CN111667064A - Hybrid neural network based on photoelectric computing unit and operation method thereof - Google Patents
Hybrid neural network based on photoelectric computing unit and operation method thereof Download PDFInfo
- Publication number
- CN111667064A CN111667064A CN202010322172.6A CN202010322172A CN111667064A CN 111667064 A CN111667064 A CN 111667064A CN 202010322172 A CN202010322172 A CN 202010322172A CN 111667064 A CN111667064 A CN 111667064A
- Authority
- CN
- China
- Prior art keywords
- neural network
- computing unit
- photoelectric
- selector
- computing
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Granted
Links
- 238000013528 artificial neural network Methods 0.000 title claims abstract description 47
- 238000000034 method Methods 0.000 title claims abstract description 11
- 238000004364 calculation method Methods 0.000 claims abstract description 49
- 210000002569 neuron Anatomy 0.000 claims abstract description 25
- 210000001787 dendrite Anatomy 0.000 claims abstract description 17
- 210000003050 axon Anatomy 0.000 claims abstract description 14
- 230000000946 synaptic effect Effects 0.000 claims abstract description 11
- 230000008878 coupling Effects 0.000 claims abstract description 5
- 238000010168 coupling process Methods 0.000 claims abstract description 5
- 238000005859 coupling reaction Methods 0.000 claims abstract description 5
- 230000005284 excitation Effects 0.000 claims description 14
- 230000004913 activation Effects 0.000 claims description 11
- 238000006243 chemical reaction Methods 0.000 claims description 10
- 230000001537 neural effect Effects 0.000 claims description 9
- 238000001208 nuclear magnetic resonance pulse sequence Methods 0.000 claims description 5
- 238000012216 screening Methods 0.000 claims description 3
- 230000005693 optoelectronics Effects 0.000 claims description 2
- 230000000737 periodic effect Effects 0.000 claims description 2
- 210000000225 synapse Anatomy 0.000 claims description 2
- 230000001360 synchronised effect Effects 0.000 claims description 2
- 230000000694 effects Effects 0.000 abstract 1
- 210000005036 nerve Anatomy 0.000 abstract 1
- 238000004422 calculation algorithm Methods 0.000 description 31
- 230000006870 function Effects 0.000 description 14
- 238000013507 mapping Methods 0.000 description 7
- 238000013527 convolutional neural network Methods 0.000 description 6
- 238000012545 processing Methods 0.000 description 6
- 230000003287 optical effect Effects 0.000 description 5
- 230000003213 activating effect Effects 0.000 description 4
- 239000000969 carrier Substances 0.000 description 4
- 238000010586 diagram Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 210000004556 brain Anatomy 0.000 description 3
- 125000004122 cyclic group Chemical group 0.000 description 3
- 238000012549 training Methods 0.000 description 3
- 230000005540 biological transmission Effects 0.000 description 2
- 238000001514 detection method Methods 0.000 description 2
- 238000005516 engineering process Methods 0.000 description 2
- 238000004088 simulation Methods 0.000 description 2
- 238000012360 testing method Methods 0.000 description 2
- 238000009825 accumulation Methods 0.000 description 1
- 210000003710 cerebral cortex Anatomy 0.000 description 1
- 238000005265 energy consumption Methods 0.000 description 1
- 230000003631 expected effect Effects 0.000 description 1
- 230000010354 integration Effects 0.000 description 1
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
- G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
- G06N3/067—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means
- G06N3/0675—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using optical means using electro-optical, acousto-optical or opto-electronic means
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02D—CLIMATE CHANGE MITIGATION TECHNOLOGIES IN INFORMATION AND COMMUNICATION TECHNOLOGIES [ICT], I.E. INFORMATION AND COMMUNICATION TECHNOLOGIES AIMING AT THE REDUCTION OF THEIR OWN ENERGY USE
- Y02D10/00—Energy efficient computing, e.g. low power processors, power management or thermal management
Abstract
The invention discloses a hybrid neural network based on a photoelectric computing unit and an operation method thereof. The nerve computing core structure of the hybrid neural network comprises a synaptic network, axons, a data receiver, a selector, dendrites and neurons, wherein an array formed by periodically arranging a plurality of photoelectric computing units is used as the synaptic network, each photoelectric computing unit comprises a light-emitting unit and a computing unit, and light emitted by the light-emitting unit is incident into the computing unit; each calculation unit comprises a carrier control area, a coupling area, a photon-generated carrier collecting area and a readout area; the carrier control regions of each row of computing units in the array are sequentially connected to serve as axons, and the carrier control regions in the axons are connected with the data receiver; the read-out areas of each column of the calculation units in the array are sequentially connected to serve as dendrites, and the output ends of the read-out areas of the dendrites are connected with the neurons. The neural network of the invention does not need to repeatedly access the off-chip memory in actual operation, thereby achieving the effect of reducing power consumption.
Description
Technical Field
The invention relates to a hybrid neural network based on a photoelectric computing unit and an operation method thereof, belonging to the field of computing and photoelectric detection.
Background
Most of the conventional computers adopt the von neumann architecture, however, because the von neumann architecture is separated from the memory unit, the great energy consumption is generated on the data transmission, and the operation speed is influenced. The photoelectric computing unit is a computing device which can carry out independent operation or operation combined with the current electronic computing technology, can realize the integrated function of high-precision storage and operation, can store the optical signal of the optical input end by a single device and store the optical signal for a long time after the optical input end is cut off, and can realize that the multiplication operation can be finished by the single device. These characteristics make the photoelectric computation unit suitable for not only the traditional convolutional neural network but also the impulse neural network similar to the brain.
The existing neural network technology mostly uses a convolution or circular deep neural network, and the operation is completed through a general computing platform, such as a central processing unit, a graphic processing unit and the like. The convolution or circulation deep neural network has large weight parameter and high calculation complexity, needs to visit the off-chip cache read weight in large quantity to complete multiple operations, and has higher requirement on the calculation power of a traditional calculation system. And due to repeated access to the off-chip cache, a large amount of energy is lost in the data transmission process, so that the computing energy efficiency is low, and the increasing application requirements cannot be met.
Disclosure of Invention
In order to improve the calculation performance of the system, the invention provides a hybrid neural network based on a photoelectric calculation unit and an operation method of the hybrid neural network.
The neural network adopts the following technical scheme:
the hybrid neural network based on the photoelectric computing units comprises a neural computing core structure, wherein the neural computing core structure comprises a synaptic network, axons, a data receiver, a selector, dendrites and neurons, an array formed by periodically arranging a plurality of photoelectric computing units is used as the synaptic network, each photoelectric computing unit comprises a light-emitting unit and a computing unit, and light emitted by the light-emitting unit is incident into the computing unit; each calculation unit comprises a carrier control area, a coupling area, a photon-generated carrier collecting area and a readout area; the carrier control regions of each row of computing units in the array are sequentially connected to serve as axons, the carrier control region of each row is connected with a first selector, and all the first selectors are connected with a data receiver; the carrier reading regions of each column of computing units in the array are sequentially connected to serve as dendrites, and each column of carrier reading regions is connected with one neuron; the single neuron comprises an analog-to-digital converter, an adder, a second selector, an activation function and a threshold comparator; the input end of the analog-to-digital converter is connected with the output end of the current carrier reading area of the computing unit, and the output end of the analog-to-digital converter is connected with the input end of the adder; the output end of the adder is respectively connected with the input ends of the activation function comparator and the threshold comparator; the output end of the activation function and threshold comparator is connected with the second selector to output the result by the second selector classification.
Furthermore, the neural network also comprises a third selector, the input ends of the third selector are respectively connected with the activation function and the threshold comparator, and the output end of the third selector is connected with the adder.
The invention relates to a photoelectric calculation unit-based hybrid neural network operation method, which comprises the following specific steps of: the light-emitting unit is driven by the driver to input and store the weight value of the neural computation core structure into the synaptic network; the data receiver receives and caches multi-bit value accurate information or binary pulse sequences from the outside, after receiving a synchronous control signal of current information, the cached current information is transmitted to the first selector of each row through an axon to carry out data type screening, and the screened current information is fed into a carrier control area of a corresponding computing unit; multiplying current information received by the carrier control area and a weight value stored in a synaptic network, wherein the calculation result of each calculation unit is connected through dendrites, so that the current of the output end of the reading area of each row of calculation units is converged and output to a neuron; the neuron performs analog-to-digital conversion on the converged current signals to form digital signals, continuously accumulates the digital signal results generated by input pulse excitation and after current conversion is calculated, and then biases and activates the accumulated results, or brings the accumulated results into a model function of a threshold comparator to judge whether output pulse signals are generated or not.
The invention fully utilizes the characteristic of the storage and calculation integration of the photoelectric calculation unit, inputs weight parameters in the mixed neural network into the array through light to be stored so as to simulate the connection weight in the human brain, completes multiplication operation by utilizing the characteristic of a single device, and controls and executes an explicit algorithm or simulates cerebral cortex by using a multi-path selector, thereby forming the neural network which is mixed with computer science and neuroscience. Because the photoelectric storage and calculation integrated device integrates storage and calculation, the neural network of the invention does not need to repeatedly access an off-chip memory in actual operation, can greatly reduce power consumption and improve calculation efficiency, and has great advantages compared with the existing deep neural network.
Drawings
FIG. 1 is a block diagram of a multi-function region of a computing unit.
Fig. 2 is a schematic diagram of a hybrid neural network computational core structure based on an optoelectronic computational unit.
Fig. 3 is a schematic diagram of an application system of a hybrid neural network based on a photoelectric computing unit.
Fig. 4 is a schematic application flow diagram of a hybrid neural network based on a photoelectric computing unit.
Detailed Description
The photoelectric calculating unit comprises a light-emitting unit and a calculating unit, wherein light emitted by the light-emitting unit is incident into the calculating unit. Each computing unit is a multi-functional area structure including three functional areas, as shown in fig. 1, the three functional areas are: a carrier control region, a coupling region, a photogenerated carrier collection region, and a readout region. The carrier control region is used for controlling and modulating carriers in the photogenerated carrier collecting region and the readout region; the collecting region in the photogenerated carrier collecting region and the reading region is used for absorbing the photons emitted by the light emitting unit and collecting the generated photogenerated carriers; the reading area in the carrier control area or the photogenerated carrier collecting area and the reading area is connected with an electric signal, and the reading area is used for outputting carriers acted by the photogenerated carriers and the electric signal; a coupling region connects the collection region and the readout region. The specific structure of the calculation unit can be the structure of embodiments 1-3 of patent application CN 201910415827.1.
Example 1
In the embodiment, a plurality of light-emitting units and a calculation unit structure are used for forming an n × n photoelectric calculation array to form a neural calculation core structure, so that a hybrid neural network is formed, and algorithms capable of processing multi-bit value accurate information, such as a convolutional neural network and a cyclic neural network, and algorithms capable of processing a binary pulse sequence, such as a pulse neural network, are compatible.
As shown in FIG. 2, the square W in the figureijThe photoelectric calculation array simulates a neural synapse network in a human brain, receives and stores weight information transmitted by optical signals, and completes corresponding operation when receiving data signals after the storage is finished. An array formed by a plurality of photoelectric computing units in a periodic arrangement mode is used as a synaptic network; the carrier control regions of each row of computing units in the array are sequentially connected to serve as axons, the connected control regions are connected with one selector, all the selectors are connected with a data receiver, and the data receiver receives external current information. The carrier reading regions of each column of computing units in the array are sequentially connected to serve as dendrites, and the connected carrier reading regions are connected with one neuron; the input end of the analog-to-digital converter in the single neuron is connected with the output end of the carrier reading area of the computing unit, and the output end of the analog-to-digital converter in the single neuron is connected with one input end of the adder; the output end of the adder is respectively connected with the input ends of the activation function comparator and the threshold comparator; the output end of the activating function or threshold comparator is connected with the selector of the result output part; and after the neuron discriminates the corresponding algorithm, the result is classified and output through a selector of the result output part. The input end of the selector between the adder and the activating function and threshold comparator is connected with the activating function bias information and the threshold comparator, the output end of the selector is connected with the adder, and the neuron can select corresponding data to perform accumulation operation according to the type of the current network algorithm.
The axon receives data excitation signals in the neural network through a data receiver in the neural computational core. The buffered current information is passed through the axon to the selector of each row for data type screening. The dendrites deliver the converged current to the neurons. The neuron carries out analog-to-digital conversion on the gathered current signals to form digital signals, continuously accumulates the digital signal results generated by input pulse excitation and after current conversion calculation, if the digital signal results are multi-bit value accurate information of a convolutional neural network, a cyclic neural network and the like, the accumulated results are subjected to operations of adding bias, activating a function and the like, and if the digital signal results are binary pulse sequences of the pulse neural network and the like, the accumulated results are brought into a threshold comparator model function formed by digital logic to judge whether output pulse signals are generated or not.
If the buffer current information screened by the selector is multi-bit value accurate information of a convolutional neural network, a cyclic neural network and the like, the model of the neuron can be described by the following formula:
in the formula Ii(t) represents the multi-valued information received by the data receiver in the network at a certain time, and the current signal is transmitted to all the contact points (i.e. photoelectric calculation units) of each row in fig. 2 through the axon. WijAnd representing the weight value stored by each photoelectric calculation unit in the n x n array. If the point has a weight value and is not 0, the point is represented to have a connection relation in the network, and the point generates current according to the corresponding weight value after receiving the pulse and transmits the current through the dendrite. The summation between i-1-n in the formula corresponds to dendrites in fig. 2, and the currents of all the protruding points in each column are superposed through dendrite convergence. bjIndicating the applied bias, relative to the network model. x is the number ofjIntermediate results before the activation function ReLU model are fed in are stored by the digital logic in the neuron. In the activation function ReLU model f (x), if the multiply-and-add multi-bit value precision data is positive, this information is retainedAnd outputting the data; and if the multi-bit value accurate data after multiplication and addition is negative, resetting.
If the buffer current information screened by the selector is a binary pulse sequence of a pulse neural network or the like, the model of the neuron can be described by the following formula:
in the formula Ii(t) indicates that at some point in the network the pulse receiver receives a binary pulse, if a pulse excitation is received, Ii(t) is then 1, and the pulse signal is transmitted through the axon to all the contact points (i.e., the photoelectric computing unit) of each row in fig. 2. WijAnd representing the weight value stored by each photoelectric calculation unit in the n x n array. If the point has a weight value and is not 0, the point is represented to have a connection relation in the network, and the point generates current according to the corresponding weight value after receiving the pulse and transmits the current through the dendrite. The summation between i-1-n in the formula corresponds to dendrites in fig. 2, and the currents of all the protruding points in each column are superposed through dendrite convergence. VjAnd (t-1) storing an intermediate result stored at the previous moment through a digital logic part in the neuron and adding the intermediate result to a calculation result at the current moment. Lambda [ alpha ]jA threshold value representing each neuron is associated with the network model. Comparing the calculation result of the current moment in the neuron with a threshold value, if the calculation result exceeds the threshold value, sending a pulse signal out and clearing the current value; and if the calculation result does not exceed the threshold value, continuing to maintain and receiving the calculation intermediate result at the next moment.
Example 2
The embodiment provides an application system and a process based on the hybrid neural network.
As shown in fig. 3, the application system includes six parts, namely an algorithm selection module, an algorithm mapping module, a parameter writing module, an excitation generation module, a network calculation module and a result conversion module. The algorithm selection module selects and uses one or more of a CNN algorithm and a SNN algorithm according to the actual requirements of users; the algorithm mapping module decomposes the corresponding algorithm according to the user requirements, so that the calculation process of the algorithm can be mapped to the corresponding hybrid neural network; the parameter writing module writes parameter information such as weight, bias and the like required by the corresponding algorithm into the photoelectric calculation array or the register according to the mapping relation according to the algorithm split by the algorithm mapping module; the excitation generating module generates corresponding data excitation according to excitation data in the actual operation of the algorithm; the network computing module is used for processing the parameter and the operation process of excitation in the hybrid neural network; the result conversion module is used for collecting the calculation result transmitted by the hybrid neural network, processing the calculation result according to the corresponding algorithm model and converting the calculation result into an explicit result of the corresponding algorithm.
As shown in fig. 4, the application flow of the hybrid neural network based on the photoelectric computing unit is as follows:
firstly, determining an application scene and selecting an algorithm. The hybrid neural network based on the photoelectric computing unit is suitable for various application scenes, such as face detection, image classification, voice recognition, denoising coding and the like; after the application category and the use scene are determined, the user selects a proper algorithm and a training test data set according to the self requirement. For example, for applications such as face recognition and image classification, a CNN algorithm can be selected; for example, for speech recognition, denoising coding and other applications, SNN algorithms can be selected;
and secondly, network training is carried out. A user builds a network structure according to the selected algorithm type and trains a network model according to the training test set to achieve the expected effect;
and thirdly, carrying out algorithm mapping and simulation. The trained algorithm is mapped into a hybrid neural network based on a photoelectric calculation array through an algorithm mapping module, the operation flow of the algorithm is converted into an information form supported by the hybrid neural network based on the photoelectric calculation array, and the mapped network is subjected to simulation, so that the mapped network can normally complete work;
and fourthly, realizing the whole hardware architecture. According to the mapping relation of the algorithm, the main control part recompiles the hybrid neural network, the light-emitting unit inputs the weight value of the neural computation core structure into the synaptic network under the drive of the driver and stores the weight value, and meanwhile, the register stores parameter information such as bias according to the requirement of the corresponding algorithm; after the writing of parameters such as weight, bias and the like is finished, the neural network receives excitation data transmitted from the outside and converts an external excitation signal into an information form which can be processed by the hybrid neural network through the excitation generating module; then, the neural network completes the operation process of parameters such as weight, bias and the like and excitation in the network calculation module and feeds the calculation result into the result conversion module; and the result conversion module integrates the calculation result of the neural network according to the corresponding algorithm model, converts the calculation result into an explicit calculation result and displays the explicit calculation result in the application system.
Claims (3)
1. Hybrid neural network based on photoelectric computing units, comprising a neural computational core structure comprising a synaptic network, axons, data receivers, selectors, dendrites and neurons,
an array formed by a plurality of photoelectric computing units in a periodic arrangement mode is used as a synapse network, each photoelectric computing unit comprises a light-emitting unit and a computing unit, and light emitted by the light-emitting unit is incident into the computing unit; each calculation unit comprises a carrier control area, a coupling area, a photon-generated carrier collecting area and a readout area; the carrier control regions of each row of computing units in the array are sequentially connected to serve as axons, the carrier control region of each row is connected with a first selector, and all the first selectors are connected with a data receiver; the carrier reading regions of each column of computing units in the array are sequentially connected to serve as dendrites, and each column of carrier reading regions is connected with one neuron; the single neuron comprises an analog-to-digital converter, an adder, a second selector, an activation function and a threshold comparator; the input end of the analog-to-digital converter is connected with the output end of the current carrier reading area of the computing unit, and the output end of the analog-to-digital converter is connected with the input end of the adder; the output end of the adder is respectively connected with the input ends of the activation function comparator and the threshold comparator; the output end of the activation function and threshold value comparator is connected with a second selector, and the second selector classifies and outputs the result.
2. The hybrid neural network based on optoelectronic computing units of claim 1, further comprising a third selector having inputs respectively connected to the activation function and the threshold comparator and an output connected to the adder.
3. The method for operating the hybrid neural network based on the photoelectric computing unit according to claim 1, wherein the specific steps comprise:
the light-emitting unit is driven by the driver to input and store the weight value of the neural computation core structure into the synaptic network;
the data receiver receives and caches multi-bit value accurate information or binary pulse sequences from the outside, after receiving a synchronous control signal of current information, the cached current information is transmitted to the first selector of each row through an axon to carry out data type screening, and the screened current information is fed into a carrier control area of a corresponding computing unit;
multiplying current information received by the carrier control area and a weight value stored in a synaptic network, wherein the calculation result of each calculation unit is connected through dendrites, so that the current of the output end of the reading area of each row of calculation units is converged and output to a neuron;
the neuron performs analog-to-digital conversion on the converged current signals to form digital signals, continuously accumulates the digital signal results generated by input pulse excitation and after current conversion is calculated, and then biases and activates the accumulated results, or brings the accumulated results into a model function of a threshold comparator to judge whether output pulse signals are generated or not.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010322172.6A CN111667064B (en) | 2020-04-22 | 2020-04-22 | Hybrid neural network based on photoelectric computing unit and operation method thereof |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010322172.6A CN111667064B (en) | 2020-04-22 | 2020-04-22 | Hybrid neural network based on photoelectric computing unit and operation method thereof |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN111667064A true CN111667064A (en) | 2020-09-15 |
| CN111667064B CN111667064B (en) | 2023-10-13 |
Family
ID=72382719
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010322172.6A Active CN111667064B (en) | 2020-04-22 | 2020-04-22 | Hybrid neural network based on photoelectric computing unit and operation method thereof |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN111667064B (en) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113420788A (en) * | 2020-10-12 | 2021-09-21 | 黑芝麻智能科技(上海)有限公司 | Integer-based fusion convolution layer in convolutional neural network and fusion convolution method |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH08235144A (en) * | 1995-02-27 | 1996-09-13 | Nec Corp | Neural network device |
| CN105095967A (en) * | 2015-07-16 | 2015-11-25 | 清华大学 | Multi-mode neural morphological network core |
| CN109478557A (en) * | 2016-08-03 | 2019-03-15 | 株式会社半导体能源研究所 | Photographic device, photographing module, electronic equipment and camera system |
-
2020
- 2020-04-22 CN CN202010322172.6A patent/CN111667064B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JPH08235144A (en) * | 1995-02-27 | 1996-09-13 | Nec Corp | Neural network device |
| CN105095967A (en) * | 2015-07-16 | 2015-11-25 | 清华大学 | Multi-mode neural morphological network core |
| CN109478557A (en) * | 2016-08-03 | 2019-03-15 | 株式会社半导体能源研究所 | Photographic device, photographing module, electronic equipment and camera system |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113420788A (en) * | 2020-10-12 | 2021-09-21 | 黑芝麻智能科技(上海)有限公司 | Integer-based fusion convolution layer in convolutional neural network and fusion convolution method |
Also Published As
| Publication number | Publication date |
|---|---|
| CN111667064B (en) | 2023-10-13 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| US11308406B2 (en) | Method of operating neural networks, corresponding network, apparatus and computer program product | |
| WO2021115262A1 (en) | Pulse convolutional neural network algorithm, integrated circuit, computing apparatus, and storage medium | |
| CN111897579B (en) | Image data processing method, device, computer equipment and storage medium | |
| CN109447241B (en) | Dynamic reconfigurable convolutional neural network accelerator architecture for field of Internet of things | |
| CN111144329A (en) | Light-weight rapid crowd counting method based on multiple labels | |
| US20230334632A1 (en) | Image recognition method and device, and computer-readable storage medium | |
| CN111464881B (en) | Full-convolution video description generation method based on self-optimization mechanism | |
| KR20220058897A (en) | Perform XNOR equivalent operations by adjusting column thresholds of a compute-in-memory array | |
| CN106845632B (en) | Method and system for converting impulse neural network information into artificial neural network information | |
| CN111898733A (en) | Deep separable convolutional neural network accelerator architecture | |
| CN112084911B (en) | Human face feature point positioning method and system based on global attention | |
| Liu et al. | RB-Net: Training highly accurate and efficient binary neural networks with reshaped point-wise convolution and balanced activation | |
| CN111860773B (en) | Processing apparatus and method for information processing | |
| CN115423081A (en) | Neural network accelerator based on CNN _ LSTM algorithm of FPGA | |
| CN111667064B (en) | Hybrid neural network based on photoelectric computing unit and operation method thereof | |
| CN110222835A (en) | A kind of convolutional neural networks hardware system and operation method based on zero value detection | |
| CN110232441A (en) | A kind of stacking-type based on unidirectional systolic arrays is from encoding system and method | |
| CN113033769B (en) | Probabilistic calculation neural network method and asynchronous logic circuit | |
| CN114821218A (en) | Target detection model searching method based on improved channel attention mechanism | |
| CN110728303B (en) | Dynamic self-adaptive computing array based on convolutional neural network data complexity | |
| CN110263926B (en) | Pulse neural network based on photoelectric computing unit, system and operation method thereof | |
| CN115587622A (en) | Multilayer perceptron device based on integrative device of photoelectric memory | |
| CN115964258A (en) | Internet of things network card abnormal behavior grading monitoring method and system based on multi-time sequence analysis | |
| CN115587616A (en) | Network model training method and device, storage medium and computer equipment | |
| Lee et al. | MPQ-YOLACT: Mixed-Precision Quantization for Lightweight YOLACT |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant | ||
| TR01 | Transfer of patent right | ||
| TR01 | Transfer of patent right |
Effective date of registration: 20240409 Address after: Floor 13, Building 04, Accelerator, No. 18 Jialing Jiangdong Street, Nanjing, Jiangsu Province, 210000 Patentee after: NANJING UNIVERSITY 5D TECHNOLOGY Co.,Ltd. Country or region after: China Address before: 9/F, Building B, No. 100, Tianjiao Road, Qilin Hi-tech Industrial Development Zone, Jiangning District, Nanjing, Jiangsu, 210000 Patentee before: Nanjing Weixin Photoelectric System Co.,Ltd. Country or region before: China |