member is integration-leakage-fire model, and second computing unit includes：Dendron expands memory cell, parameter storage unit and product Divide leak calculation unit, the expansion memory cell is connected with the integration leak calculation unit, stores transmitting terminal neuron membrane Potential value, the parameter storage unit are connected with the integration leak calculation unit, store the film potential of neuron, threshold value and let out Leakage value, the integration leak calculation unit perform integration-leakage-ignition operation.
- 9. multi-modal neuromorphic network core as claimed in claim 8, it is characterised in that described under impulsive neural networks pattern The operational process of one time step of multi-modal neuromorphic network core comprises the following steps：S21, after detecting trigger signal, for the trigger signal counter cycle from adding 1, the controller starts multi-modal nerve Form network core is run；S22, the aixs cylinder input block read current time step aixs cylinder input vector according to the value of the trigger signal counter, And the dendron summing elements are sent to, the synapse weight memory cell sequential reads out the dendron cynapse power of 1~n neurons Weight vector, and it is sent to the dendron summing elements；S23, the dendron summing elements calculate the aixs cylinder input vector and 1~n nerves according to input sequence successively First dendron synapse weight vector product, and it is sent to second computing unit；S24, second computing unit read 1~No. No. n neuron parameter from the parameter storage unit successively, and with coming Calculated from the input value of the dendron summing elements, output spike pulse signal and current film potential；S25, the controller stop the multi-modal neuromorphic network core operation, and return to step S21.
- 10. multi-modal neuromorphic network core as claimed in claim 1, it is characterised in that the multi-modal neuromorphic net There is network core a bars input aixs cylinder, b bars dendron and b pericaryon, every aixs cylinder to be connected respectively with b bar dendrons, each connection It is a cynapse at point, shares a × b cynapse, the corresponding pericaryon of every dendron, wherein a, b are to be whole more than 0 Number.
- 11. multi-modal neuromorphic network core as claimed in claim 10, it is characterised in that the a=b=256.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510419465.5A CN105095967B (en) | 2015-07-16 | 2015-07-16 | A kind of multi-modal neuromorphic network core |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201510419465.5A CN105095967B (en) | 2015-07-16 | 2015-07-16 | A kind of multi-modal neuromorphic network core |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN105095967A CN105095967A (en) | 2015-11-25 |
| CN105095967B true CN105095967B (en) | 2018-02-16 |
Family
ID=54576341
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201510419465.5A Active CN105095967B (en) | 2015-07-16 | 2015-07-16 | A kind of multi-modal neuromorphic network core |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN105095967B (en) |
Families Citing this family (33)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109242094B (en) * | 2016-01-20 | 2020-05-08 | 中科寒武纪科技股份有限公司 | Apparatus and method for performing artificial neural network forward operations |
| CN105719000B (en) * | 2016-01-21 | 2018-02-16 | 广西师范大学 | A kind of neuron hardware unit and the method with this unit simulation impulsive neural networks |
| US10878318B2 (en) * | 2016-03-28 | 2020-12-29 | Google Llc | Adaptive artificial neural network selection techniques |
| US10990872B2 (en) * | 2016-03-31 | 2021-04-27 | International Business Machines Corporation | Energy-efficient time-multiplexed neurosynaptic core for implementing neural networks spanning power- and area-efficiency |
| CN108510064B (en) * | 2016-04-18 | 2021-12-10 | 中国科学院计算技术研究所 | Processing system and method for artificial neural network comprising multiple core processing modules |
| US20170330069A1 (en) * | 2016-05-11 | 2017-11-16 | Kneron Inc. | Multi-layer artificial neural network and controlling method thereof |
| CN106022468B (en) * | 2016-05-17 | 2018-06-01 | 成都启英泰伦科技有限公司 | the design method of artificial neural network processor integrated circuit and the integrated circuit |
| CN106056211B (en) * | 2016-05-25 | 2018-11-23 | 清华大学 | Neuron computing unit, neuron computing module and artificial neural networks core |
| CN106201651A (en) * | 2016-06-27 | 2016-12-07 | 鄞州浙江清华长三角研究院创新中心 | The simulator of neuromorphic chip |
| DE102016216944A1 (en) * | 2016-09-07 | 2018-03-08 | Robert Bosch Gmbh | Method for calculating a neuron layer of a multilayer perceptron model with a simplified activation function |
| US11270193B2 (en) * | 2016-09-30 | 2022-03-08 | International Business Machines Corporation | Scalable stream synaptic supercomputer for extreme throughput neural networks |
| CN106875005B (en) * | 2017-01-20 | 2019-09-20 | 清华大学 | Adaptive threshold neuronal messages processing method and system |
| CN106845633B (en) * | 2017-01-25 | 2021-07-09 | 北京灵汐科技有限公司 | Neural network information conversion method and system |
| CN106909969B (en) * | 2017-01-25 | 2020-02-21 | 清华大学 | Neural network information receiving method and system |
| CN106845632B (en) * | 2017-01-25 | 2020-10-16 | 清华大学 | Method and system for converting impulse neural network information into artificial neural network information |
| CN106971229B (en) * | 2017-02-17 | 2020-04-21 | 清华大学 | Neural network computing core information processing method and system |
| CN106971228B (en) * | 2017-02-17 | 2020-04-07 | 北京灵汐科技有限公司 | Method and system for sending neuron information |
| US10909449B2 (en) | 2017-04-14 | 2021-02-02 | Samsung Electronics Co., Ltd. | Monolithic multi-bit weight cell for neuromorphic computing |
| CN107657312B (en) * | 2017-09-18 | 2021-06-11 | 东南大学 | Binary network implementation system for speech common word recognition |
| US20190332924A1 (en) * | 2018-04-27 | 2019-10-31 | International Business Machines Corporation | Central scheduler and instruction dispatcher for a neural inference processor |
| KR102650660B1 (en) * | 2018-09-03 | 2024-03-25 | 삼성전자주식회사 | Neuromorphic apparatus and method for processing multi-bits neuromorphic operation thereof |
| CN109858620B (en) * | 2018-12-29 | 2021-08-20 | 北京灵汐科技有限公司 | Brain-like computing system |
| CN109901878B (en) * | 2019-02-25 | 2021-07-23 | 北京灵汐科技有限公司 | Brain-like computing chip and computing equipment |
| CN110263926B (en) * | 2019-05-18 | 2023-03-24 | 南京惟心光电系统有限公司 | Pulse neural network based on photoelectric computing unit, system and operation method thereof |
| CN110322010B (en) * | 2019-07-02 | 2021-06-25 | 深圳忆海原识科技有限公司 | Pulse neural network operation system and method for brain-like intelligence and cognitive computation |
| KR20210012839A (en) * | 2019-07-26 | 2021-02-03 | 에스케이하이닉스 주식회사 | Semiconductor device for performing an operation |
| CN111565152B (en) * | 2020-03-27 | 2022-04-29 | 中国人民解放军国防科技大学 | Brain-like chip routing system data communication method based on routing domain division |
| CN111667064B (en) * | 2020-04-22 | 2023-10-13 | 南京惟心光电系统有限公司 | Hybrid neural network based on photoelectric computing unit and operation method thereof |
| CN112966814B (en) * | 2021-03-17 | 2023-05-05 | 上海新氦类脑智能科技有限公司 | Information processing method of fusion impulse neural network and fusion impulse neural network |
| CN113554162B (en) * | 2021-07-23 | 2022-12-20 | 上海新氦类脑智能科技有限公司 | Axon input extension method, device, equipment and storage medium |
| CN113313240B (en) * | 2021-08-02 | 2021-10-15 | 成都时识科技有限公司 | Computing device and electronic device |
| CN114861864A (en) * | 2022-02-24 | 2022-08-05 | 天津大学 | Neuron network modeling method and device with dendritic morphology |
| CN114611686B (en) * | 2022-05-12 | 2022-08-30 | 之江实验室 | Synapse delay implementation system and method based on programmable neural mimicry core |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6269352B1 (en) * | 1995-10-13 | 2001-07-31 | Stmicroelectronics S.R.L. | Low-voltage, very-low-power conductance mode neuron |
| CN1516070A (en) * | 2003-01-08 | 2004-07-28 | 剑 王 | Associative memory neural network |
| CN102457419A (en) * | 2010-10-22 | 2012-05-16 | 中国移动通信集团广东有限公司 | Method and device for optimizing transmission network route based on neural network model |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| GB0811057D0 (en) * | 2008-06-17 | 2008-07-23 | Univ Ulster | Artificial neural network architecture |
-
2015
- 2015-07-16 CN CN201510419465.5A patent/CN105095967B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6269352B1 (en) * | 1995-10-13 | 2001-07-31 | Stmicroelectronics S.R.L. | Low-voltage, very-low-power conductance mode neuron |
| CN1516070A (en) * | 2003-01-08 | 2004-07-28 | 剑 王 | Associative memory neural network |
| CN102457419A (en) * | 2010-10-22 | 2012-05-16 | 中国移动通信集团广东有限公司 | Method and device for optimizing transmission network route based on neural network model |
Non-Patent Citations (2)
| Title |
|---|
| "A Spike-Timing-Based Integrated Model";Hu Jun etal.;《Neural Computation》;20121130;第451-472页 * |
| "Model-Based Online Learning With Kernels";Li Guoqi etal.;《IEEE TRANSACTIONS ON NEURAL NETWORKS AND LEARNING SYSTEMS》;20130331;第356-369页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN105095967A (en) | 2015-11-25 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN105095967B (en) | A kind of multi-modal neuromorphic network core | |
| CN105095961B (en) | A kind of hybrid system of artificial neural network and impulsive neural networks | |
| CN105095965B (en) | The mixed communication method of artificial neural network and impulsive neural networks nerve | |
| CN105095966B (en) | The hybrid system of artificial neural network and impulsive neural networks | |
| Ferreira et al. | An approach to reservoir computing design and training | |
| Ishibuchi et al. | Fuzzy neural networks with fuzzy weights and fuzzy biases | |
| CN107423839A (en) | A kind of method of the intelligent building microgrid load prediction based on deep learning | |
| Ernoult et al. | Updates of equilibrium prop match gradients of backprop through time in an RNN with static input | |
| CN106570597A (en) | Content popularity prediction method based on depth learning under SDN architecture | |
| CN107909206A (en) | A kind of PM2.5 Forecasting Methodologies based on deep structure Recognition with Recurrent Neural Network | |
| CN104077438B (en) | Power network massive topologies structure construction method and system | |
| Igiri et al. | Effect of learning rate on artificial neural network in machine learning | |
| Velasco et al. | Back propagation artificial neural networks for the analysis of quality control charts | |
| Zong et al. | Price forecasting for agricultural products based on BP and RBF Neural Network | |
| CN104915714A (en) | Predication method and device based on echo state network (ESN) | |
| CN105260556B (en) | The overhead crane modeling method of hair clip mutation operation RNA genetic algorithm | |
| Prado et al. | FPGA based implementation of a Fuzzy Neural Network modular architecture for embedded systems | |
| Jing et al. | Study on characteristic of fractional master-slave neural network | |
| Wang et al. | Convolution neural network based load model parameter selection considering short-term voltage stability | |
| Qu et al. | An Optimization Method of SNNs for Shortest Path Problem | |
| Mayrhofer et al. | Devs simulation of spiking neural networks | |
| CN106570586A (en) | Field detection vehicle path planning method for electric energy metering device | |
| Kwon et al. | Nonlinear mapping of interval vectors by neural networks | |
| Ann et al. | Calculation of hybrid multi-layered perceptron neural network output using matrix multiplication | |
| Sambatti et al. | Self-configured neural network for data assimilation using FPGA for ocean circulation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| C06 | Publication | ||
| PB01 | Publication | ||
| C10 | Entry into substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20180212 Address after: 100036 Beijing city Haidian District West Sanhuan Road No. 10 wanghailou B block two layer 200-30 Patentee after: Beijing Ling Xi Technology Co. Ltd. Address before: 100084 Beijing Beijing 100084-82 mailbox Patentee before: Tsinghua University |
|
| TR01 | Transfer of patent right |