WO2018133568A1

WO2018133568A1 - Compound-mode neuron information processing method and system, and computer device

Info

Publication number: WO2018133568A1
Application number: PCT/CN2017/114661
Authority: WO
Inventors: 裴京; 邓磊; 施路平; 吴臻志; 李国齐
Original assignee: 清华大学
Priority date: 2017-01-20
Filing date: 2017-12-05
Publication date: 2018-07-26
Also published as: CN106875004A; CN106875004B

Abstract

A compound-mode neuron information processing method and system, and a computer device. The method comprises: reading a neuron working mode configuration parameter, wherein the neuron working mode configuration parameter comprises an artificial neuron working mode configuration parameter or a pulse neuron working mode configuration parameter; configuring the current neuron working mode according to the neuron working mode configuration parameter; receiving front-end neuron output information; reading the current neuron information; calculating the current neuron output information according to the front-end neuron output information and the current neuron information; and outputting the current artificial neuron output information or the current pulse neuron output information. The method simultaneously supports an artificial neural network and a pulse neural network under the same architecture, enriches the information processing type of a brain-like computing platform, reduces the cost of multi-task execution in different neural network working modes, and improves the multi-task execution efficiency in different neural network working modes.

Description

Composite mode neuron information processing method, system and computer equipment

Related application

The present application claims priority to Chinese Patent Application Serial No. JP-A------------

Technical field

The present invention relates to the field of artificial neural network technology, and in particular to a composite mode neuron information processing method, system and computer device.

Background technique

Most of today's artificial neural network research is still implemented in von Neumann computer software and with the high-performance GPGPU (General Purpose Graphic Processing Units) platform, the hardware overhead, energy consumption and information of the whole process. Processing speed is not optimistic. To this end, in recent years, the field of neuromorphic computing has developed rapidly, that is, the use of hardware circuits to directly construct neural networks to simulate the function of the brain, trying to achieve a massively parallel, low-energy, computing platform that can support complex mode learning.

However, the neurons of the traditional neuromorphic system are limited to support a single information processing and transmission mode: artificial neural network or pulsed neural network, resulting in a high cost of neural network construction for a single task, and low efficiency of neural networks for multiple tasks.

Summary of the invention

Based on this, it is necessary to provide a composite mode neuron information processing method and system for the problem of constructing a neural network with different working modes for tasks in different neural network working modes, the method comprising:

Reading a neuron working mode configuration parameter, the neuron working mode configuration parameter comprising an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter;

And configuring a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode;

Receiving front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output information or front-end pulse neuron output information;

Reading current neuron information, the current neuron information including current artificial neuron information or current pulsed neuron information;

Calculating current neuron output information according to the front end neuron output information and the current neuron information, including Calculating current artificial neuron output information according to the front end artificial neuron output information and the current artificial neuron information, or calculating current pulse neurons according to the front end pulse neuron output information and the current pulsed neuron information Output information;

Outputting the current artificial neuron output information or the current pulsed neuron output information.

In one of the embodiments,

The front end artificial neuron output information includes: membrane potential information output by the front end artificial neuron, and a connection weight index of the front end artificial neuron and the current artificial neuron;

The current artificial neuron information includes: current artificial neuron offset information;

And calculating, according to the membrane potential information output by the front end artificial neuron and the current artificial neuron information, the current artificial neuron output information, including:

And according to the connection weight index of the front end artificial neuron and the current artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron is read;

And calculating the current state according to the membrane potential information output by the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information by using a preset artificial neuron activation function. Artificial neurons output information.

In one of the embodiments,

The membrane potential information output according to the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information are calculated by a preset artificial neuron activation function. After the step of outputting information by the artificial neurons, the method further includes:

Reading a potential extremum and an extremum neuron identifier corresponding to the potential extremum;

Comparing the current artificial neuron output information with the potential extreme value, and if the current artificial neuron output information is greater than or equal to the potential extreme value,

Updating the potential extremum to the current artificial neuron output information, and updating the extremum neuron identifier to an identifier of the current artificial neuron.

In one of the embodiments,

The outputting the current artificial neuron output information includes:

Reading the first issuance enable identifier, where the first issue enable identifier includes allowing data to be released or not allowing data to be issued;

Reading the artificial neuron issuing data type parameter when the first issuance enable identifier is allowed to issue data, the artificial neuron issuing the data type, including: issuing the current artificial neuron output information, and issuing the potential pole And extracting one of the extreme value neuron identifiers corresponding to the potential extreme values;

Determining, according to the artificial neuron, a data type parameter, determining a final output information of the current artificial neuron;

Outputting the final output information of the current artificial neuron.

In one embodiment, the front-end pulse neuron output information includes: pulse tip information output by the front-end pulse neuron, and a connection weight index of the front-end pulse neuron and the current pulse neuron;

The current pulse neuron information includes: a current time window width, a pulse tip information sequence in a current time window, historical membrane potential information, and membrane potential leakage information;

And calculating, according to the front-end pulse neuron output information and the current pulse neuron information, the current pulse neuron output information, including:

And according to the connection weight index of the front-end pulse neuron and the current pulse neuron, the connection weight of the front-end pulse neuron and the current pulse neuron is read;

And updating the pulse tip information sequence in the current time window according to the pulse tip information output by the front-end pulse neuron and the pulse tip information sequence in the current time window, and acquiring a pulse tip information update sequence in the current time window;

And calculating, according to the current time window width, the pulse tip information update sequence in the current time window, the front end pulse neuron input information by using an attenuation function;

Calculating the current pulsed nerve by using a pulse neuron calculation model according to the front end pulse neuron input information, the connection weight of the front end pulse neuron and the current pulse neuron, the historical membrane potential information, and the membrane potential leakage information Meta output information.

In one embodiment, the pulsed nerve is passed through the pulse according to the front-end pulse neuron input information, the connection weight of the front-end pulse neuron and the current pulsed neuron, the historical membrane potential information, and the membrane potential leakage information. The meta-calculation model, after the step of calculating the current pulse neuron output information, before the step of outputting the current pulse neuron output information, the method further includes:

Obtaining a threshold potential;

Comparing the current pulsed neuron output information with the threshold potential, and determining the issuance triggering flag information according to the comparison result, where the issuing triggering flag information includes: issuing a trigger or issuing a non-trigger; and when the issuing triggering flag information is issued When triggered,

The refractory period timer is reset, and the historical membrane potential information is updated to a preset reset membrane potential information.

In one embodiment, after the step of comparing the current pulsed neuron output information with the threshold potential and determining the issuance of the trigger flag information according to the comparison result, the method further includes:

When the issue trigger flag information is not triggered by the issue, the current time step of the refractory period width and the refractory period timer is read;

Determining whether the current time is within the refractory period according to the refractory period width and the current time step of the refractory period timer, and if the current time is within the refractory period, accumulating the refractory period timer Timing a time step without updating the historical membrane potential information;

If the current time is not within the refractory period, accumulating the refractory period timer for one time step and updating the Historical membrane potential information is the current pulsed neuron output information.

In one embodiment, the obtaining a threshold potential includes:

Reading a random threshold mask potential, a threshold offset, and a random threshold;

Performing a bitwise AND operation on the random threshold and the random threshold mask potential to obtain a threshold random superposition amount;

The threshold potential is determined based on the threshold random amount and the threshold offset.

In one embodiment, the outputting the current pulse neuron output information comprises:

Reading a second issuance enablement identifier, where the second issuance enablement identifier includes allowing data to be issued or not allowing data to be issued, and when the second issuance enablement identifier is allowed to issue data,

Reading the issue trigger flag information, when the issue trigger flag information is an issue trigger;

Outputting the current pulse neuron output information.

In one embodiment, the artificial neuron output information is compared with the potential extremum, and if the artificial neuron output information is greater than or equal to the potential extremum, the potential extremum is updated to The output information of the artificial neurons can make the neural network complete the operation of the pooling step directly in the convolutional layer, without allocating physical space for the pooled layer neurons, greatly saving the neuron resources, thereby reducing the construction cost of the neural network.

In one embodiment, the current artificial neuron output information is determined by setting the release enable identifier and the artificial neuron release data type parameter, so that the output of the artificial neuron is more controllable, and the release enable flag can be configured with The neurons are not allowed to issue data, but only used as intermediate auxiliary computational neurons, which is necessary for some functions that require multiple neurons to work together. In addition, the release enable flag and the release type can work together to set the area. The neuron outputs information to a current neuron output information (the neuron output information is the maximum value in each neuron output information of the set region), and the direct maximum pooling operation is completed.

In one embodiment, the pulse tip information sequence in the current time window is updated according to the pulse tip information output by the front end pulse neuron and the pulse tip information sequence in the current time window, and the current time window pulse is acquired. The tip information update sequence can support the spatiotemporal pulse neural network model with time depth according to the current time window width, the connection weight of the front end pulse neuron and the current pulse neuron, and the front end pulse neuron input information calculated by the attenuation function. Compared with the neural network technology scheme with only one time depth, the space-time information coding ability of the pulse neural network can be greatly improved, and the application space of the pulse neural network is enriched.

In one embodiment, the threshold potential is determined by reading a random threshold mask potential and a threshold offset and receiving a configuration value given by a configuration register such that the neuron issues pulse tip information with a probability of randomness, Regardless of whether the membrane potential exceeds a fixed threshold bias, due to the existence of a positive or negative threshold random superposition amount, the neuron cell body may issue pulses, which improves the computational power and information processing capability of the pulse neural network model. .

The present invention also provides a computer device comprising a memory, a processor, and a computer program stored on the memory and operable on the processor, the processor executing the computer program to implement the method of any of the above embodiments A step of

The invention also provides a composite mode neuron information processing system, comprising:

a neuron working mode reading module, configured to read a neuron working mode configuration parameter, where the neuron working mode configuration parameter comprises an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter;

a neuron working mode configuration module, configured to configure a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode;

The front-end neuron output information receiving module is configured to receive front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output information or front-end pulse neuron output information;

a current neuron information reading module, configured to read current neuron information, where the current neuron information includes current artificial neuron information or current pulsed neuron information;

The current neuron output information calculation module is configured to calculate current neuron output information according to the front end neuron output information and the current neuron information, including outputting information according to the front end artificial neuron and the current artificial neuron Information, calculating current artificial neuron output information, or calculating current pulse neuron output information according to the front end pulse neuron output information and the current pulse neuron information;

The current neuron output information output module is configured to output the current artificial neuron output information or the current pulse neuron output information.

The composite mode neuron information processing method, the computer device and the system provided by the present invention configure a corresponding artificial neuron working mode or a pulsed neuron working mode according to a preset neuron working mode configuration parameter, and the corresponding neuron In the working mode, the current artificial neuron output information or the current pulsed neuron output information is calculated and outputted by receiving the front-end neuron output information and reading the current neuron information. The composite mode neuron information processing system provided by the invention can configure the corresponding neuron working mode according to the requirements of the task, and only needs to modify the working mode of the neuron when switching between different tasks in different neural network working modes. The configuration parameters can be different from the neural network accelerator scheme that only supports the artificial neural network mode, and the neuromorphic scheme that only supports the pulse neural network. The present invention can support the artificial neural network mode based on the same architecture. Machine learning applications and computational neuroscience applications based on pulsed neural networks enrich the type of processing information for brain-like computing platforms, reduce the cost of multi-task execution in different neural network operating modes, and improve the working mode of different neural networks. The efficiency of multitasking.

DRAWINGS

In order to clearly explain the technical solutions of the present application, the drawings to be used in the embodiments will be briefly described below. Obviously, the drawings in the following description are only some of the embodiments of the present application, and those skilled in the art can obtain other drawings according to the drawings without any creative work.

1 is a schematic flow chart of a multi-mode neural network information processing method according to an embodiment;

2 is a schematic flow chart of a multi-mode neural network information processing method according to another embodiment;

3 is a schematic flow chart of a multi-mode neural network information processing method according to still another embodiment;

4 is a schematic flow chart of a multi-mode neural network information processing method according to still another embodiment;

FIG. 5 is a schematic flowchart diagram of a multi-mode neural network information processing method according to still another embodiment; FIG.

6 is a schematic flow chart of a multi-mode neural network information processing method according to still another embodiment;

7 is a schematic structural diagram of a multi-mode neural network information processing system according to an embodiment;

8 is a schematic structural diagram of a multi-mode neural network information processing system according to another embodiment;

9 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment;

FIG. 10 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment; FIG.

11 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment;

FIG. 12 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment.

detailed description

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

FIG. 1 is a schematic flowchart of a multi-mode neural network information processing method according to an embodiment, and the multi-mode neural network information processing method shown in FIG. 1 includes:

Step S100, reading a neuron working mode configuration parameter, where the neuron working mode configuration parameter comprises an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter.

Specifically, the present invention provides a multi-mode and multi-functional composite neuromorphic cell unit based on an all-digital circuit, integrating an artificial neural network (ANN) and a pulsed neural network (SNN). Two kinds of information processing and transmission modes. The neuron cell unit can be set to the working mode of the ANN or SNN according to the processing requirements of different tasks.

The neuron working mode configuration parameter is a working mode of a well-defined neural network.

Step S200: Configure a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode.

Specifically, according to the read neuron working mode configuration parameter, the neuron cell unit may be set to an ANN or SNN working mode.

Step S300, receiving front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output Output information or front-end pulse neuron output information.

Specifically, according to the configured neuron working mode, receiving different front-end neuron output information, if configured as an artificial neuron working mode, receiving front-end artificial neuron output information, if configured as a pulsed neuron working mode, receiving The front-end pulsed neuron outputs information.

Step S400, reading current neuron information, the current neuron information including current artificial neuron information or current pulsed neuron information.

Specifically, according to the configured neuron working mode, reading different current neuron information, if configured as an artificial neuron working mode, reading current artificial neuron information, if configured as a pulsed neuron working mode, reading Take the current pulse neuron information.

Step S500, calculating current neuron output information according to the front-end neuron output information and the current neuron information, including calculating current artificial neurons according to the front-end artificial neuron output information and the current artificial neuron information Outputting information, or calculating current pulse neuron output information according to the front end pulse neuron output information and the current pulse neuron information.

Specifically, according to the configured neuron working mode, the calculation process of the artificial neuron or the calculation process of the pulse neuron is performed, and the current artificial neuron output information or the current pulse neuron output information is obtained.

Step S600, outputting the current artificial neuron output information or the current pulse neuron output information.

Specifically, according to the configured neuron working mode, when working in the artificial neuron mode, outputting the current artificial neuron output information according to an output manner of the artificial neural network, when working in the pulsed neuron mode, The current pulsed neuron output information is output according to an output manner of the pulsed neural network.

The composite mode neuron information processing method provided by the invention provides a corresponding artificial neuron working mode or a pulsed neuron working mode according to a preset neuron working mode configuration parameter, and passes the corresponding neuron working mode. Receiving front-end neuron output information, and reading current neuron information, calculating current artificial neuron output information or current pulsed neuron output information, and outputting. The composite mode neuron information processing method provided by the invention can configure the corresponding neuron working mode according to the requirements of the task, and only needs to modify the neuron working mode when switching between different tasks in different neural network working modes. The configuration parameters can be different from the neural network accelerator scheme that only supports the artificial neural network mode, and the neuromorphic scheme that only supports the pulse neural network. The present invention can support the artificial neural network mode based on the same architecture. Machine learning applications and computational neuroscience applications based on pulsed neural networks enrich the type of processing information for brain-like computing platforms, reduce the cost of multi-task execution in different neural network operating modes, and improve the working mode of different neural networks. The efficiency of multitasking.

FIG. 2 is a schematic flowchart diagram of a multi-mode neural network information processing method according to another embodiment, and the multi-mode neural network information processing method shown in FIG. 2 is the method for configuring the neuron working mode in the method shown in FIG. When the parameters are configured for the artificial neuron working mode, the refining steps of step S300 to step S500 include:

Step S100a, receiving the membrane potential information output by the front end artificial neuron, and the connection weight index of the front end artificial neuron and the current artificial neuron.

Specifically, the connection weight index of the front end artificial neuron and the current artificial neuron is a weight index sent by the front end neuron together with the front end artificial neuron output information, and is used to indicate the extraction of the current neuron weight.

In step S200a, the current artificial neuron offset information is read.

Specifically, the offset information of the artificial neuron is a membrane potential bias value.

Step S300a: The connection weight of the front end artificial neuron and the current artificial neuron is read according to the connection weight index of the front end artificial neuron and the current artificial neuron.

Specifically, the connection weight index of the front end artificial neuron and the current artificial neuron is an address information, and the current neuron is indexed according to the received connection weight of the front end artificial neuron and the current artificial neuron, and the current nerve is In the memory in the element, the connection weight of the front end artificial neuron and the current artificial neuron is read, and according to the connection weight information, the output information of the front end neuron can be used in the calculation process of participating in the current neuron output information. More accurately reflects the weight of the output information of the front-end neurons.

Step S400a, according to the membrane potential information output by the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information, by a preset artificial neuron activation function , calculate the current artificial neuron output information.

specific,

Where Y _i is the membrane potential information of the current neuron output, X _i is the front-end neuron output information, W _ji is the connection weight of the front-end artificial neuron and the current artificial neuron, V _bias is the current artificial neuron bias, f( ·) is the default artificial neuron activation function, commonly used mainly for ReLU function

Sigmoid function

Tanh function

Wait. If applied to chip design, the basic model at the cell body can be simplified as: V _ANN = f (V _input + V _bias ), V _input is the _input of the current beat accumulation, equivalent to the above

The composite mode neuron information processing system provided by the embodiment is configured as an artificial neuron working mode according to the requirements of the task, and can realize most of the current artificial neural network models and support the application of the brain computing platform in the field of machine learning.

In one embodiment, the membrane potential information output according to the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information are preset. After the step of calculating the artificial neuron output information by the artificial neuron activation function, the method further includes:

Reading a potential extremum, and an extremum neuron identifier corresponding to the potential extremum; comparing the current artificial neuron output information with the potential extremum, if the current artificial neuron output information is greater than or And equal to the potential extremum, updating the potential extremum to the current artificial neuron output information, and updating the extremum neuron identifier to an identifier of the current artificial neuron.

Specifically, in the ANN mode, each cell body compares the current artificial neuron output information, that is, the current membrane potential V _ANN with the stored potential extreme value, that is, the transverse maximum potential value V _max between the cells. If V _ANN ≥ V _max , the inter-cell lateral maximum potential value is updated to the current membrane potential V _ANN , V _max =V _ANN , and the maximum cell number N _max is updated to the current cell identification number.

In this embodiment, the current artificial neuron output information is compared with the potential extreme value, and if the current artificial neuron output information is greater than or equal to the potential extreme value, the potential extreme value is updated to The current artificial neuron output information enables the neural network to complete the operation of the pooling step directly in the convolutional layer without allocating physical space for the pooled layer neurons, thereby greatly saving neuron resources, thereby reducing the construction of the neural network. cost.

FIG. 3 is a schematic flowchart of a multi-mode neural network information processing method according to still another embodiment. The multi-mode neural network information processing method shown in FIG. 3 is a refinement process of step S600 of FIG. 1, and includes:

Step S610a, reading the first issuance enablement identifier, where the first issuance enablement identifier includes allowing data to be issued or not allowing data to be issued, when the first issuance enable identifier is allowed to issue data.

Specifically, the first issuance enablement identifier is control information that is determined in the task to determine whether the final neuron output information is issued. When the first release enable identifier is not allowed to issue data, the The calculated artificial neuron output information is not allowed to be issued, and the process ends. When the first issuance enable identifier is allowed to issue data, step S620a is followed.

In step S620, the artificial neuron issuance data type parameter is read, and the artificial neuron issuance data type includes: issuing the current artificial neuron output information, issuing the potential extremum, and issuing an extremum corresponding to the potential extremum A neuron identifies one of them.

Specifically, when the first issuance enable identifier is allowed to issue data, the artificial neuron issuance data type parameter is read, and according to the subsequent calculation requirement, different data types may be selected, for example, to satisfy subsequent convolution. The maximum pooling operation in the neural network requires the output of the potential extremum.

Step S630a, determining, according to the artificial neuron issuing data type parameter, that the current artificial neuron finally loses Information.

Step S640a, outputting the final output information of the current artificial neuron.

In this embodiment, by setting the release enable identifier and the artificial neuron release data type parameter, the current artificial neuron output information is determined, so that the output of the artificial neuron is more controllable, and the release enable flag can be configured. Neurons are not allowed to issue data, but only as intermediate auxiliary computing neurons, which is necessary for some functions that require multiple neurons to work together. In addition, the release enable flag and the release type can work together to set the area of the nerve. The meta-output information is reduced to a current neuron output information (the neuron output information is a maximum value in each neuron output information of the set region), and the direct maximum pooling operation is completed.

4 is a schematic flow chart of a multi-mode neural network information processing method according to still another embodiment, and the multi-mode neural network information processing method shown in FIG. 4 is a configuration parameter of the neuron working mode in the method shown in FIG. When the parameters are configured for the pulsed neuron working mode, the refinement process of step S300 to step S500 includes:

Step S100b: Receive pulse tip information output by the front end pulse neuron, and a connection weight index of the front end pulse neuron and the current pulse neuron.

Specifically, the connection weight index of the front-end pulse neuron and the current pulse neuron is a weight index sent by the front-end neuron together with the output information of the front-end pulse neuron, and is used to indicate the extraction of the current neuron weight. The pulse tip information output by the front end pulse neuron is a pulse tip signal sent by the front end pulse neuron.

Step S200b, reading the current time window width, the pulse tip information sequence in the current time window, the historical membrane potential information, and the membrane potential leakage information.

Specifically, the pulse tip information sequence in the current time window refers to a sequence of information in which the pulse tip information received in a time step within a certain range in the current time window width is sequentially buffered in chronological order.

Step S300b, reading the connection weight of the front-end pulse neuron and the current pulse neuron according to the connection weight index of the front-end pulse neuron and the current pulse neuron.

Specifically, the connection weight index of the front-end pulse neuron and the current pulse neuron is an address information, and the current neuron is indexed according to the received connection weight of the front-end pulse neuron and the current pulse neuron. In the memory in the element, the connection weight of the front-end pulse neuron and the current pulse neuron is read, and according to the connection weight information, the output information of the front-end neuron can be used in the calculation process of participating in the current neuron output information. More accurately reflects the weight of the output information of the front-end neurons, carrying more abundant information.

Step S400b, updating the pulse tip information sequence in the current time window according to the pulse tip information output by the front-end pulse neuron and the pulse tip information sequence in the current time window, and acquiring the pulse tip information in the current time window. Update the sequence.

Specifically, the pulse tip information sequence, after storing a new pulse tip information in the operation step of each pulse neuron, deletes the pulse tip information at the tail position of the sequence, and updates the pulse tip sequence once. .

Step S500b: Calculate the front-end pulse neuron input information by using the attenuation function according to the current time window width and the pulse tip information update sequence in the current time window.

Specifically, use

Calculating the front-end pulse neuron input information, wherein T _w is the time window width, and δ _j is a pulse front information update sequence in the current time window after the front-end neuron j issues a spike in the current time window Time step. K(Δt) is an attenuation function that decreases rapidly as Δt increases.

Step S600b, calculating, according to the front-end pulse neuron input information, the connection weight of the front-end pulse neuron and the current pulsed neuron, the historical membrane potential information, and the membrane potential leakage information, by calculating a pulse neuron calculation model Current pulsed neuron output information.

Specifically, the calculation of the input information of the front-end pulse neuron is represented by the following formula:

Where W _ij is the connection weight of the front-end pulse neuron j and the current pulse neuron i, T _w is the time window width, and δ _j is the front-end neuron j after the spike is issued in the current time window, at the current The time step within the time window of the pulse tip information update sequence. K(Δt) is an attenuation function that decreases rapidly as Δt increases. The basic model at the cell body can be simplified to:

V _SNN =f(V+V _input +V _leak )

The release model and the reset model are unchanged, where V is the historical membrane potential information stored in the memory, and V _input is the _input of the current beat accumulation, equivalent to the above

V _leak is the leak value information.

In this embodiment, the pulse tip information sequence in the current time window is updated according to the pulse tip information output by the front end pulse neuron and the pulse tip information sequence in the current time window, and the pulse tip in the current time window is acquired. The information update sequence can support the spatio-temporal pulse neural network model with time depth according to the current time window width, the connection weight of the front-end pulse neuron and the current pulsed neuron, and the front-end pulse neuron input information through the attenuation function. Compared with the neural network technology scheme with only one time depth, the space-time information coding ability of the pulse neural network can be greatly improved, and the application space of the pulse neural network is enriched.

5 is a schematic flowchart of a multi-mode neural network information processing method according to still another embodiment, and the multi-mode neural network information processing method shown in FIG. 5 is after all the steps shown in FIG. 4, before the step of S600 in FIG. ,include:

In step S100c, a threshold potential is acquired.

Specifically, the acquiring a threshold potential includes: reading a random threshold mask potential, a threshold offset, and a random threshold; The random threshold and the random threshold mask potential are subjected to bitwise AND operation to obtain a threshold random superposition amount; and the threshold potential is determined according to the threshold random superimposition amount and the threshold offset.

The pseudo-random number generator generates a random threshold V _rand , and uses the random threshold and the preset random threshold mask potential V _{mask to perform} a bitwise AND operation to generate a threshold random superposition amount, and then the threshold random superposition amount and the pre- The set threshold offset _{Vth0 is} added to produce a true threshold potential _Vth . Wherein, the initial seed of the pseudo random number generator is given by the configuration register V _seed .

The mask potential V _{mask is} used to limit the range of threshold increments: if V _mask =0, the threshold random superposition amount is also 0, the release mode degenerates to a fixed threshold release, and the fixed threshold is V _th0 ; if V _mask ≠ 0, then The issuance mode is a partial probability threshold issue. When the extreme case V _th0 =0, the issuance mode is a full probability threshold issuance.

Step S200c: Comparing the current pulse neuron output information with the threshold potential, and determining the issuance trigger flag information according to the comparison result, where the issuing trigger flag information includes: issuing a trigger or issuing a non-trigger. When the issuance triggering flag information is the issuance trigger, the process proceeds to step S300c, and when the issuance triggering flag information is not issued, the process proceeds to step S400c.

Specifically, according to the obtained threshold potential, the current pulse neuron output information is compared, and the trigger trigger flag information is determined according to the comparison result. Only when the current pulse neuron output information is greater than or equal to the threshold potential, the current pulse neuron output information may be transmitted.

Step S300c, resetting the refractory period timer, and updating the historical membrane potential information to a preset reset membrane potential information.

Specifically, when the issue trigger flag information is an issue trigger, the current pulse neuron output information may be sent, the refractory period timer is reset, and the historical membrane potential information is updated to a preset membrane potential information. And the historical membrane potential information is updated, and the membrane potential is selectively reset to the current membrane potential, the current membrane potential and the threshold potential difference, or a fixed reset voltage according to the configured reset type Reset_type.

Step S400c, reading the current time step of the refractory period width and the refractory period timer.

Specifically, when the issuing trigger flag information is not triggered by the issuance, the current pulse neuron output information is not sent, and further determining whether the current period is within the refractory period. The refractory period width is greater than the value of the refractory period timer, which is timed by means of a time step.

Step S500c: Determine whether the current time is within the refractory period according to the refractory period width and the current time step of the refractory period timer. If the current time is within the refractory period, step S600c is followed, and if the current time is outside the refractory period, step S700c is followed.

Specifically, according to the cumulative calculation of the current time step of the refractory period timer, it can be determined whether the current time step is still in the refractory period.

Step S600c, accumulating the refractory period timer for one time step, and not updating the historical film potential information.

Specifically, if the current time is within the refractory period, according to the biomimetic characteristics of the pulsed neural network, no response is made to the pulsed nerve output information, and the historical membrane potential information is not updated, and the historical membrane potential information is the next The pulsed neurons of the time step need to read the information, that is, during the refractory period, the pulsed neuron output information calculated this time does not participate in the calculation of the next time step.

Step S700c, accumulating the refractory period timer for one time step, and updating the historical membrane potential information as the current pulse neuron output information.

Specifically, as in the refractory period, the historical membrane potential information is updated to the current pulsed neuron output information to participate in the calculation of the next time step.

In this embodiment, the threshold potential is determined by reading the random threshold mask potential and the threshold offset and receiving the configuration value given by the configuration register, so that the neuron issues the pulse tip information with a certain probability of randomness, regardless of Whether the membrane potential exceeds a fixed threshold bias, and because there is a positive or negative threshold random superposition amount, the neuron cell body may issue pulses, which improves the computational power and information processing capability of the pulse neural network model.

FIG. 6 is a schematic flowchart diagram of a multi-mode neural network information processing method according to still another embodiment, and the multi-mode neural network information processing method shown in FIG. 6 is configured for the neuron working mode in the method shown in FIG. When configuring parameters for the pulsed neuron working mode, the refinement process of step S600 includes:

In step S610b, the second issuance enablement identifier is read, and the second issuance enablement identifier includes permission to issue data or disallowing to issue data, and when the second issuance enable identifier is allowed to issue data.

Specifically, the second issuance enablement identifier is control information that determines whether the final neuron output information is issued. When the second issuance enable identifier is not allowed to issue data, the calculated artificial neuron The output information is not allowed to be issued and the process ends. When the second issuance enable identifier is allowed to issue data, step S620b is followed.

Step S620b, the issuing trigger flag information is read, when the issuing trigger flag information is an issue trigger.

Specifically, when the second issuance enable identifier is the permission method data, it is further required to further determine whether the issue trigger flag is an issue trigger.

Step S630b, outputting the current pulse neuron output information.

In this embodiment, by setting the release enable identifier and the issue trigger flag, the current pulse neuron output information is determined, so that the output of the pulse neuron is more controllable, and the release enable flag can be configured with neurons that are not allowed. Issuing data, but only as an intermediate auxiliary computing neuron, is necessary for some functions that require multiple neurons to work together.

Based on the same inventive concept, an embodiment of the present invention further provides a computer device, including a memory and a processor. And a computer program stored on the memory and operable on the processor, wherein the processor executes the computer program to implement the steps of the method mentioned in the above embodiments.

A person skilled in the art can understand that all or part of the process of implementing the foregoing embodiment can be completed by a computer program or instruction related hardware, and the program can be stored in a computer readable storage medium. When executed, the flow of an embodiment of the methods as described above may be included. Any reference to a memory, storage, database or other medium used in the various embodiments provided herein may include non-volatile and/or volatile memory. Non-volatile memory can include read only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in a variety of formats, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronization chain. Synchlink DRAM (SLDRAM), Memory Bus (Rambus) Direct RAM (RDRAM), Direct Memory Bus Dynamic RAM (DRDRAM), and Memory Bus Dynamic RAM (RDRAM).

FIG. 7 is a schematic structural diagram of a multi-mode neural network information processing system according to an embodiment. The multi-mode neural network information processing system shown in FIG. 7 includes:

The neuron working mode reading module 100 is configured to read a neuron working mode configuration parameter, where the neuron working mode configuration parameter comprises an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter.

The neuron working mode configuration module 200 is configured to configure a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode.

The front-end neuron output information receiving module 300 is configured to receive front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output information or front-end pulse neuron output information.

The current neuron information reading module 400 is configured to read current neuron information, where the current neuron information includes current artificial neuron information or current pulsed neuron information.

The current neuron output information calculation module 500 is configured to calculate current neuron output information according to the front end neuron output information and the current neuron information, including outputting information according to the front end artificial neuron and the current artificial neural network. Meta-information, calculating current artificial neuron output information, or calculating current pulse neuron output information according to the front-end pulse neuron output information and the current pulsed neuron information.

The current neuron output information output module 600 is configured to output the current artificial neuron output information or the current pulse neuron output information.

The composite mode neuron information processing system provided by the invention provides a corresponding artificial neuron working mode or a pulsed neuron working mode according to a preset neuron working mode configuration parameter, and passes the corresponding neuron working mode. Receiving front-end neuron output information, and reading current neuron information, calculating current artificial neuron output information or current pulsed neuron output information, and outputting. The composite mode neuron information processing system provided by the invention can configure the corresponding neuron working mode according to the requirements of the task, and only needs to modify the working mode of the neuron when switching between different tasks in different neural network working modes. The configuration parameters can be different from the neural network accelerator scheme that only supports the artificial neural network mode, and the neuromorphic scheme that only supports the pulse neural network. The present invention can support the artificial neural network mode based on the same architecture. Machine learning applications and computational neuroscience applications based on pulsed neural networks enrich the type of processing information for brain-like computing platforms, reduce the cost of multi-task execution in different neural network operating modes, and improve the working mode of different neural networks. The efficiency of multitasking.

FIG. 8 is a schematic structural diagram of a multi-mode neural network information processing system according to another embodiment. The multi-mode neural network information processing system shown in FIG. 8 includes:

The front end artificial neuron output information receiving unit 100a is configured to receive the membrane potential information output by the front end artificial neuron, and the connection weight index of the front artificial liver neuron and the current artificial neuron.

The current artificial neuron information reading unit 200a is configured to read current artificial neuron offset information.

The artificial neuron connection weight reading unit 300a is configured to read the connection weight of the pre-artificial neuron and the current artificial neuron according to the connection weight index of the front-end artificial neuron and the current artificial neuron.

The current artificial neuron output information calculation unit 400a is configured to: according to the membrane potential information output by the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron offset information, The current artificial neuron output information is calculated by a preset artificial neuron activation function.

FIG. 9 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment. The multi-mode neural network information processing system shown in FIG. 9 includes:

a potential extremum reading unit 500a for reading a potential extremum and an extremum neuron identifier corresponding to the potential extremum;

The potential extreme value comparison unit 600a is configured to compare the current artificial neuron output information with the potential extreme value, and if the current artificial neuron output information is greater than or equal to the potential extreme value,

The potential extremum updating unit 700a is configured to update the potential extremum to the current artificial neuron output information, and update the extremum neuron identifier to an identifier of the current artificial neuron.

In this embodiment, the current artificial neuron output information is compared with the potential extreme value, and if the current artificial neuron output information is greater than or equal to the potential extreme value, the potential extreme value is updated to The artificial neuron outputs information, so that the neural network completes the operation of the pooling step directly in the convolution layer, without allocating physical space for the pooled layer neurons, greatly saving neuron resources, thereby reducing the construction cost of the neural network. .

FIG. 10 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment. The multi-mode neural network information processing system shown in FIG. 10 includes:

The first issuance enablement reading unit 800a is configured to read the first issuance enablement identifier, where the first issue enablement identifier includes allowing data to be released or not allowing data to be issued, when the first issue enablement identifier is When data is allowed to be issued.

The artificial neuron issue data type reading unit 900a is configured to read an artificial neuron release data type parameter, and the artificial neuron issue data type includes: issuing the current artificial neuron output information, and issuing the potential extremum, One of the extreme neuron identifiers corresponding to the potential extremum is issued.

The artificial neuron issuance data type determining unit 910a is configured to determine the final output information of the current artificial neuron according to the artificial neuron issuing data type parameter.

The current artificial neuron output information output unit 920a is configured to output the final output information of the current artificial neuron.

In this embodiment, by setting the release enable identifier and the artificial neuron release data type parameter, the current artificial neuron output information is determined, so that the output of the artificial neuron is more controllable, and the release enable flag can be configured. Neurons are not allowed to issue data, but only as intermediate auxiliary computational neurons, which is necessary for some functions that require multiple neurons to work together. In addition, the release enable flag and the release type can work together to set the area. The neuron output information is reduced to one current neuron output information (the neuron output information is a maximum value in each neuron output information of the set region), and the direct maximum pooling operation is completed.

11 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment. The multi-mode neural network information processing system shown in FIG. 11 includes:

The front end pulse neuron output information receiving unit 100b is configured to receive pulse tip information output by the front end pulse neuron, and a connection weight index of the front end pulse neuron and the current pulse neuron.

The current pulse neuron information reading unit 200b is configured to read the current time window width, the current time window intrapulse information sequence, the historical membrane potential information, and the membrane potential leakage information.

The pulse neuron connection weight reading unit 300b is configured to be based on the front end pulse neuron and the current pulse neuron The weight index is connected, and the connection weight of the front-end pulse neuron and the current pulse neuron is read.

The time-window pulse tip information sequence updating unit 400b is configured to update the current time window intra-pulse tip information sequence according to the pulse tip information output by the front-end pulse neuron and the current time-window pulse tip information sequence. Get the pulse tip information update sequence in the current time window.

The front end pulse neuron input information calculation unit 500b is configured to calculate front end pulse neuron input information by using an attenuation function according to the current time window width and the current time window intrapulse information update sequence.

The pulse neuron output information calculation unit 600b is configured to: according to the front end pulse neuron input information, a connection weight of the front end pulse neuron and the current pulse neuron, the historical membrane potential information, the membrane potential leakage information, The current pulsed neuron output information is calculated by a pulsed neuron calculation model.

12 is a schematic structural diagram of a multi-mode neural network information processing system according to still another embodiment. The multi-mode neural network information processing system shown in FIG. 12 includes:

a threshold potential obtaining unit 700b, configured to acquire a threshold potential; comprising a threshold information receiving subunit for reading a random threshold mask potential, a threshold offset, and a random threshold; and a threshold random superposition amount obtaining subunit, configured to: The threshold value and the random threshold mask potential are bitwise AND operated to obtain a threshold random superposition amount; the threshold potential determination subunit is configured to determine the threshold potential according to the threshold random superposition amount and the threshold offset.

The issue trigger determination unit 800b is configured to compare the current pulse neuron output information with the threshold potential, and determine the issue trigger flag information according to the comparison result, where the issue trigger flag information includes: issuing a trigger or issuing a non-trigger.

When the issuing trigger flag information is an issue trigger,

The triggering action unit 910b is configured to reset the refractory period timer and update the historical membrane potential information to a preset reset membrane potential information.

When the issue trigger flag information is not triggered by the issue,

a non-trigger action unit 920b, including a refractory period sub-unit for reading a current time step of the refractory period width and the refractory period timer; and according to the refractory period width and the refractory period timer Current time step, determine current time Whether it is within the refractory period, if the refusal period judges that the subunit judges that the current time is within the refractory period, the subunit should not be run during the period, and is used to accumulate the refractory period timer for one time step. The historical membrane potential information is not updated; if the refusal period judges that the subunit determines that the current time is not within the refractory period, the subunit may not be run out of time, for accumulating the refractory period timer A time step and updating the historical membrane potential information for the current pulsed neuron output information.

The second issuance enablement reading unit 930b is configured to read the second issuance enablement identifier, where the second issue enablement identifier includes allowing data to be issued or not allowed to be issued, and when the second issue enablement identifier is When the data is allowed to be distributed, the issuing trigger flag information reading unit is configured to read the issuing trigger flag information, and when the issuing trigger flag information is an issue trigger; the current pulse neuron output information output unit is configured to output The current pulsed neuron outputs information.

The technical features of the above-described embodiments may be arbitrarily combined. For the sake of brevity of description, all possible combinations of the technical features in the above embodiments are not described. However, as long as there is no contradiction between the combinations of these technical features, All should be considered as the scope of this manual.

The above-mentioned embodiments are merely illustrative of several embodiments of the present invention, and the description thereof is more specific and detailed, but is not to be construed as limiting the scope of the invention. It should be noted that a number of variations and modifications may be made by those skilled in the art without departing from the spirit and scope of the invention. Therefore, the scope of the invention should be determined by the appended claims.

Claims

A composite mode neuron information processing method, characterized in that the method comprises:

Reading a neuron working mode configuration parameter, the neuron working mode configuration parameter comprising an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter;

And configuring a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode;

Receiving front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output information or front-end pulse neuron output information;

Reading current neuron information, the current neuron information including current artificial neuron information or current pulsed neuron information;

Calculating current neuron output information according to the front-end neuron output information and the current neuron information, including calculating current artificial neuron output information according to the front-end artificial neuron output information and the current artificial neuron information, Or calculating current pulse neuron output information according to the front end pulse neuron output information and the current pulse neuron information;

Outputting the current artificial neuron output information or the current pulsed neuron output information.
The composite mode neuron information processing method according to claim 1, wherein:

The front end artificial neuron output information includes: membrane potential information output by the front end artificial neuron, and a connection weight index of the front end artificial neuron and the current artificial neuron;

The current artificial neuron information includes: current artificial neuron offset information;

And calculating, according to the membrane potential information output by the front end artificial neuron and the current artificial neuron information, the current artificial neuron output information, including:

And according to the connection weight index of the front end artificial neuron and the current artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron is read;

And calculating the current state according to the membrane potential information output by the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information by using a preset artificial neuron activation function. Artificial neurons output information.
The composite mode neuron information processing method according to claim 2, wherein the membrane potential information outputted by the front end artificial neuron, the connection weight of the front end artificial neuron and the current artificial neuron, The current artificial neuron bias information, after the step of calculating the current artificial neuron output information by using a preset artificial neuron activation function, the method further includes:

Reading a potential extremum and an extremum neuron identifier corresponding to the potential extremum;

Comparing the current artificial neuron output information with the potential extremum if the current artificial neuron output If the information is greater than or equal to the potential extreme value, then

Updating the potential extremum to the current artificial neuron output information, and updating the extremum neuron identifier to an identifier of the current artificial neuron.
The composite mode neuron information processing method according to claim 3, wherein the outputting the current artificial neuron output information comprises:

Reading the first issuance enable identifier, where the first issue enable identifier includes allowing data to be released or not allowing data to be issued;

Reading the artificial neuron issuing data type parameter when the first issuance enable identifier is allowed to issue data, the artificial neuron issuing the data type, including: issuing the current artificial neuron output information, and issuing the potential pole And extracting one of the extreme value neuron identifiers corresponding to the potential extreme values;

Determining, according to the artificial neuron, a data type parameter, determining a final output information of the current artificial neuron;

Outputting the final output information of the current artificial neuron.
The composite mode neuron information processing method according to claim 1, wherein:

The front-end pulse neuron output information includes: pulse tip information output by the front-end pulse neuron, and a connection weight index of the front-end pulse neuron and the current pulse neuron;

The current pulse neuron information includes: a current time window width, a pulse tip information sequence in a current time window, historical membrane potential information, and membrane potential leakage information;

And calculating, according to the front-end pulse neuron output information and the current pulse neuron information, the current pulse neuron output information, including:

And according to the connection weight index of the front-end pulse neuron and the current pulse neuron, the connection weight of the front-end pulse neuron and the current pulse neuron is read;

And updating the pulse tip information sequence in the current time window according to the pulse tip information output by the front-end pulse neuron and the pulse tip information sequence in the current time window, and acquiring a pulse tip information update sequence in the current time window;

And calculating, according to the current time window width, the pulse tip information update sequence in the current time window, the front end pulse neuron input information by using an attenuation function;

Calculating the current pulsed nerve by using a pulse neuron calculation model according to the front end pulse neuron input information, the connection weight of the front end pulse neuron and the current pulse neuron, the historical membrane potential information, and the membrane potential leakage information Meta output information.
The composite mode neuron information processing method according to claim 5, characterized in that, according to the front end pulse neuron input information, the connection weight of the front end pulse neuron and the current pulse neuron, the historical membrane potential The information, the membrane potential leakage information, the step of calculating the current pulsed neuron output information by the pulse neuron calculation model, and the step of outputting the current pulse neuron output information, the method further includes:

Obtaining a threshold potential;

Comparing the current pulsed neuron output information with the threshold potential, and determining the issuance triggering flag information according to the comparison result, where the issuing triggering flag information includes: issuing a trigger or issuing a non-trigger; and when the issuing triggering flag information is issued When triggered,

The refractory period timer is reset, and the historical membrane potential information is updated to a preset reset membrane potential information.
The composite mode neuron information processing method according to claim 6, wherein after the current pulse neuron output information is compared with the threshold potential, and the step of issuing trigger flag information is determined based on the comparison result, The method also includes:

When the issue trigger flag information is not triggered by the issue, the current time step of the refractory period width and the refractory period timer is read;

Determining whether the current time is within the refractory period according to the refractory period width and the current time step of the refractory period timer, and if the current time is within the refractory period, accumulating the refractory period timer Timing a time step without updating the historical membrane potential information;

If the current time is not within the refractory period, the refractory period timer is cumulatively counted for one time step, and the historical membrane potential information is updated to the current pulsed neuron output information.
The composite mode neuron information processing method according to claim 6, wherein the acquiring the threshold potential comprises:

Reading a random threshold mask potential, a threshold offset, and a random threshold;

Performing a bitwise AND operation on the random threshold and the random threshold mask potential to obtain a threshold random superposition amount;

The threshold potential is determined based on the threshold random amount and the threshold offset.
The composite mode neuron information processing method according to claim 6, wherein the outputting the current pulse neuron output information comprises:

Reading a second issuance enablement identifier, where the second issuance enablement identifier includes allowing data to be issued or not allowing data to be issued, and when the second issuance enablement identifier is allowed to issue data,

Reading the issue trigger flag information, when the issue trigger flag information is an issue trigger;

Outputting the current pulse neuron output information.
A composite mode neuron information processing system, comprising:

a neuron working mode reading module, configured to read a neuron working mode configuration parameter, where the neuron working mode configuration parameter comprises an artificial neuron working mode configuration parameter or a pulsed neuron working mode configuration parameter;

a neuron working mode configuration module, configured to configure a current neuron working mode according to the neuron working mode configuration parameter, where the current neuron working mode includes an artificial neuron working mode or a pulsed neuron working mode;

The front-end neuron output information receiving module is configured to receive front-end neuron output information, where the front-end neuron output information includes front-end artificial neuron output information or front-end pulse neuron output information;

a current neuron information reading module, configured to read current neuron information, where the current neuron information includes current artificial neuron information or current pulsed neuron information;

The current neuron output information calculation module is configured to calculate current neuron output information according to the front end neuron output information and the current neuron information, including outputting information according to the front end artificial neuron and the current artificial neuron Information, calculating current artificial neuron output information, or calculating current pulse neuron output information according to the front end pulse neuron output information and the current pulse neuron information;

The current neuron output information output module is configured to output the current artificial neuron output information or the current pulse neuron output information.
A computer device, comprising: a memory, a processor, and a computer program stored on the memory and operable on the processor, the processor executing the computer program to implement any one of claims 1-9 The steps of the method.