[go: up one dir, main page]

WO2020121023A1 - Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée - Google Patents

Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée Download PDF

Info

Publication number
WO2020121023A1
WO2020121023A1 PCT/IB2018/059878 IB2018059878W WO2020121023A1 WO 2020121023 A1 WO2020121023 A1 WO 2020121023A1 IB 2018059878 W IB2018059878 W IB 2018059878W WO 2020121023 A1 WO2020121023 A1 WO 2020121023A1
Authority
WO
WIPO (PCT)
Prior art keywords
input values
stream
arithmetic unit
neuron
input
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.)
Ceased
Application number
PCT/IB2018/059878
Other languages
English (en)
Inventor
Sebastien Delerse
Ludovic Larzul
Benoit Chappet De Vangel
Taoufik Chouta
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
Mipsology SAS
Original Assignee
Mipsology SAS
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
Application filed by Mipsology SAS filed Critical Mipsology SAS
Priority to EP18827254.6A priority Critical patent/EP3895071A1/fr
Priority to PCT/IB2018/059878 priority patent/WO2020121023A1/fr
Priority to PCT/IB2019/060631 priority patent/WO2020121203A1/fr
Priority to PCT/IB2019/060630 priority patent/WO2020121202A1/fr
Priority to EP19835488.8A priority patent/EP3895072A1/fr
Priority to EP19835489.6A priority patent/EP3895073A1/fr
Priority to US16/709,507 priority patent/US11126912B2/en
Publication of WO2020121023A1 publication Critical patent/WO2020121023A1/fr
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Classifications

    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; 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 OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0464—Convolutional networks [CNN, ConvNet]
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/04—Architecture, e.g. interconnection topology
    • G06N3/0495—Quantised networks; Sparse networks; Compressed networks
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/06—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • G—PHYSICS
    • G06—COMPUTING OR CALCULATING; COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/09—Supervised learning

Definitions

  • the present disclosure relates generally to data processing and, more particularly, to system and method for accelerating artificial neural network
  • ANNs Artificial Neural Networks
  • the human brain contains 10-20 billion neurons connected through synapses. Electrical and chemical messages are passed from neurons to neurons based on input information and their resistance to passing information.
  • a neuron can be represented by a node performing a simple operation of addition coupled with a saturation function.
  • a synapse can be represented by a connection between two nodes. Each of the connections can be associated with an operation of multiplication by a constant.
  • the ANNs are particularly useful for solving problems that cannot be easily solved by classical computer programs.
  • ANNs While forms of the ANNs may vary, they all have the same basic elements similar to the human brain.
  • a typical ANN can be organized into layers, each of the layers may include many neurons sharing similar functionality. The inputs of a layer may come from a previous layer, multiple previous layers, any other layers or even the layer itself.
  • Major architectures of ANNs include Convolutional Neural Network (CNN), Recurrent Neural Network (RNN) and Long Term Short Memory (LTSM) network, but other architectures of ANN can be developed for specific applications. While some operations have a natural sequence, for example a layer depending on previous layers, most of the operations can be carried out in parallel within the same layer. The ANNs can then be computed in parallel on many different computing elements similar to neurons of the brain.
  • a single ANN may have hundreds of layers. Each of the layers can involve millions of connections. Thus, a single ANN may potentially require billions of simple operations like multiplications and additions.
  • ANNs can result in a very heavy load for processing units (e.g., CPU), even ones running at high rates.
  • processing units e.g., CPU
  • GPUs graphics processing units
  • GPUs can be used to process large ANNs because GPUs have a much higher throughput capacity of operations in comparison to CPUs. Because this approach solves, at least partially, the throughput limitation problem, GPUs appear to be more efficient in the computations of ANNs than the CPUs.
  • GPUs are not well suited to the computations of ANNs because the GPUs have been specifically designed to compute graphical images.
  • the GPUs may provide a certain level of parallelism in computations.
  • GPUs are constraining the computations in long pipes implying latency and lack of reactivity.
  • very large GPUs can be used which may involving excessive power consumption, a typical issue of GPUs.
  • CPUs provide a very generic engine that can execute very few sequences of instructions with a minimum effort in terms of programming, but lack the power of computing for ANN.
  • GPUs are slightly more parallel and require a larger effort of programming than CPUs, which can be hidden behind libraries with some performance costs, but are not very well suitable for ANNs.
  • FPGAs Field Programmable Gate Arrays
  • the FPGAs can be configured to perform computations in parallel. Therefore, FPGAs can be well suited to compute ANNs.
  • One of the challenges of FPGAs is the programming, which requires a much larger effort than programming CPUs and GPUs. Adaption of FPGAs to perform ANN computations can be more challenging than for CPUs and GPUs.
  • a system for accelerating ANN computation may include a controller, a selector
  • the controller can be configured to dynamically control the selection, based on a criterion, of an input value from a stream of input values of a neuron.
  • the controller can configure the selector to provide dynamically the selected input value to the arithmetic unit.
  • the controller can provide, to the arithmetic unit, information of the selected input value.
  • the information may include an offset of the input value in the stream. The offset cannot be statically computed before the computation of the neuron since the input values in the stream can be generated as a result of previous neuron computation.
  • the arithmetic unit can be configured to acquire, based on the offset, a weight from a set of weights.
  • the arithmetic unit can further perform a mathematical operation on the selected input value and the weight to obtain a result to be used in computing an output of the neuron.
  • the mathematical operation performed by the arithmetic unit in its simplest form may include a multiplication product.
  • the criterion applied by the controller may include comparison of the input value to a reference value.
  • the comparison may include an equality comparison to zero.
  • the comparison may also include a more-than comparison to a threshold.
  • the selector may include a multiplexer.
  • the selector may also include an information related to a memory address.
  • Input values not selected from the subset of input values are not provided to the arithmetic unit, therefore, a count of mathematical operations performed by the arithmetic unit on selected input values from the subset of input values is less than a count of mathematical operations required to be performed by the arithmetic unit on all input values from the subset of the input values.
  • a system for realigning streams of neuron outputs may include an arithmetic unit, at least one further arithmetic unit, and a synchronization module communicatively coupled to the arithmetic unit and the further arithmetic unit.
  • the arithmetic unit can generate a stream of neuron outputs and the further arithmetic unit can generate a further stream of further neuron outputs.
  • the synchronization module can receive the stream of neuron outputs from the arithmetic unit and the further stream of further neuron outputs from the further arithmetic unit.
  • the stream of the neuron outputs and the stream of further neuron outputs can be unaligned with respect to each other.
  • the synchronization module can dynamically reorder the neuron outputs and the further neuron outputs to obtain an ordered sequence.
  • the synchronization module can write the ordered sequence to a local or global memory storage.
  • the ordered sequence can be used as further input values for further computations required by the ANN.
  • the neuron outputs can be computed by the arithmetic unit based on first input values.
  • the first input values can be selected from a stream of input values.
  • the further neuron outputs can be computed by the further arithmetic unit based on second input values.
  • the second input values can be selected from the further stream of further input values.
  • a count of the input values in stream can be equal to a count of further input values in the further stream.
  • First indexes of the first input values in the stream of input values can be different from second indexes of the second input values in the further stream of further input values.
  • a method for accelerating ANN computation may further include dynamically selecting, by a controller communicatively coupled to a selector and an arithmetic unit and based on a criterion, an input value from a stream of input values of a neuron of ANN.
  • the method may include dynamically configuring, by the controller, the selector to provide the selected input value to the arithmetic unit.
  • the method may include dynamically providing, by the controller to the arithmetic unit, an offset of the selected input value.
  • the method may include dynamically acquiring, by the arithmetic unit and based on the offset, a weight from a set of weights.
  • the method may further include performing, by the arithmetic unit a mathematical operation on the selected input value and the weight to obtain a result, wherein the result is to be used to compute an output of the neuron.
  • FIG. 1 is a block diagram showing an example system wherein a method for acceleration of ANN computations can be implemented, according to some example embodiments.
  • FIG. 2 shows an ANN, neuron, and transfer function, according to an example embodiment.
  • FIG. 3 is a flow chart showing training and inference of ANN, according to some example embodiments.
  • FIG. 4 is a block diagram showing a system for acceleration of ANN computations, according to some example embodiments.
  • FIG. 5 is a block diagram showing a system for selecting input values for processing by ANN computations, according to an example embodiment.
  • FIG. 6 is a block diagram showing a system for selecting input data for processing by ANN computations, according to another example embodiment.
  • FIG. 7 is a block diagram showing a system for synchronization of outputs of neurons in ANN computations, according to another example embodiment.
  • FIG. 8 is a flow chart showing steps of a method for acceleration of ANN computations, according to some example embodiments.
  • FIG. 9 shows a computing system that can be used to implement
  • Embodiments of this disclosure are concerned with methods and systems for acceleration of ANN computations.
  • Embodiments of present disclosure may facilitate selection of input values for processing by neurons of an ANN in order to avoid unnecessary mathematical operations in computation of outputs of neurons and, thereby, accelerating of computations of the ANN.
  • the input values equal to zero are not processed by arithmetic units configured to compute the neurons of ANN.
  • the selection of input values can be also based on another criterion.
  • the sequence of operations performed by the arithmetic units may depend dynamically on the stream of input values and the criterion used for selection. For example, a select-all (no selection) criterion would result in an identical sequence of operations for all inputs, wherein essentially all operations of the neurons (and ANN) are performed. In this case the computation of ANN will be similar to ANN
  • module shall be construed to mean a hardware device, software, or a combination of both.
  • a hardware-based module can use one or more microprocessors, FPGAs, application-specific integrated circuits (ASICs),
  • module can constitute computer programs, computer program procedures, computer program functions, and the like.
  • a module of a system can be implemented by a computer or server, or by multiple computers or servers interconnected into a network.
  • module may also refer to a subpart of a computer system, a hardware device, an integrated circuit, or a computer program.
  • Technical effects of certain embodiments of the present disclosure can include configuring integrated circuits, FPGAs, or computer systems to perform ANN computations without execution of redundant and unnecessary mathematical operations, thereby accelerating the ANN computations. Further technical effects of some embodiments of the present disclosure can facilitate configuration of integrated circuits, FPGAs, or computer systems to dynamically qualify data on which
  • FIG. 1 is a block diagram showing an example system 100, wherein a method for accelerating ANN computation can be implemented, according to some example embodiments.
  • the system 100 can be part of a computing system, such as a personal computer, a server, a cloud-based computing recourse, and the like.
  • the system 100 may include one or more FPGA boards 105 and a chipset 135 including a least one CPU.
  • the chipset 135 can be communicatively connected to the FPGA boards 105 via a communication interface.
  • the communication interface may include a Peripheral Component Interconnect Express (PCIE) standard 130.
  • PCIE Peripheral Component Interconnect Express
  • the communication interface may also include an Ethernet connection 131.
  • the FPGA board 105 may include an FPGA 115, a volatile memory 110, and a non-volatile memory 120.
  • the volatile memory 110 may include a double data rate synchronous dynamic random-access memory (DDR SDRAM), High Bandwidth Memory (HBM), or any other type of memory.
  • the volatile memory 110 may include the host memory.
  • the non-volatile memory 120 may include Electrically Erasable Programmable Read-Only Memory (EEROM), a solid-state drive (SSD), a flash memory, and so forth.
  • EEROM Electrically Erasable Programmable Read-Only Memory
  • SSD solid-state drive
  • flash memory and so forth.
  • the FPGA 115 can include blocks.
  • the blocks may include a set of
  • the blocks may further include registers retaining bit
  • Programming of FPGA 115 may include configuring each of the blocks to have an expected behavior and connecting the blocks by routing information between the blocks. Programming of FPGA 115 can be carried out using a result from a compiler taking as input schematic description, gate-level description, hardware languages like Verilog, System Verilog, or Very High Speed Integrated Circuit
  • VHDF Hardware Description Fanguage
  • the non-volatile memory 120 may be configured to store instructions in a form of bit file 125 to be executed by the FPGA 115.
  • the FPGA 115 can be configured by the instructions to perform one or more floating point operations including multiplication and addition to calculate sum of products that can be used in neural network computations.
  • the volatile memory 110 may be configured to store weights W[i] for neurons of one or more ANNs, input values V[i] to be processed for the ANNs, and results of ANNs computation including any intermediate results of computations of layers of the ANNs.
  • FIG. 2 shows ANN 210, neuron 220, and transfer function 230, according to some example embodiments.
  • the ANN 210 may include one or more input layers 240, one or more hidden layers 250, and one or more output layers 260.
  • Each of the input layers, hidden layers, and output layers may include one or more (artificial) neurons 220. The number of neurons can be different for different layers.
  • Each of neurons 220 may represent a calculation of a mathematical function
  • V[i] are neuron input values
  • W[i] are weights assigned to input values at neuron
  • F(X) is a transfer function.
  • the transfer function 230 F(X) is selected to be zero for X ⁇ 0 and have a limit of zero as X approaches zero.
  • the transfer function F(X) can be in the form of a sigmoid. The result of calculation of a neuron propagates as an input value of further neurons in the ANN.
  • the further neurons can belong to either a next layer, a previous layer or the same layer.
  • ANN 210 illustrated in FIG. 2 can be referred to as a feedforward neural network
  • embodiments of the present disclosure can be also used in computations of convolution neural networks, recurrent neural networks, long short-term memory networks, and other types of ANNs.
  • FIG. 3 is a flow chart showing training 310 and inference 325 of an ANN, according to some example embodiments.
  • the training 310 (also known as learning) is a process of teaching ANN 305 to output a proper result based on a given set of training data 315.
  • the process of training may include determining weights 320 of neurons of the ANN 305 based on training data 315.
  • the training data 315 may include samples. Each of the samples may be represented as a pair of input values and an expected output.
  • the training data 315 may include hundreds to millions of samples. While the training 310 is required to be performed only once, it may require a significant amount of computations and take a considerable time.
  • the ANNs can be configured to solve different tasks including, for example, image recognition, speech recognition, handwriting recognition, machine translation, social network filtering, video games, medical diagnosis, and so forth.
  • the inference 325 is a process of computation of an ANN.
  • the inference 325 uses the trained ANN weights 320 and new data 330 including new sets of input values. For each new set of input values, the computation of the ANN provides a new output which answer the problem that the ANN is supposed to solve.
  • an ANN can be trained to recognize various animals in images.
  • the ANN can be trained on millions of images of animals. Submitting a new image to the ANN would provide the information for animals in the new image (this process being known as image tagging). While the inference for each image takes less computations than training, number of inferences can be large because new images can be received from billions of sources.
  • the inference 325 includes multiple computations of sum of products:
  • V[i] are new input values and W[i] are weights associated with neurons of ANN.
  • Some previous approaches for performing inference include inspection of the weights W[i] and replacing some of the weights W[i] with zero values if a value of the weight is relatively small when compared to other weights of the ANN. In FIG. 3, this process is shown as pruning 335.
  • the pruning 335 generates new weights 340 that then can be used in inference 325 instead of the weights 320.
  • the weights 310 may remain unchanged in inference 325, while
  • Multiplications V ⁇ i ⁇ x W ⁇ L ⁇ are not carried out if a predetermined criterion is satisfied with respect to input value V[i]. For example, multiplication V[i ⁇ x W[i ⁇ can be skipped if the input value V[i] is substantially zero.
  • the criterion used to select the operations to be done may be different from a
  • embodiments of present disclosure allow dynamic selection of operations to be performed.
  • FIG. 4 is a block diagram showing a system 400 for accelerating ANN computation, according to some example embodiments.
  • the system may include an arithmetic unit 425, a controller 415, and a selector 420.
  • the controller 415 may receive a set ⁇ T[i 0 ], Tfq], , V[i x-1 ] ⁇ of X input values of data 405.
  • the controller 415 may optionally receive further input values 406 which are different from the input values 405.
  • the further input values 406 can be related to the neuron, the layer, the ANN, the weights, the operation to be carried or any other kind of values.
  • the controller 415 may provide, based on the input values 405 and the further values 406, an indication to the selector 420 as to which of the X input values are to be selected in the stream.
  • the controller 415 may also provide, to the arithmetic unit 425, an offset or an index or bit enables of one or multiple selected value(s) in the set
  • FIG. 5 is a block diagram showing a controller 415, according to some example embodiments.
  • the controller 415 can compare the input values to reference value(s) (ref).
  • the reference value(s) can be included in the further values 406. In some embodiments, the controller 415 may not use the further values 406. In some embodiments, the reference value(s) can be equal to zero.
  • the controller 415 may provide the selector 420 with one or more index(es) of the input value to be selected. In other embodiments, the controller 415 can perform a selection of an input value based on different criteria.
  • one or multiple input values 405 and multiple further values 406 can be used for the selection. In certain embodiments, other selection operations could be done involving one or multiple X input values 405 or one or multiple further values 406. For example, the input value may be selected if the input value is less than a threshold.
  • the selector 420 may receive the set of input values ⁇ V[i 0 ], T[ii F[G-i] ⁇ and the indication from the controller 415 as to which of the input values to select.
  • the selector 420 may select a value V[i] and provide the selected input value V[i] to the arithmetic unit 425.
  • the information of the selected input value may be represented in any form.
  • the controller 415 and the selector 420 may be carried out as a single unit configured to perform functionalities of both controller 415 and selector 420.
  • the selector 420 can be also configured to select weights 410 based on indications from the controller 415.
  • the arithmetic unit 425 can be configured to compute sums, multiplications, accumulations, or other operations.
  • the arithmetic unit 425 may receive the selected value V[i] from the selector 420 and the offset of index of selected value V[i] from the controller 415.
  • the arithmetic unit 425 may be further configured to select, based on the offset, a weight W[i] corresponding to value V[i].
  • the arithmetic unit 425 may further determine product V[i] x W[i ⁇ and add the product to corresponding sum. Because the multiplication is performed only for selected values of data and selected weights, the computation of sum, and hence computation of the ANN, can be accelerated.
  • the arithmetic unit 425 can determine products V[j]xW[k], wherein j and k are determined based on the input values 405, further input values 406, and the weights 410 specified by the controller 415. In some embodiments, the arithmetic unit 425 can perform further mathematical operations different from products and sums, independently, prior to, or after performing the products or sums.
  • FIG. 6 is a block diagram showing a system 600 for selecting input values to be processed in an ANN computation, according to another example embodiment.
  • the system may include a controller 415, an arithmetic unit 425, and a memory 610.
  • the memory 610 can be configured to store a set of X input values ⁇ T[i 0 ], Tfi , ... V[i x-1 ] ⁇ .
  • the controller 415 can be configured to determine an address of an input value to be selected for multiplication in arithmetic unit 425.
  • the arithmetic unit 425 may read the selected input value based on the address received from the controller 415.
  • FIG. 7 is a block diagram showing a system 700 for synchronization of results of parallel calculations of neurons in the ANN, according to an example embodiment.
  • the system 700 may include arithmetic units 710, a synchronization module 715, and a memory 720.
  • the arithmetic units 710 can process input values of different data 705.
  • the data 705 can be related to each other or be in part or generally the same input data.
  • the data 705 may represent red, green, and blue color components of the color images.
  • the data 705 may represent input values from different receptive fields selected by filters. Because the amount of input values selected for processing by different arithmetic units 710 can be different, arithmetic units 710 may finish processing of the input data at different rates.
  • results of neuron computations (performed in parallel) within a layer of the ANN can be unaligned.
  • data 0 and datcq of the data 705 may include equal numbers of input values.
  • the number and indexes of input values selected from data 0 to compute i-th neuron output by a first arithmetic unit of the arithmetic units 710 can be different from the number and indexes of input values selected from datcq to compute i-th neuron output by a second arithmetic unit of the arithmetic units 710.
  • i-th neuron output of the first arithmetic unit can be generated substantially prior to an i-th neuron output of the second arithmetic unit; or 2) the i-th neuron output of the first arithmetic unit can be generated substantially after the i-th neuron output of the second arithmetic. Therefore, a first stream of neuron outputs from the first arithmetic unit can be substantially unaligned with a second stream of neuron outputs from the second arithmetic unit even when the first arithmetic unit and the second arithmetic unit are configured to process the data 0 and data 1 in parallel.
  • the first stream of neuron outputs and the second stream of neuron outputs can also be unaligned due to the difference in the number of neuron outputs from the first arithmetic unit and the second arithmetic unit because the first arithmetic unit and the second arithmetic unit can be configured to generate different numbers of neuron outputs based on the same number of input values.
  • the synchronization module 715 can be configured to re-align outputs from arithmetic units 710.
  • the synchronization module can receive neuron outputs from arithmetic units 710, one neuron output at time, and write the neuron outputs to memory storage 720, in proper order.
  • the synchronization module 715 may receive partial results of neurons or multiple results of different neurons.
  • the neuron outputs may be received by the synchronization module 715 in an unordered time sequence O[0,0], O[0,l], O[l,0], O[0,2], 0[1,1], ... .
  • the synchronization module 715 can write the neuron outputs in a proper order: O[0,0], O[l,0], ..., O[N,0], O[0,l], 0[1,1], ..., 0[N,1], O[0,2], 0[1,2], ... .
  • the neuron outputs can be read in a proper order as inputs for neurons at further layer(s) of the ANN.
  • Data 705 may include labels 730 indicating the last input value of the stream for input values.
  • the arithmetic units 710 can be configured to send, to the synchronization module 715, an indication 735 that the last part of the input data is being processed.
  • the synchronization module 715 may finish writing neuron output to memory storage 720 and stand by ready for the next series of neuron outputs.
  • the indication 730 can be indicative of the first part of a new input data being processed rather than the last part of the input data.
  • the indication may include a number of input values to be processed. The indication can be part of the data or a separate information, as shown on FIG. 7 with marks 730 and 735.
  • FIG. 8 is a flow chart illustrating a method 800 for accelerating of ANN computations, in accordance with some example embodiments.
  • the operations may be combined, performed in parallel, or performed in a different order.
  • the method 800 may also include additional or fewer operations than those illustrated.
  • the method 800 may be performed by system 100 described above with reference to in FIG. 1.
  • the method 800 may select, by a controller communicatively coupled to a selector and an arithmetic unit and based on a criterion, an input value from a stream of input values of a neuron.
  • the controller can be configured to select input values by comparison of the input values to a reference value.
  • the reference value can be equal to zero.
  • the method 800 may configure, by the controller, the selector to provide the selected input value to the arithmetic unit.
  • the method 800 may provide, by the controller, to the arithmetic unit, information for the selected input value.
  • the information may include an offset of the selected input values in the stream.
  • the method 800 may acquire, by the arithmetic unit and based on the information, a weight from a set of weights.
  • the method 800 may perform, by the arithmetic unit, a mathematical operation on the selected input value and the weight to obtain a result, wherein the result is to be used in computing an output of the neuron.
  • the arithmetic unit may determine a multiplication product of the selected input values and weight, and summate the multiplication product into a sum.
  • a count of the input values in the stream may be greater than a count of mathematical operations performed by the arithmetic units, wherein the operations are performed on the input values selected from the stream.
  • FIG. 9 illustrates an example computing system 900 that may be used to implement embodiments described herein.
  • the example computing system 900 of FIG. 9 may include one or more processors 910 and memory 920.
  • Memory 920 may store, in part, instructions and data for execution by the one or more processors 910.
  • Memory 920 can store the executable code when the exemplary computing system 900 is in operation.
  • the processor 910 may include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar accelerators that may be suitable for use with embodiments described herein.
  • the memory 920 may include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar accelerators that may be suitable for use with embodiments described herein.
  • the example computing system 900 of FIG. 9 may further include a mass storage 930, portable storage 940, one or more output devices 950, one or more input devices 960, a network interface 970, and one or more peripheral devices 980.
  • the components shown in FIG. 9 are depicted as being connected via a single bus 990.
  • the components may be connected through one or more data transport means.
  • the one or more processors 910 and memory 920 may be connected via a local microprocessor bus, and the mass storage 930, one or more peripheral devices 980, portable storage 940, and network interface 970 may be connected via one or more input/output buses.
  • Mass storage 930 which may be implemented with a magnetic disk drive, an optical disk drive or a solid state drive, is a non-volatile storage device for storing data and instructions for use by a magnetic disk, an optical disk drive or SSD, which in turn may be used by one or more processors 910. Mass storage 930 can store the system software for implementing embodiments described herein for purposes of loading that software into memory 920.
  • the mass storage 930 may also include internal accelerators like a graphical processing unit, a Field Programmable Gate Array, or similar
  • Portable storage 940 may operate in conjunction with a portable non-volatile storage medium, such as a compact disk (CD) or digital video disc (DVD), to input and output data and code to and from the computing system 900 of FIG. 9.
  • a portable non-volatile storage medium such as a compact disk (CD) or digital video disc (DVD)
  • CD compact disk
  • DVD digital video disc
  • the system software for implementing embodiments described herein may be stored on such a portable medium and input to the computing system 900 via the portable storage 940.
  • One or more input devices 960 provide a portion of a user interface.
  • the one or more input devices 960 may include an alphanumeric keypad, such as a keyboard, for inputting alphanumeric and other information, or a pointing device, such as a mouse, a trackball, a stylus, or cursor direction keys.
  • the computing system 900 as shown in FIG. 9 includes one or more output devices 950. Suitable one or more output devices 950 include speakers, printers, network interfaces, and monitors.
  • Network interface 970 can be utilized to communicate with external devices, external computing devices, servers, and networked systems via one or more communications networks such as one or more wired, wireless, or optical networks including, for example, the Internet, intranet, LAN, WAN, cellular phone networks (e.g., Global System for Mobile communications network, packet switching
  • Network interface 970 may be a network interface card, such as an Ethernet card, optical transceiver, radio frequency transceiver, or any other type of device that can send and receive
  • network interfaces may include Bluetooth®, 3G, 4G, and WiFi® radios in mobile computing devices as well as a USB.
  • One or more peripheral devices 980 may include any type of computer support device to add additional functionality to the computing system.
  • the one or more peripheral devices 980 may include a modem or a router.
  • the example computing system 900 of FIG. 9 may also include one or more accelerator devices 985.
  • the accelerator devices 985 may include PCIe-form-f actor boards or storage-form-factor boards, or any electronic board equipped with a specific electronic component like a Graphical Processing Unit, a Neural Processing Unit, a Multi-CPU component, a Field Programmable Gate Array component, or similar accelerators electronic or photonic components, that may be suitable for use with embodiments described herein.
  • the components contained in the exemplary computing system 900 of FIG. 9 are those typically found in computing systems that may be suitable for use with embodiments described herein and are intended to represent a broad category of such computer components that are well known in the art.
  • the exemplary computing system 900 of FIG. 9 can be a personal computer, hand held computing device, telephone, mobile computing device, workstation, server, minicomputer, mainframe computer, or any other computing device.
  • the computer can also include different bus configurations, networked platforms, multi-processor platforms, and so forth.
  • Various operating systems (OS) can be used including UNIX, Linux, Windows, Macintosh OS, Palm OS, and other suitable operating systems.
  • Some of the above-described functions may be composed of instructions that are stored on storage media (e.g., computer-readable medium).
  • the instructions may be retrieved and executed by the processor.
  • Some examples of storage media are memory devices, tapes, disks, and the like.
  • the instructions are operational when executed by the processor to direct the processor to operate in accord with the example embodiments. Those skilled in the art are familiar with instructions, processor(s), and storage media.
  • Non-volatile media include, for example, optical or magnetic disks, such as a fixed disk.
  • Volatile media include dynamic memory, such as RAM.
  • Transmission media include coaxial cables, copper wire, and fiber optics, among others, including the wires that include one embodiment of a bus. Transmission media can also take the form of acoustic or light waves, such as those generated during radio frequency and infrared data
  • Common forms of computer-readable media include, for example, a floppy disk, a flexible disk, a hard disk, magnetic tape, any other magnetic medium, SSD, a CD-read-only memory (ROM) disk, DVD, any other optical medium, any other physical medium with patterns of marks or holes, a RAM, a PROM, an EPROM, an EEPROM, a FLASHEPROM, any other memory chip or cartridge, a carrier wave, or any other medium from which a computer can read.
  • Various forms of computer-readable media may be involved in carrying one or more sequences of one or more instructions to a CPU for execution.
  • a bus carries the data to system RAM, from which a CPU retrieves and executes the instructions.
  • the instructions received by system RAM can optionally be stored on a fixed disk either before or after execution by a CPU.
  • the instructions or data may not be used by the CPU but be accessed in writing or reading from the other devices without having the CPU directing them.

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)
  • Complex Calculations (AREA)

Abstract

L'invention concerne des systèmes et des procédés pour accélérer un calcul de réseau neuronal artificiel. Un exemple peut comprendre la sélection, par un contrôleur connecté en communication à un sélecteur et à une unité arithmétique et sur la base d'un critère, d'une valeur d'entrée à partir du flux de valeurs d'entrée d'un neurone, la configuration, par le contrôleur, du sélecteur pour fournir, de manière dynamique, la valeur d'entrée sélectionnée à l'unité arithmétique, la fourniture, par le contrôleur à l'unité arithmétique, d'une information de la valeur d'entrée sélectionnée, l'acquisition, par l'unité arithmétique et sur la base des informations, d'un poids à partir d'un ensemble de poids, et la réalisation, par l'unité arithmétique, d'une opération mathématique sur la valeur d'entrée sélectionnée et le poids afin d'obtenir un résultat, le résultat étant destiné à être utilisé pour calculer une sortie du neurone. Le critère peut comprendre une comparaison entre la valeur d'entrée et une valeur de référence. La valeur de référence peut comprendre zéro.
PCT/IB2018/059878 2018-12-11 2018-12-11 Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée Ceased WO2020121023A1 (fr)

Priority Applications (7)

Application Number Priority Date Filing Date Title
EP18827254.6A EP3895071A1 (fr) 2018-12-11 2018-12-11 Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée
PCT/IB2018/059878 WO2020121023A1 (fr) 2018-12-11 2018-12-11 Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée
PCT/IB2019/060631 WO2020121203A1 (fr) 2018-12-11 2019-12-10 Réalignement de flux de sorties de neurone dans des calculs de réseau de neurones artificiels
PCT/IB2019/060630 WO2020121202A1 (fr) 2018-12-11 2019-12-10 Réalignement de flux de sorties de neurone dans des calculs de réseau neuronal artificiel
EP19835488.8A EP3895072A1 (fr) 2018-12-11 2019-12-10 Réalignement de flux de sorties de neurone dans des calculs de réseau neuronal artificiel
EP19835489.6A EP3895073A1 (fr) 2018-12-11 2019-12-10 Réalignement de flux de sorties de neurone dans des calculs de réseau de neurones artificiels
US16/709,507 US11126912B2 (en) 2018-12-11 2019-12-10 Realigning streams of neuron outputs in artificial neural network computations

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/IB2018/059878 WO2020121023A1 (fr) 2018-12-11 2018-12-11 Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée

Related Child Applications (1)

Application Number Title Priority Date Filing Date
US16/215,685 Continuation-In-Part US10769527B2 (en) 2018-12-11 2018-12-11 Accelerating artificial neural network computations by skipping input values

Publications (1)

Publication Number Publication Date
WO2020121023A1 true WO2020121023A1 (fr) 2020-06-18

Family

ID=64901618

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/IB2018/059878 Ceased WO2020121023A1 (fr) 2018-12-11 2018-12-11 Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée

Country Status (2)

Country Link
EP (1) EP3895071A1 (fr)
WO (1) WO2020121023A1 (fr)

Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180121377A1 (en) * 2016-10-27 2018-05-03 Google Inc. Exploiting input data sparsity in neural network compute units
US20180218518A1 (en) * 2017-02-01 2018-08-02 Nvidia Corporation Data compaction and memory bandwidth reduction for sparse neural networks

Family Cites Families (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
GB2560600B (en) * 2017-11-06 2020-03-04 Imagination Tech Ltd Nueral Network Hardware

Patent Citations (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US20180121377A1 (en) * 2016-10-27 2018-05-03 Google Inc. Exploiting input data sparsity in neural network compute units
US20180218518A1 (en) * 2017-02-01 2018-08-02 Nvidia Corporation Data compaction and memory bandwidth reduction for sparse neural networks

Also Published As

Publication number Publication date
EP3895071A1 (fr) 2021-10-20

Similar Documents

Publication Publication Date Title
US12254400B2 (en) Optimizing artificial neural network computations based on automatic determination of a batch size
US12205006B2 (en) Accelerating neuron computations in artificial neural networks by skipping bits
US10990525B2 (en) Caching data in artificial neural network computations
WO2020165629A1 (fr) Surveillance de qualité et quantification cachée dans des calculs de réseau neuronal artificiel
US11068784B2 (en) Generic quantization of artificial neural networks
US11568255B2 (en) Fine tuning of trained artificial neural network
US11494624B2 (en) Accelerating neuron computations in artificial neural networks with dual sparsity
US10769527B2 (en) Accelerating artificial neural network computations by skipping input values
US11645510B2 (en) Accelerating neuron computations in artificial neural networks by selecting input data
US11126912B2 (en) Realigning streams of neuron outputs in artificial neural network computations
WO2020121203A1 (fr) Réalignement de flux de sorties de neurone dans des calculs de réseau de neurones artificiels
EP3895024A1 (fr) Mise en mémoire cache de données dans des calculs de réseau neuronal artificiel
WO2020121023A1 (fr) Accélération de calculs de réseau neuronal artificiel par saut de valeurs d'entrée
WO2020194032A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels par saut de bits
US20220222519A1 (en) Optimizing operations in artificial neural network
WO2020152504A1 (fr) Quantification générique de réseaux neuronaux artificiels
WO2022153078A1 (fr) Optimisation d'opérations dans un réseau de neurones artificiels
US11748623B2 (en) Modifying structure of artificial neural networks by collocating parameters
US20210117800A1 (en) Multiple locally stored artificial neural network computations
EP4211610A1 (fr) Réglage de précision d'un réseau neuronal artificiel entraîné
WO2020208396A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels par sélection des données d'entrée
WO2021079168A1 (fr) Multiples calculs stockés localement de réseau neuronal artificiel
EP3908981A1 (fr) Optimisation de calculs de réseau neuronal artificiel sur la base d'une détermination automatique d'une taille de lot
WO2021209789A1 (fr) Structure modificatrice de réseaux neuronaux artificiels en colocalisant des paramètres
EP3973464A1 (fr) Accélération de calculs neuronaux dans des réseaux neuronaux artificiels avec double dispersion

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: 18827254

Country of ref document: EP

Kind code of ref document: A1

NENP Non-entry into the national phase

Ref country code: DE

ENP Entry into the national phase

Ref document number: 2018827254

Country of ref document: EP

Effective date: 20210712