+

WO2018133567A1 - Procédé et système de traitement d'informations de poids de neurone, procédé et système de traitement d'informations neuronales, et dispositif informatique - Google Patents

Procédé et système de traitement d'informations de poids de neurone, procédé et système de traitement d'informations neuronales, et dispositif informatique Download PDF

Info

Publication number
WO2018133567A1
WO2018133567A1 PCT/CN2017/114659 CN2017114659W WO2018133567A1 WO 2018133567 A1 WO2018133567 A1 WO 2018133567A1 CN 2017114659 W CN2017114659 W CN 2017114659W WO 2018133567 A1 WO2018133567 A1 WO 2018133567A1
Authority
WO
WIPO (PCT)
Prior art keywords
neuron
information
weight
neurons
group
Prior art date
Application number
PCT/CN2017/114659
Other languages
English (en)
Chinese (zh)
Inventor
裴京
邓磊
施路平
吴臻志
田雷
李国齐
Original Assignee
清华大学
Priority date (The priority date 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 date listed.)
Filing date
Publication date
Priority claimed from CN201710042090.4A external-priority patent/CN106815638B/zh
Priority claimed from CN201710042087.2A external-priority patent/CN106875010B/zh
Application filed by 清华大学 filed Critical 清华大学
Publication of WO2018133567A1 publication Critical patent/WO2018133567A1/fr

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons

Definitions

  • the receiving neuron output information of the front-end neuron, and the weight index corresponding to the neuron output information include:
  • the current neuron information when the collaborative group is a pulse cooperative group, includes historical membrane potential information;
  • the historical membrane potential information of the pulsed effective neurons is updated.
  • the neuron is set to allow data to be issued or not allowed to be distributed, and a predetermined number of consecutive neurons are grouped into a coordinated group, and the cooperative group can be flexibly configured according to requirements. .
  • the pulse synergy group after the pulse effective neuron outputs the cooperative output information, the historical membrane potential information of the pulsed effective neurons is updated, so that the entire cooperative group completes subsequent information processing, and the pulse The cooperative neurons do not update the historical membrane potential information, and in the subsequent information processing, the function of weight expansion is completed, and the information processing capability of the entire pulse neural network is improved by the pulse synergy group.
  • the weight information obtaining module is configured to: according to the weight index, read the correspondence of the weight index information, and obtain the weight information, where the weight index information correspondence relationship is a correspondence between the weight index and the weight information;
  • the front end neuron comprises an artificial neuron or a pulsed neuron.
  • the weight index includes a storage address of weight information corresponding to the weight index.
  • the front-end neuron outputs an information receiving module, including:
  • the invention also provides a neuron information processing system with input weight expansion, comprising:
  • a collaborative group determining module configured to determine a preset number of consecutive neurons as a collaborative group, and determine a last neuron in the collaborative group as a valid neuron, except the effective neurons in the collaborative group
  • the neurons are identified as synergistic neurons
  • FIG. 2 is a schematic flow chart of a method for processing a neuron weight information according to another embodiment
  • FIG. 3 is a schematic structural diagram of a neuron weight information processing system according to an embodiment
  • FIG. 4 is a schematic structural diagram of a neuron weight information processing system of another embodiment
  • FIG. 5 is a schematic flow chart of a neuron information processing method for input weight expansion according to an embodiment
  • FIG. 6 is a schematic flow chart of a neuron information processing method for input weight expansion in another embodiment
  • FIG. 7 is a schematic structural diagram of a neuron information processing system with an input weight extension according to an embodiment
  • FIG. 1 is a schematic flowchart of a method for processing a neuron weight information according to an embodiment.
  • the method for processing a neuron weight information as shown in FIG. 1 includes:
  • the neuron output information of the front-end neuron is output information calculated by the front-end neuron;
  • the weight index is index information for the current neuron to retrieve the weight information corresponding to the front-end neuron output information.
  • the weight indexing method can occupy a smaller information transmission space in the process of information transmission, which not only reduces the processing requirements of the hardware, but also needs to change the index information, so that the change of the weight information can be more flexibly and conveniently performed.
  • the update makes it easier to update the weight information in the neural network.
  • Step S200 The weight index is read according to the weight index, and the weight index information is obtained, and the weight index information correspondence relationship is a correspondence between the weight index and the weight information.
  • the weight index information correspondence relationship may be stored locally in the current neuron or may be stored in other locations in the neural network as long as the current neuron can be read.
  • Step S300 acquiring input information of the front-end neuron according to the weight information and the neuron output information.
  • the read weight information is subjected to corresponding operation processing according to different neuron models to obtain input information of the front end neurons.
  • the front-end neuron including artificial neurons or pulsed neurons, ie, front-end neurons and current neurons, may be artificial neurons or pulsed neurons, ie, artificial neural networks or pulsed nerves.
  • the meta-network can use the weight indexing method in the transmission of the weight information.
  • the neuron output information of the front-end neuron and the weight index corresponding to the neuron output information include: 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; the reading the weight index information corresponding relationship according to the weight index, and obtaining the weight information, including: according to the front end artificial neuron and the current artificial neuron Connect the weight index to read the connection weight between the front-end artificial neurons and the current artificial neurons.
  • the neuron output information of the receiving front end neuron, and the weight index corresponding to the neuron output information include: pulse tip information output by the front end pulse neuron, front end pulse The weighting index of the connection between the neuron and the current pulsed neuron; reading the weighted index information correspondence according to the weight index, and obtaining the weight information, including: reading 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 pulsed neuron is taken.
  • the use of the weight index is applicable not only to the artificial neural network but also to the pulse neural network, and improves the information processing capability of the artificial neural network and the pulsed neural network.
  • the weight index includes a storage address of weight information corresponding to the weight index.
  • the storage address of the weight information is used as the index information, so that the neuron receiving the index information directly uses the storage address information to query the weight information, thereby improving the extraction efficiency of the weight information, thereby improving the entire neural network. Information processing efficiency.
  • the receiving neuron output information of the front-end neuron and the weight index corresponding to the neuron output information include: receiving routing information output by the front-end neuron, the routing information including the front-end neural network Meta-outputted neuron output information, and a weight index corresponding to the neuron output information; parsing the routing information, and acquiring the neuron output information and the weight index.
  • the weight index is set in the routing information, and the weight index information is transmitted by using the information transmission data in the existing neural network.
  • the index information may be stored using fixed length or variable length information bits.
  • the weight index and the neuron output information are set and transmitted in the routing information, and the existing routing data is fully utilized, thereby improving the information use efficiency between the neurons.
  • the weight information calculated by the weight reduction algorithm is used to limit the weight to some fixed value according to the preset weight value range and the initial weight information.
  • the preset weight value range is a range formed by a maximum and a minimum weight required in the network; and the weight reduction algorithm may discretize the weight under the premise of ensuring the accuracy of the algorithm. For some fixed values, such as weight binarization, binarization algorithm.
  • the weight information calculated by the weight reduction algorithm is reduced for the storage weight.
  • the storage space of the hardware of the information, and the accuracy of the weight information is maintained.
  • FIG. 2 is a schematic flowchart diagram of a method for processing a neuron weight information according to another embodiment.
  • the method for processing a neuron weight information as shown in FIG. 2 includes:
  • Step S100 Receive neuron output information of the front end neuron, and a weight index corresponding to the neuron output information.
  • Step S200 The weight index is read according to the weight index, and the weight index information is obtained, and the weight index information correspondence relationship is a correspondence between the weight index and the weight information.
  • Step S300 acquiring input information of the front-end neuron according to the weight information and the neuron output information.
  • Step S400 calculating current neuron output information according to the neuron output algorithm according to the input information of the front end neuron and the read current neuron information.
  • the current artificial neuron information includes: a current artificial nerve Meta-bias information; then, according to the input information of the front-end neuron and the read current neuron information, calculating current neuron output information according to a neuron output algorithm, including: outputting according to the front-end artificial neuron
  • the membrane potential information, the connection weight of the front end artificial neuron and the current artificial neuron, and the current artificial neuron bias information are used to calculate the current artificial neuron output information through a preset artificial neuron activation function.
  • the current pulsed neuron information includes: historical membrane potential information and membrane potential leakage information; then the input information according to the front-end neuron and the read current neuron information Calculating, according to the neuron output algorithm, the current neuron output information, including: 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, calculated by pulse neuron calculation model, current pulse neuron output information
  • Step S500 determining destination information of the current neuron output information, and searching for a destination index correspondence relationship according to the destination information, and acquiring a weight index of the destination information, where the destination index correspondence relationship includes between the destination information and the weight index Correspondence.
  • the connection relationship between the current neuron and the back-end neuron has been determined, and the destination information of the current neuron output information has also been determined.
  • the destination information of the current neuron output information the destination index correspondence relationship is searched, and the corresponding weight index can be used.
  • Step S600 outputting the current neuron output information and the weight index.
  • outputting the current neuron output information and the weight index to the neurons of the back end can complete the entire transfer process of the weight index.
  • the current neuron calculates the output information of the current neuron for output, according to the target neuron of the output information, after searching for the corresponding weight index, the current neuron output information and the weight are output. index.
  • the current neuron sends the weight index to the neurons in the back end, and the weight index information is completely transmitted in the neural network, which improves the information processing capability of the neural network.
  • FIG. 3 is a schematic structural diagram of a neuron weight information processing system according to an embodiment.
  • the neuron weight information processing system shown in FIG. 3 includes:
  • the front-end neuron output information receiving module 100 is configured to receive neuron output information of the front-end neuron, and a weight index corresponding to the neuron output information; the front-end neuron includes an artificial neuron or a pulsed neuron.
  • the weight index includes a storage address of the weight information corresponding to the weight index.
  • the front-end neuron output information receiving module includes: a routing information receiving unit, configured to receive routing information output by the front-end neuron, where the routing information includes neuron output information output by the front-end neuron, and output with the neuron A weight index corresponding to the information; a routing information parsing unit, configured to parse the routing information, and obtain the neuron output information and the weight index.
  • the front-end neuron input information obtaining module 300 is configured to acquire input information of the front-end neuron according to the weight information and the neuron output information.
  • the output information of the front-end neuron received by the current neuron carries a weight index of the weight information between the front-end neuron and the current neuron, and the current neuron reads the weight information according to the received weight index information.
  • corresponding operation processing is performed to obtain input information of the front-end neurons.
  • the weight information is no longer transmitted directly between the neurons, but the index of the weight information is transmitted, which not only saves the amount of information transmitted between the networks, but also can change the setting of the weight information of each neuron more flexibly, and improves the setting.
  • Information processing capabilities of neural networks are examples of information from the neural networks.
  • weight index is applicable not only to artificial neural networks, but also to pulsed neural networks, which improves the information processing capabilities of artificial neural networks and pulsed neural networks.
  • the storage address of the weight information is used as the index information, so that the neuron receiving the index information directly uses the storage address information to query the weight information, thereby improving the extraction efficiency of the weight information, thereby improving the information processing efficiency of the entire neural network.
  • the weight index and the neuron output information are set and transmitted in the routing information, and the existing routing data is fully utilized, thereby improving the information use efficiency between the neurons.
  • the weight information calculated by the weight reduction algorithm has a weight value within a preset value range, which reduces the storage space of the hardware for storing the weight information, and maintains the weight. The accuracy of the information.
  • FIG. 4 is a schematic structural diagram of a neuron weight information processing system according to another embodiment, and the neuron weight information processing system shown in FIG. 4 includes:
  • the front-end neuron output information receiving module 100 is configured to receive neuron output information of the front-end neuron, and a weight index corresponding to the neuron output information.
  • the weight information obtaining module 200 is configured to read the weight index information corresponding relationship according to the weight index, and obtain weight information, where the weight index information correspondence relationship is a correspondence between the weight index and the weight information.
  • the front-end neuron input information obtaining module 300 is configured to acquire input information of the front-end neuron according to the weight information and the neuron output information.
  • the current neuron output information obtaining module 400 is configured to calculate current neuron output information according to the neuron output algorithm according to the input information of the front end neuron and the read current neuron information.
  • the weight index determining module 500 is configured to determine destination information of the current neuron output information, and search for a destination index correspondence according to the destination information, and obtain a weight index of the destination information, where the destination index correspondence includes the destination information. The correspondence between the weight index and the weight index.
  • the weight index sending module 600 is configured to output the current neuron output information and the weight index.
  • the current neuron calculates the output information of the current neuron for output, according to the target neuron of the output information, after searching for the corresponding weight index, the current neuron output information and the weight are output. index.
  • the current neuron sends the weight index to the neurons at the back end, and the weight index information is completely transmitted in the neural network. Improve the information processing capabilities of neural networks.
  • FIG. 5 is a schematic flowchart of a method for processing neuron information of an input weight extension according to an embodiment, and the method for processing neuron information of the input weight extension shown in FIG. 1 includes:
  • Step S400' the valid neuron acquires cooperative output information according to the received front-end neuron information, the read current neuron information of the effective neuron, and the horizontal accumulation information.
  • step S500' the effective neurons output the coordinated output information.
  • the weight information of the plurality of input groups breaks the shortcoming of the limited type of input weight of the existing single neuron, and improves the information processing capability of the neural network.
  • the setting is issued with an enable identifier for setting a determined preset number of consecutive neurons as a cooperative group, and setting only the last neuron to output information.
  • a predetermined number of consecutive artificial neurons are determined as an artificial cooperative group, and the last artificial neuron in the artificial cooperative group is determined as an artificial effective neuron, and the artificial cooperative group is divided.
  • the artificial neurons outside the artificial effective neurons are determined as artificial cooperative neurons; or a preset number of consecutive pulse neurons are determined as a pulse cooperative group, and the last pulse neuron in the pulse cooperative group is determined as a pulse An effective neuron, wherein the pulsed neurons other than the pulsed effective neurons in the pulse synergy group are determined as pulse cooperative neurons.
  • a predetermined number of consecutive artificial neurons are determined as an artificial cooperative group, or a preset number of consecutive pulse neurons are determined as a pulse cooperative group, in an artificial neural network or a pulsed neural network,
  • the cooperative group can be determined to expand the input weight of a single neuron, and improve the information processing capability of the artificial neural network or the pulse neural network.
  • the current neuron information includes historical membrane potential information; after the step of outputting the cooperative output information by the effective neuron, the method further includes: updating the pulse effective The historical membrane potential information of neurons.
  • the pulse effective neurons output the coordinated output information
  • the historical membrane potential information of the pulsed effective neurons is updated, so that the entire collaborative group completes subsequent information processing, and the pulse synergy
  • the neurons do not update the historical membrane potential information, and in the subsequent information processing, the function of weight expansion is completed, and the information processing capability of the entire pulse neural network is improved through the pulse synergy group.
  • FIG. 6 is a schematic flowchart diagram of a method for processing neuron information of an input weight extension according to another embodiment, and the method for processing neuron information of the input weight extension shown in FIG. 2 includes:
  • the front-end neuron information includes: front-end neuron output information, a connection weight index of the front-end neuron and the current neuron.
  • the front-end neuron information includes: membrane potential information outputted by the front-end artificial neurons, and a connection weight index of the front-end artificial neurons and the current artificial neurons.
  • the front-end neuron information includes: pulse tip information outputted by the front-end pulse neuron, and a connection weight index of the front-end pulse neuron and the current pulse neuron.
  • Step S100a' determining a preset number of consecutive neurons as a cooperative group, determining a last neuron in the cooperative group as a valid neuron, and selecting a neuron other than the effective neuron in the cooperative group Determined as a synergistic neuron.
  • Step S200a' the first coordinated neuron in the cooperative group reads the connection weight of the front-end neuron and the current neuron according to the connection weight index of the front-end neuron and the current neuron; according to the front-end neuron and The connection weight of the current neuron, the front-end neuron information, and the horizontal accumulation intermediate information of the first coordinated neuron.
  • connection weight index of the front end artificial neuron and the current artificial neuron is an address information
  • the current neuron is indexed according to the received connection weight of the front end artificial neuron and the current artificial neuron
  • the current nerve is
  • 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.
  • the front-end neuron information includes membrane potential information output by the front-end artificial neuron, based on the membrane potential information output by the front-end artificial neuron, and the read front-end neuron and current nerve
  • the weight of the connection of the element after multiplication, obtains the horizontal accumulation intermediate information of the first artificial coordinated neuron and puts it into the accumulator.
  • the front-end neuron information includes pulse tip information output by the front-end pulse neuron, according to pulse tip information output by the front-end pulse neuron, and the read front-end neuron and current nerve
  • the weight of the connection of the element after multiplication, obtains the horizontal accumulation intermediate information of the first pulse cooperative neuron and puts it into the accumulator.
  • Step S300a' the subsequent cooperative neuron in the collaborative group sequentially reads the connection weight of the front-end neuron and the current neuron according to the connection weight index of the front-end neuron and the current neuron; according to the front-end neuron Obtaining intermediate information of the connection weights of the current neurons, the front-end neuron information, and the lateral accumulation of the front-end cooperative neurons, acquiring the horizontal accumulation intermediate information of the cooperative neurons, and the last synergistic nerve in the synergistic group The horizontal accumulation intermediate information of the element is determined as the horizontal accumulation information.
  • the subsequent cooperative neurons in the collaborative group separately calculate the received front-end neuron output information and the connected front-end neuron and the current neuron, and calculate according to the preset neuron mode, such as After multiplying, the horizontal accumulation intermediate information of the coordinated neurons of the front end connected thereto is accumulated to obtain the horizontal accumulation intermediate information of the current coordinated neurons. Until the last coordinated neuron obtains the horizontal accumulation intermediate information, it is confirmed as the horizontal accumulation information.
  • Step S400a' the valid neuron acquires cooperative output information according to the received front-end neuron information, the read current neuron information of the effective neuron, and the horizontal accumulation information.
  • the current neuron information includes current artificial neuron offset information.
  • the effective neuron obtains collaborative output information according to the received front-end neuron information, the read current neuron information of the effective neuron, and the horizontal accumulation information, including: a membrane output according to the front-end artificial neuron Potential information,
  • the connection weight of the front-end neuron and the current neuron, the current artificial neuron bias information, and the cooperative output information of the artificial effective neuron are calculated by a preset artificial neuron activation function.
  • the current neuron information includes historical membrane potential information and membrane potential leakage information.
  • the effective neuron obtains cooperative output information according to the received front-end neuron information, the read current neuron information of the effective neuron, and the horizontal accumulation information, including: according to the pulse output by the front-end pulse neuron
  • the cutting-edge information, the connection weight of the front-end neuron and the current neuron, the historical membrane potential information, and the membrane potential leakage information are calculated by a pulsed neuron calculation model to calculate cooperative output information of the pulsed effective neurons.
  • connection weight of the front-end neuron and the current neuron in the received front-end neuron information is indexed, and the connection weight of the front-end neuron and the current neuron is read, it is used to calculate the horizontal accumulation intermediate information, and
  • the weight information of each coordinated neuron in a collaborative group is fully utilized, and in the cooperative output information of the effective neuron output, the weight information of each coordinated neuron is embodied, which is equivalent to the weight information of the effective neurons. It has been expanded to improve the information processing capabilities of neural networks.
  • the collaborative group determining module 100' is configured to determine a preset number of consecutive neurons as a cooperative group, and determine a last neuron in the collaborative group as a valid neuron, and save the effective neural group in the cooperative group
  • the neurons outside the element are determined as cooperative neurons; and are used for setting an issue enablement identifier of the neurons in the collaborative group, the issue enablement identifier includes allowing data to be released or not allowing data to be released, and the effective neurons are
  • the issuance enablement flag is set to allow the release of data, and the issue enablement identification of all of the coordinated neurons is set to not allow the release of data.
  • the method includes: an artificial neuron determining unit, configured to determine a preset number of consecutive artificial neurons as an artificial cooperative group, and determine a last artificial neuron in the artificial cooperative group as an artificial effective neuron, and the artificial synergy
  • the artificial neurons other than the artificial effective neurons in the group are determined as artificial cooperative neurons
  • the pulsed neuron determining unit is configured to determine a preset number of consecutive pulse neurons as a pulse synergy group, and the pulse The last pulsed neuron in the cooperative group is determined as a pulsed effective neuron, and the pulsed neurons other than the pulsed effective neurons in the pulse cooperative group are determined as pulse cooperative neurons.
  • the horizontally accumulated information obtaining module 200 ′ is configured to: acquire, by the first coordinated neuron in the collaborative group, the horizontal accumulated intermediate information of the first coordinated neuron according to the received front-end neuron information; Each of the cooperative neurons sequentially acquires the horizontal accumulation intermediate information of the cooperative neurons according to the received front-end neuron information and the lateral accumulation intermediate information of the front-end cooperative neurons, and the last synergistic nerve in the synergistic group
  • the horizontal accumulation intermediate information of the element is determined as horizontal accumulation information
  • the horizontal accumulation intermediate information of the last coordinated neuron in the cooperation group is determined as horizontal accumulation information
  • the front end neuron information includes: front end neuron output information, front end a weighted index of the connection between the neuron and the current neuron
  • the horizontal accumulation information acquisition module 200' is used to coordinate the first coordinated neuron in the group, the root According to the connection weight index of the front-end neuron and the current neuron, the connection weight of the front-end neuron and the current neuron is read; according to the connection weight of
  • the horizontal accumulation information acquisition module 200 ′ when implementing the hardware circuit by using specific components, the horizontal accumulation intermediate information generated by each coordinated neuron in the cooperation group is transmitted to the next through the shared register. Synergistic neurons or effective neurons are used for membrane potential accumulation, and this way of feedback addition can be achieved with an accumulator. More specifically, the cooperative neuron obtains the horizontal accumulation intermediate information of the front-end cooperative neurons by reading the shared register. After the valid neuron outputs the information, the shared register needs to be cleared to wait for the next or next collaborative group to work properly.
  • the input circuit circuits of the cooperative neurons in the cooperative group and the last effective neurons may be the same, that is, the same as the effective neurons, and the synergistic neurons are also
  • the input circuit has the function of reading the current neuron information, and the current neuron input information of each cooperative neuron is set to 0 by using the software design method.
  • the collaborative output information output module 400' is configured to output the coordinated output information by the valid neurons.
  • the neurons in the back end are equivalent to all the neurons in the cooperative group, and the multiple inputs correspond to one effective output.
  • the weight information of the multiple inputs can be fully utilized, and the existing neuron input weights are broken.
  • the shortcomings of the limited type improve the information processing capability of the neural network.
  • each of the cooperative groups After receiving the weight of the connection between the front-end neuron and the current neuron in the received front-end neuron information, reading the connection weight of the front-end neuron and the current neuron, and calculating the horizontal accumulation intermediate information, each of the cooperative groups
  • the weight information of the cooperative neurons is fully utilized, and in the cooperative output information of the effective neuron output, the weight information of each coordinated neuron is embodied, which is equivalent to expanding the weight information of the effective neurons, thereby improving The information processing capability of the neural network.
  • Determining a preset number of consecutive artificial neurons as an artificial cooperative group, or determining a preset number of consecutive pulse neurons as a pulse synergistic group, and determining a cooperative group in an artificial neural network or a pulsed neural network The expansion of input weights of individual neurons improves the information processing capability of artificial neural networks or pulsed neural networks.
  • FIG. 8 is a schematic structural diagram of a neuron information processing system for input weight expansion according to another embodiment.
  • the input weight-expanded neuron information processing system shown in FIG. 4 includes:
  • the collaborative group determining module 100' is configured to determine a preset number of consecutive neurons as a cooperative group, and determine a last neuron in the collaborative group as a valid neuron, and save the effective neural group in the cooperative group Neurons outside the element are identified as synergistic neurons.
  • the horizontally accumulated information obtaining module 200 ′ is configured to: acquire, by the first coordinated neuron in the collaborative group, the horizontal accumulated intermediate information of the first coordinated neuron according to the received front-end neuron information; Each of the cooperative neurons sequentially acquires the horizontal accumulation intermediate information of the cooperative neurons according to the received front-end neuron information and the lateral accumulation intermediate information of the front-end cooperative neurons, and the last synergistic nerve in the synergistic group The horizontal accumulation intermediate information of the element is determined as the horizontal accumulation information.
  • the collaborative output information obtaining module 300' is configured to obtain the collaborative output information according to the received front-end neuron information, the read current neuron information of the effective neuron, and the horizontal accumulated information.
  • the historical membrane potential update module 500' is configured to update the historical membrane potential information of the pulsed effective neurons.
  • the pulse effective neurons output the coordinated output information
  • the historical membrane potential information of the pulsed effective neurons is updated, so that the entire collaborative group completes subsequent information processing, and the pulse synergy
  • the neurons do not update the historical membrane potential information, and in the subsequent information processing, the function of weight expansion is completed, and the information processing capability of the entire pulse neural network is improved through the pulse synergy group.
  • an embodiment of the present invention further provides a computer device including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer
  • a computer device including a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor executes the computer
  • 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.
  • 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.
  • SRAM static RAM
  • DRAM dynamic RAM
  • SDRAM synchronous DRAM
  • DDRSDRAM double data rate SDRAM
  • ESDRAM enhanced SDRAM
  • Synchlink DRAM SLDRAM
  • Memory Bus Radbus
  • RDRAM Direct RAM
  • DRAM Direct Memory Bus Dynamic RAM
  • RDRAM Memory Bus Dynamic RAM

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Biomedical Technology (AREA)
  • Biophysics (AREA)
  • General Health & Medical Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Evolutionary Computation (AREA)
  • Computational Linguistics (AREA)
  • Molecular Biology (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • General Physics & Mathematics (AREA)
  • Mathematical Physics (AREA)
  • Software Systems (AREA)
  • Artificial Intelligence (AREA)
  • Neurology (AREA)
  • Management, Administration, Business Operations System, And Electronic Commerce (AREA)

Abstract

La présente invention concerne un procédé de traitement d'informations de poids de neurone, un procédé et un système de traitement d'informations neuronales pour extension de poids d'entrée, et un dispositif informatique. Le procédé de traitement d'informations de poids de neurone consiste à : recevoir des informations de sortie de neurone au sujet d'un neurone d'extrémité et un indice de poids correspondant aux informations de sortie de neurone (S100) ; lire, selon l'indice de poids, une corrélation d'informations d'indice de poids pour acquérir des informations de poids, la corrélation d'informations d'indice de poids étant une corrélation entre l'indice de poids et les informations de poids (S200) ; et acquérir des informations d'entrée au sujet du neurone d'extrémité selon les informations de poids et les informations de sortie de neurone (S300). Des informations de poids ne sont plus transmises directement entre des neurones et un indice des informations de poids est transmis, de sorte que non seulement la quantité d'informations transmises entre des réseaux est réduite, mais la définition des informations de poids peut être également changée de manière flexible.
PCT/CN2017/114659 2017-01-20 2017-12-05 Procédé et système de traitement d'informations de poids de neurone, procédé et système de traitement d'informations neuronales, et dispositif informatique WO2018133567A1 (fr)

Applications Claiming Priority (4)

Application Number Priority Date Filing Date Title
CN201710042090.4A CN106815638B (zh) 2017-01-20 2017-01-20 输入权重拓展的神经元信息处理方法和系统
CN201710042087.2 2017-01-20
CN201710042090.4 2017-01-20
CN201710042087.2A CN106875010B (zh) 2017-01-20 2017-01-20 神经元权重信息处理方法和系统

Publications (1)

Publication Number Publication Date
WO2018133567A1 true WO2018133567A1 (fr) 2018-07-26

Family

ID=62907871

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/CN2017/114659 WO2018133567A1 (fr) 2017-01-20 2017-12-05 Procédé et système de traitement d'informations de poids de neurone, procédé et système de traitement d'informations neuronales, et dispositif informatique

Country Status (1)

Country Link
WO (1) WO2018133567A1 (fr)

Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN104036451A (zh) * 2014-06-20 2014-09-10 深圳市腾讯计算机系统有限公司 基于多图形处理器的模型并行处理方法及装置
CN106022468A (zh) * 2016-05-17 2016-10-12 成都启英泰伦科技有限公司 人工神经网络处理器集成电路及该集成电路的设计方法
CN106104406A (zh) * 2014-03-06 2016-11-09 前进公司 神经网络及神经网络训练的方法
CN106815638A (zh) * 2017-01-20 2017-06-09 清华大学 输入权重拓展的神经元信息处理方法和系统
CN106875010A (zh) * 2017-01-20 2017-06-20 清华大学 神经元权重信息处理方法和系统

Patent Citations (5)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN106104406A (zh) * 2014-03-06 2016-11-09 前进公司 神经网络及神经网络训练的方法
CN104036451A (zh) * 2014-06-20 2014-09-10 深圳市腾讯计算机系统有限公司 基于多图形处理器的模型并行处理方法及装置
CN106022468A (zh) * 2016-05-17 2016-10-12 成都启英泰伦科技有限公司 人工神经网络处理器集成电路及该集成电路的设计方法
CN106815638A (zh) * 2017-01-20 2017-06-09 清华大学 输入权重拓展的神经元信息处理方法和系统
CN106875010A (zh) * 2017-01-20 2017-06-20 清华大学 神经元权重信息处理方法和系统

Similar Documents

Publication Publication Date Title
US11942191B2 (en) Compound property prediction method and apparatus, computer device, and readable storage medium
WO2018133568A1 (fr) Procédé et système de traitement d'informations neuronales en mode composé, et dispositif informatique
CN110232095B (zh) 一种数据同步方法、装置、存储介质和服务器
US20110320767A1 (en) Parallelization of Online Learning Algorithms
CN110457431A (zh) 基于知识图谱的问答方法、装置、计算机设备和存储介质
CN111914936B (zh) 语料数据的数据特征增强方法、装置及计算机设备
CN109583594B (zh) 深度学习训练方法、装置、设备及可读存储介质
CN106845632B (zh) 脉冲神经网络信息转换为人工神经网络信息的方法和系统
CN110442574B (zh) 一种数据处理方法、电子设备和计算机可读存储介质
CN109408590B (zh) 分布式数据库的扩容方法、装置、设备及存储介质
JP2019145111A (ja) データに基づく関数モデルを計算するためのモデル計算ユニット、制御装置、及び方法
WO2018133570A1 (fr) Procédé de traitement d'informations neuronales à seuil auto-adaptatif, procédé et système de traitement d'informations neuronales à valeur de fuite auto-adaptative, et dispositif informatique et support d'informations lisible
CN106875010B (zh) 神经元权重信息处理方法和系统
CN106815638B (zh) 输入权重拓展的神经元信息处理方法和系统
WO2018133569A1 (fr) Procédé et système de traitement d'informations neuronales ayant un fenêtrage temporel profond
WO2018133567A1 (fr) Procédé et système de traitement d'informations de poids de neurone, procédé et système de traitement d'informations neuronales, et dispositif informatique
CN117785490B (zh) 一种图神经网络模型的训练架构、方法、系统及服务器
CN103078946A (zh) 一种海量实时数据分布方法及其访问方法
US20190155330A1 (en) Information processing system
CN111666340A (zh) 基于大数据的同步数据校对方法、装置和计算机设备
CN105391590A (zh) 一种自动获取numa架构下系统路由表的方法及系统
CN116009889A (zh) 深度学习模型的部署方法、装置、电子设备及存储介质
CN116226553A (zh) 轨迹查询方法和装置、存储介质
CN114564878A (zh) 一种基于图神经网络的结构动力学模型生成方法
CN106790620B (zh) 一种分布式大数据处理方法

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 17892935

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 17892935

Country of ref document: EP

Kind code of ref document: A1

点击 这是indexloc提供的php浏览器服务，不要输入任何密码和下载