CN113516171A - Image classification method based on Bayesian neural network random addition decomposition structure - Google Patents

Image classification method based on Bayesian neural network random addition decomposition structure Download PDF

Info

Publication number
CN113516171A
CN113516171A CN202110544590.4A CN202110544590A CN113516171A CN 113516171 A CN113516171 A CN 113516171A CN 202110544590 A CN202110544590 A CN 202110544590A CN 113516171 A CN113516171 A CN 113516171A
Authority
CN
China
Prior art keywords
scaling factor
neural network
inner product
bayesian neural
bit stream
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
Application number
CN202110544590.4A
Other languages
Chinese (zh)
Other versions
CN113516171B (en
Inventor
姜书艳
孙召曦
许怡楠
黄乐天
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.)
University of Electronic Science and Technology of China
Original Assignee
University of Electronic Science and Technology of China
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 University of Electronic Science and Technology of China filed Critical University of Electronic Science and Technology of China
Priority to CN202110544590.4A priority Critical patent/CN113516171B/en
Publication of CN113516171A publication Critical patent/CN113516171A/en
Application granted granted Critical
Publication of CN113516171B publication Critical patent/CN113516171B/en
Active legal-status Critical Current
Anticipated expiration legal-status Critical

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/24—Classification techniques
    • G06F18/241—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches
    • G06F18/2415—Classification techniques relating to the classification model, e.g. parametric or non-parametric approaches based on parametric or probabilistic models, e.g. based on likelihood ratio or false acceptance rate versus a false rejection rate
    • G06F18/24155—Bayesian classification
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06F—ELECTRIC DIGITAL DATA PROCESSING
    • G06F18/00—Pattern recognition
    • G06F18/20—Analysing
    • G06F18/29—Graphical models, e.g. Bayesian networks
    • 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
    • G06N3/042—Knowledge-based neural networks; Logical representations of neural networks
    • 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
    • G06N3/045—Combinations of networks
    • 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
    • G06N3/047—Probabilistic or stochastic networks
    • 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
    • G06N3/063—Physical realisation, i.e. hardware implementation of neural networks, neurons or parts of neurons using electronic means
    • 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
    • G06N3/08—Learning methods
    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N5/00—Computing arrangements using knowledge-based models
    • G06N5/04—Inference or reasoning models
    • G06N5/046—Forward inferencing; Production systems
    • 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

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Data Mining & Analysis (AREA)
  • General Physics & Mathematics (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • General Engineering & Computer Science (AREA)
  • Artificial Intelligence (AREA)
  • Evolutionary Computation (AREA)
  • Biomedical Technology (AREA)
  • Computational Linguistics (AREA)
  • Software Systems (AREA)
  • Biophysics (AREA)
  • Computing Systems (AREA)
  • Mathematical Physics (AREA)
  • Health & Medical Sciences (AREA)
  • Molecular Biology (AREA)
  • General Health & Medical Sciences (AREA)
  • Probability & Statistics with Applications (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Evolutionary Biology (AREA)
  • Neurology (AREA)
  • Image Processing (AREA)
  • Image Analysis (AREA)

Abstract

The invention discloses an image classification method based on a Bayesian neural network random addition decomposition structure, which comprises the steps of scaling attribute parameters of a Bayesian neural network to obtain scaling attribute parameters, converting the scaling attribute parameters into random bit stream data, determining the input number of a reference multiplexer based on a network structure of the Bayesian neural network, determining the input number and the number of a middle multiplexer according to the input number of the reference multiplexer, determining an inner product scaling factor of an inner product operation unit based on a parameter scaling factor and the input number and the number of the middle multiplexer, determining an inner product operation output result of the inner product operation unit according to the random bit stream data and the middle multiplexer, determining an output scaling factor based on the inner product scaling factor and the parameter scaling factor, and outputting the final output result of the Bayesian neural network to finish image classification, the image classification is carried out by the classification method, so that the hardware implementation cost during classification is greatly reduced.

Description

Image classification method based on Bayesian neural network random addition decomposition structure
Technical Field
The invention belongs to the technical field of image classification, and particularly relates to an image classification method based on a Bayesian neural network random addition decomposition structure.
Background
Image classification is the hottest research direction in the fields of artificial intelligence and computer vision at present, and the problem of processing image classification by using a neural network becomes the mainstream in recent years.
In the prior art, a bayesian neural network is generally adopted for image classification, and in the prior art, the bayesian neural network is adopted for image classification, a gaussian random number generator, a multiplier and an adder need to work together, so that the hardware implementation cost is too high when the image classification is carried out.
Therefore, how to reduce the hardware implementation cost of image classification when using the bayesian neural network to classify images and improve the user experience is a technical problem to be solved by those skilled in the art.
Disclosure of Invention
The invention aims to solve the technical problem that the hardware implementation cost is too high when a Bayesian neural network is used for image classification in the prior art, and provides an image classification method based on a Bayesian neural network random addition decomposition structure.
The technical scheme of the invention is as follows: the image classification method based on the Bayesian neural network random addition decomposition structure comprises the following steps:
s1, scaling the attribute parameters of the Bayesian neural network according to parameter scaling factors to obtain scaling attribute parameters, and converting the scaling attribute parameters into random bit stream data, wherein the attribute parameters comprise input data, weight parameters and bias parameters, and the input data is specifically first image data to be classified in an image data set to be classified;
s2, determining the input number of the reference multi-path selector based on the network structure of the Bayesian neural network;
s3, decomposing the reference multiplexer into a plurality of intermediate multiplexers, and determining the input number and the number of the intermediate multiplexers according to the input number of the reference multiplexer;
s4, determining an inner product scaling factor of an inner product operation unit based on the parameter scaling factor and the input number and the number of the intermediate multi-path selector, and determining an inner product operation output result of the inner product operation unit according to the random bit stream data and the intermediate multi-path selector;
s5, determining an output scale factor of the Bayes neural network based on the inner product scaling factor and the parameter scaling factor, and outputting a final output result of the Bayes neural network according to the inner product operation output result and the output scale factor so as to complete image classification of first image data to be classified;
and S6, carrying out image classification on the rest image data to be classified in the image data set to be classified according to the Bayesian neural network with the determined quantity of the intermediate multi-path selectors.
Further, the parameter scaling factors specifically include an input data scaling factor, a weight parameter scaling factor, and a bias parameter scaling factor, the scaling attribute parameters specifically include floating-point input data, a floating-point weight parameter, and a floating-point bias parameter, and the random bitstream data includes an input data bitstream, a weight parameter bitstream, and a bias parameter bitstream.
Further, the step S2 is specifically to determine the number of neurons in each layer of the bayesian neural network according to the network structure of the bayesian neural network, where the input number of the reference multiplexer is the number of the neurons in the previous layer, and each of the neurons in each layer of the network except the input layer of the bayesian neural network has one reference multiplexer.
Further, in the step S3, the number of the intermediate multiplexers is smaller than the number of the inputs of the reference multiplexer, and a value obtained by dividing the number of the inputs of the reference multiplexer by the number of the intermediate multiplexers is a positive integer and equal to the number of the inputs of the intermediate multiplexers.
Further, in step S4, the weight parameter scaling factor, the input data parameter scaling factor, the number of inputs of the intermediate multiplexers, and the number of intermediate multiplexers are multiplied to obtain the inner product scaling factor of the inner product operation unit.
Further, the determining the output scale factor of the bayesian neural network in the step S5 specifically includes the following sub-steps:
s51, taking the least common multiple of the inner product scaling factor and the bias parameter scaling factor as a common scaling factor;
s52, determining a neuron scaling factor of a single neuron according to the common scaling factor;
s53, determining the neuron scaling factors corresponding to all neurons in different layers of the Bayesian neural network;
and S54, taking the largest neuron scaling factor in different layers of the Bayesian neural network as the layer scaling factor of the corresponding layer, and taking the layer scaling factor of the last layer of the Bayesian neural network as the output scaling factor.
Further, the step S5 of outputting the final output result of the bayesian neural network specifically includes the following sub-steps:
s55, inputting the inner product operation output result and the bias parameter bit stream into a bias addition unit for operation, and determining the output bit stream of the bias addition unit;
s56, inputting the output bit stream of the bias addition unit into an activation function for activation to obtain a neuron output result of the neuron;
s57, performing the Bayesian neural network forward reasoning based on the neuron output result and the layer scaling factor, thereby obtaining output bit stream data of the Bayesian neural network;
and S58, converting the output bit stream data of the Bayesian neural network into floating point result data represented by a floating point through a counter, and amplifying the floating point result data according to the output scale factor to obtain the final output result.
Further, between the step S51 and the step S52, the method further includes determining magnitudes of the inner product scaling factor and the offset parameter scaling factor, if the inner product scaling factor is smaller than the offset parameter scaling factor, updating the inner product scaling factor according to the common scaling factor, rescaling the inner product operation result according to the updated inner product scaling factor, if the inner product scaling factor is larger than the offset parameter scaling factor, updating the offset parameter scaling factor according to the common scaling factor, and rescaling the offset parameter bitstream according to the updated offset parameter scaling factor.
Further, in step S57, the bayesian neural network forward inference is performed, specifically, the scaling factor of the neuron is updated based on the layer scaling factor, and the output result of the neuron is rescaled according to the updated neuron scaling factor. And performing forward reasoning of the Bayesian neural network based on the rescaled neuron output result.
Further, the determining of the inner product operation output result of the inner product operation unit in step S4 is specifically to input the random bit stream data into all intermediate multiplexers, determine the output result of each intermediate multiplexer, and accumulate the output results corresponding to all the intermediate multiplexers through the accumulation multiplexers to obtain the inner product operation output result, where the inner product operation output result is a bit stream reduced by the inner product scaling factor.
Further, between the step S5 and the step S6, the method further includes:
a1, performing the Bayesian neural network forward reasoning for preset times in a random calculation mode, and simulating to obtain a first neuron output value distribution obtained by random calculation;
a2, performing the Bayesian neural network forward reasoning for preset times in a floating point operation mode, and simulating to obtain a second neuron output value distribution obtained by calculating random sampling values of weight and bias distribution;
a3, judging whether the difference value of the second prediction accuracy rate corresponding to the second neuron output value distribution minus the first prediction accuracy rate corresponding to the first neuron output value distribution is smaller than a preset threshold value, if so, keeping the number of the intermediate multi-path selectors unchanged, and if not, sequentially increasing or decreasing the number of the intermediate multi-path selectors until the difference value of the second prediction accuracy rate minus the first prediction accuracy rate is smaller than the preset threshold value.
Compared with the prior art, the invention has the following beneficial effects:
(1) the invention classifies the images in a Bayesian neural network random calculation mode, greatly reduces the hardware implementation cost of image classification and improves the use experience of users compared with the traditional image classification by utilizing the Bayesian neural network.
(2) The invention sets the reference multi-path selector in the Bayes neural network, determines the input number and the number of the middle multi-path selector according to the reference multi-path selector, determines the inner product scaling factor of the inner product operation unit through the middle multi-path selector and the parameter scaling factor, determines the inner product operation output result of the inner product operation unit according to the middle multi-path selector and the random bit stream data, and uses the middle multi-path selector with more number to input data, thereby reducing the error introduced by the direct sampling of the reference multi-path selector, effectively reducing the error of the Bayes neural network, and improving the accuracy of the random addition operation of the Bayes neural network.
(3) According to the invention, through adjusting the number and the input number of the intermediate multi-path selectors, additional complex structure design is not required to be introduced, the accuracy of the Bayesian neural network random addition is improved, the simplified structure of the Bayesian neural network is maintained, the complex calculation amount is avoided, and the working efficiency of the Bayesian neural network is ensured.
(4) According to the invention, the first neuron output value distribution is obtained in a random calculation mode, the second neuron output value distribution is obtained in a floating point operation mode, and the number of the intermediate multi-path selectors is adjusted by comparing the first neuron output value distribution with the second neuron output value distribution, so that the error level of the input Bayesian neural network can be controlled.
Drawings
Fig. 1 is a schematic flowchart of an image classification method based on a bayesian neural network random additive decomposition structure according to an embodiment of the present invention;
fig. 2 is a schematic structural diagram of an inner product operation unit according to an embodiment of the present invention.
Detailed Description
The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application, and it is obvious that the described embodiments are only a part of the embodiments of the present application, and not all of the embodiments. All other embodiments, which can be derived by a person skilled in the art from the embodiments given herein without making any creative effort, shall fall within the protection scope of the present application.
As described in the background art, in the prior art, when a bayesian neural network is used for image classification, a gaussian random number generator, a multiplier and an adder need to work together, and the hardware implementation cost is high when image classification is performed.
Therefore, the application provides an image classification method based on a Bayesian neural network random addition decomposition structure, which is used for solving the technical problem that hardware implementation cost is high when the Bayesian neural network is used for image classification in the prior art.
Fig. 1 is a schematic flow chart of an image classification method based on a bayesian neural network random additive decomposition structure according to an embodiment of the present invention, and the method includes the following steps:
step S1, scaling the attribute parameters of the bayesian neural network according to parameter scaling factors to obtain scaling attribute parameters, and converting the scaling attribute parameters into random bit stream data, where the attribute parameters include input data, weight parameters, and bias parameters, and the input data is specifically the first image data to be classified in the image data set to be classified.
The attribute parameters specifically include input data in a floating point form, a weight parameter and a bias parameter, the parameter scaling factors specifically include an input data scaling factor, a weight parameter scaling factor and a bias parameter scaling factor, the scaling attribute parameters specifically include floating point input data, a floating point weight parameter and a floating point bias parameter, and the random bit stream data includes an input data bit stream, a weight parameter bit stream and a bias parameter bit stream.
In practical applications, the data that is usually required to be classified is a data set, that is, an image data set to be classified, and the data set includes a plurality of image data to be classified.
Specifically, the attribute parameter may be reduced by first using an integer power of 2 as a scaling factor, and the attribute parameter is limited to a range representable by random calculation, where the data range of the input data x is [ m [ ]x,nx]The data range of the weight parameter is [ m ]w,nw]The data range of the bias parameter is [ m ]b,nb]Wherein the parameter scaling factor may be determined by the following formula:
input data scaling factor:
Figure BDA0003073064380000051
weight parameter scaling factor:
Figure BDA0003073064380000052
bias parameter scaling factor:
Figure BDA0003073064380000053
scaled floating point input data xfA floating point weight parameter wfFloating point bias parameterbfAre respectively represented by the following formula:
Figure BDA0003073064380000054
Figure BDA0003073064380000055
Figure BDA0003073064380000056
wherein x represents input data, w represents a weight parameter, and b represents a bias parameter; x is the number offRepresenting scaled floating-point input data, wfRepresenting scaled floating-point weight parameters, bfRepresenting the scaled floating point bias parameters; sxA scaling factor, s, representing the input data xwA scaling factor, s, representing a weight parameter wbRepresenting the scaling factor of the bias parameter b, m representing the left limit of the data range and n representing the right limit of the data range.
The scaling attribute parameters can be converted into a random bit stream form by using a linear feedback shift register as a random number generator and matching with a comparator, so as to obtain an input data bit stream x ', a weight parameter bit stream w ' and a bias parameter bit stream b '.
It should be noted that the above-mentioned scaling of the attribute parameter and the conversion of the scaled attribute parameter into the random bit stream are only one specific implementation manner in the present application, and those skilled in the art may flexibly select other manners according to actual situations, which does not affect the protection scope of the present application.
And step S2, determining the input number of the reference multiplexer based on the network structure of the Bayesian neural network.
The method specifically comprises the steps of firstly determining the number of the neurons in each layer of the Bayesian neural network according to the network structure of the Bayesian neural network, wherein the input number of the reference multiplexer is the number of the neurons in the previous layer, and one reference multiplexer is arranged in each neuron in each layer of the network except the input layer of the Bayesian neural network.
And step S3, decomposing the reference multiplexer into a plurality of intermediate multiplexers, and determining the input number and the number of the intermediate multiplexers according to the input number of the reference multiplexer.
In the above step S3, the number of the intermediate multiplexers is smaller than the number of the reference multiplexers, and a value obtained by dividing the number of the reference multiplexers by the number of the intermediate multiplexers is a positive integer and equal to the number of the intermediate multiplexers.
It should be noted that, in the present application, the reference multiplexer is decomposed into a plurality of intermediate multiplexers, and the number and number of the intermediate multiplexers in the present application are determined according to the number of the input of the reference multiplexer, which can reduce the information loss of the multiplexer in the sampling process in the conventional technology, and greatly reduce the operation error of the bayesian neural network during the random calculation.
Step S4, determining an inner product scaling factor of an inner product operation unit based on the parameter scaling factor and the input number and number of the intermediate multiplexer, and determining an inner product operation output result of the inner product operation unit according to the random bit stream data and the intermediate multiplexer.
Multiplying the weight parameter scaling factor, the input data parameter scaling factor, the input number of the intermediate multiplexers, and the number of the intermediate multiplexers to obtain the inner product scaling factor of the inner product operation unit.
Multiplying the weight parameter scaling factor, the input data parameter scaling factor and the input number of the intermediate multi-path selector to obtain the scaling factor s of the intermediate multi-path selectordot1The scaling factor of the intermediate multiplexer can be expressed by the following formula:
sdot1＝N/K·sw·sx
multiplying the scaling factor of the intermediate multiplexer by the number of the intermediate multiplexers to obtain an inner product scaling factor sdot of the inner product operation unit, where the inner product scaling factor can be expressed by the following formula:
sdot＝K·sdot1
where N is the input number of the reference multiplexers, K is the number of the intermediate multiplexers, swAs a scaling factor of the weight parameter, sxA scaling factor is parameterized for the input data.
Determining the inner product operation output result of the inner product operation unit in step S4, specifically, inputting the random bit stream data into all intermediate multiplexers, determining the output result of each intermediate multiplexer, and accumulating the output results corresponding to all intermediate multiplexers by an accumulation multiplexer to obtain the inner product operation output result, where the inner product operation output result is a bit stream reduced by the inner product scaling factor.
The product operation of the input data bit stream and the weight parameter bit stream can be realized through an exclusive-OR gate circuit in the Bayes neural network, and the product result is accumulated by adopting an accumulation multiplexer based on equivalent weight input to be used as the inner product output result of an inner product operation unit.
Wherein, inputting the random bit stream data into all the intermediate multiplexers, the method also comprises the step of performing product operation on the input data and the network weight parameter bit stream by adopting an exclusive-or gate circuit.
In this embodiment, as shown in fig. 2, which is a schematic structural diagram of the inner product operation unit in this embodiment, first, the input number N of the reference multiplexer is determined according to the numbers of the neurons in different layers, then, the number K and the input number N/K of the multiplexers are determined, and the intermediate output results y' of the intermediate multiplexers are respectively calculated, where the intermediate output results can be obtained by the following formula:
Figure BDA0003073064380000071
then, the output results y' corresponding to all the intermediate multiplexers are accumulated through the accumulation multiplexer to obtain inner product operation output bit stream data y, and the inner product operation output bit stream data can be obtained through the following formula:
Figure BDA0003073064380000072
wherein j represents the count value of the number of the intermediate multiplexers, i represents the count value of the number of the inputs of the intermediate multiplexers, w 'represents the weight parameter bit stream, x' represents the input data bit stream, w 'x' represents the multiplication of the input data bit stream and the weight parameter bit stream by the exclusive-or gate circuit, K represents the number of the intermediate multiplexers, and N/K represents the number of the inputs of the intermediate multiplexers.
Step S5, determining an output scale factor of the bayesian neural network based on the inner product scaling factor and the parameter scaling factor, and outputting a final output result of the bayesian neural network according to the inner product operation output result and the output scale factor, thereby completing image classification of the first image data to be classified.
Wherein, the determining the output scale factor of the bayesian neural network in the step S5 specifically includes the following sub-steps:
s51, taking the least common multiple of the inner product scaling factor and the bias parameter scaling factor as a common scaling factor;
s52, determining a neuron scaling factor of a single neuron according to the common scaling factor;
s53, determining the neuron scaling factors corresponding to all neurons in different layers of the Bayesian neural network;
and S54, taking the largest neuron scaling factor in different layers of the Bayesian neural network as the layer scaling factor of the corresponding layer, and taking the layer scaling factor of the last layer of the Bayesian neural network as the output scaling factor.
Specifically, the expression of the common scaling factor s' is as follows:
s′＝max{sdot,sb}
neuronal scaling factor szThe expression of (a) is as follows:
sz＝2·s′
wherein s isdotIs an inner product scaling factor, sbIs a bias parameter scaling factor.
Layer scaling factor sLThe expression of (a) is as follows:
Figure BDA0003073064380000081
wherein s iszFor the neuron scaling factor, i represents the network layer where the neuron is located, and n represents the number of neurons in the ith layer.
In this embodiment of the application, the outputting of the final output result of the bayesian neural network in step S5 specifically includes the following sub-steps:
s55, inputting the inner product operation output result and the bias parameter bit stream into a bias addition unit for operation, and determining the output bit stream of the bias addition unit;
s56, inputting the output bit stream of the bias addition unit into an activation function for activation to obtain a neuron output result of the neuron;
s57, performing the Bayesian neural network forward reasoning based on the neuron output result and the layer scaling factor, thereby obtaining output bit stream data of the Bayesian neural network;
and S58, converting the output bit stream data of the Bayesian neural network into floating point result data represented by a floating point through a counter, and amplifying the floating point result data according to the output scale factor to obtain the final output result.
Specifically, the determined inner product operation output result and the bias parameter bit stream are input into the bias addition unit for operation to obtain a bias operation result, that is, the output bit stream of the bias addition unit, the bias operation result is input into an activation function for activation to obtain a neuron output result corresponding to a neuron, forward reasoning of the bayesian neural network is performed based on the neuron output result and a layer scaling factor to determine output bit stream data of the bayesian neural network, in a specific application scenario, a counter can be used as a backward conversion circuit to convert the output bit stream data into floating point result data represented by a floating point, and the floating point result data is amplified according to the output scaling factor to obtain the final output result.
In this embodiment, between step S51 and step S52, the method further includes determining magnitudes of the inner product scaling factor and the offset parameter scaling factor, if the inner product scaling factor is smaller than the offset parameter scaling factor, updating the inner product scaling factor according to the common scaling factor, rescaling the inner product operation result according to the updated inner product scaling factor, and if the inner product scaling factor is larger than the offset parameter scaling factor, updating the offset parameter scaling factor according to the common scaling factor, and rescaling the offset parameter bitstream according to the updated offset parameter scaling factor.
Specifically, when the inner product operation output result is input to the offset addition unit, the sizes of the inner product scaling factor and the offset parameter scaling factor need to be determined first, then the inner product scaling factor or the offset parameter scaling factor is updated, and the inner product operation output result or the offset parameter bit stream is rescaled, where the update formula of the inner product scaling factor is as follows:
Figure BDA0003073064380000091
the update formula of the bias parameter scaling factor is as follows:
Figure BDA0003073064380000092
wherein sdot' is the updated inner product scaling factor，sb'is the updated bias parameter scaling factor and s' is the common scaling factor.
In addition, when the bayesian neural network performs the forward inference in step S57, it is necessary to update the scaling factor of the neuron based on the layer scaling factor, and rescale the output result of the neuron according to the updated neuron scaling factor, where an update formula of the neuron scaling factor is as follows:
Figure BDA0003073064380000093
in the formula, sz' is the updated neuron scaling factor, szTo be a pre-update neuron scaling factor, sLIs the layer scaling factor.
The purpose of this step is to ensure that when the network performs the next layer of inference, the network has input data with the same scaling factor, that is, the output of the current layer of neurons is the input of the next layer of neurons, and rescales the output of the neurons, so that the input data of the next layer of neurons has the same scaling.
In the embodiment of the present application, between step S5 and step S6, the method further includes:
a1, performing the Bayesian neural network forward reasoning for preset times in a random calculation mode, and simulating to obtain a first neuron output value distribution obtained by random calculation;
a2, performing the Bayesian neural network forward reasoning for preset times in a floating point operation mode, and simulating to obtain a second neuron output value distribution obtained by calculating random sampling values of weight and bias distribution;
a3, judging whether the difference value of the second prediction accuracy rate corresponding to the second neuron output value distribution minus the first prediction accuracy rate corresponding to the first neuron output value distribution is smaller than a preset threshold value, if so, keeping the number of the intermediate multi-path selectors unchanged, and if not, sequentially increasing or decreasing the number of the intermediate multi-path selectors until the difference value of the second prediction accuracy rate minus the first prediction accuracy rate is smaller than the preset threshold value.
Specifically, a first neuron output value distribution and a second neuron output value distribution are respectively obtained through a random calculation mode and a floating point operation mode, the prediction accuracy rates corresponding to the first neuron output value distribution and the second neuron output value distribution are compared, if the difference value of the first neuron output value distribution and the second neuron output value distribution is higher than a set threshold value, the number K of intermediate multi-path selectors used for calculating intermediate results is used as a reference, and the intermediate multi-path selectors are based on the K<N and N/K ∈ N+The value of K is sequentially increased or decreased, and the number of intermediate multiplexers used to calculate intermediate results is further adjusted to balance the scaling error introduced based on random addition with the error introduced by the sampling operation.
After the K value is finally selected, the prediction process of the Bayesian neural network is implemented for a plurality of times, so that errors introduced in the whole random calculation process meet the random distribution expected by the Bayesian neural network.
And step S6, carrying out image classification on the rest image data to be classified in the image data set to be classified according to the Bayesian neural network with the determined quantity of the intermediate multi-path selectors.
Specifically, the intermediate multi-path selector structure in the Bayesian neural network is determined through the first image data to be classified, and the other image data to be classified in the image data set to be classified can be subjected to image classification based on the determined Bayesian neural network, so that the cost of image classification by using the Bayesian neural network is greatly reduced.
It will be appreciated by those of ordinary skill in the art that the embodiments described herein are intended to assist the reader in understanding the principles of the invention and are to be construed as being without limitation to such specifically recited embodiments and examples. Those skilled in the art can make various other specific changes and combinations based on the teachings of the present invention without departing from the spirit of the invention, and these changes and combinations are within the scope of the invention.

Claims (10)

1. The image classification method based on the Bayesian neural network random addition decomposition structure is characterized by comprising the following steps of:
s1, scaling the attribute parameters of the Bayesian neural network according to parameter scaling factors to obtain scaling attribute parameters, and converting the scaling attribute parameters into random bit stream data, wherein the attribute parameters comprise input data, weight parameters and bias parameters, and the input data is specifically first image data to be classified in an image data set to be classified;
s2, determining the input number of the reference multi-path selector based on the network structure of the Bayesian neural network;
s3, decomposing the reference multiplexer into a plurality of intermediate multiplexers, and determining the input number and the number of the intermediate multiplexers according to the input number of the reference multiplexer;
s4, determining an inner product scaling factor of an inner product operation unit based on the parameter scaling factor and the input number and the number of the intermediate multi-path selector, and determining an inner product operation output result of the inner product operation unit according to the random bit stream data and the intermediate multi-path selector;
s5, determining an output scale factor of the Bayes neural network based on the inner product scaling factor and the parameter scaling factor, and outputting a final output result of the Bayes neural network according to the inner product operation output result and the output scale factor so as to complete image classification of first image data to be classified;
and S6, carrying out image classification on the rest image data to be classified in the image data set to be classified according to the Bayesian neural network with the determined quantity of the intermediate multi-path selectors.
2. The image classification method based on the Bayesian neural network stochastic addition decomposition structure as recited in claim 1, wherein the parameter scaling factors specifically include an input data scaling factor, a weight parameter scaling factor, and a bias parameter scaling factor, wherein the scaling attribute parameters specifically include floating-point input data, a floating-point weight parameter, and a floating-point bias parameter, and wherein the random bit stream data includes an input data bit stream, a weight parameter bit stream, and a bias parameter bit stream.
3. The image classification method based on the bayesian neural network stochastic additive decomposition structure as claimed in claim 2, wherein said step S2 is to determine the number of neurons in each layer of the bayesian neural network according to the network structure of the bayesian neural network, and the input number of the reference multiplexer is the number of the neurons in the previous layer, wherein the number of the reference multiplexers in each of the neurons in each layer of the network except the input layer of the bayesian neural network is one.
4. The image classification method based on the bayesian neural network stochastic additive factorization structure as claimed in claim 2, wherein in the step S3, the number of the intermediate multiplexers is smaller than the number of the inputs of the reference multiplexer, and a value of the number of the inputs of the reference multiplexer divided by the number of the intermediate multiplexers is a positive integer and equal to the number of the inputs of the intermediate multiplexers.
5. The image classification method based on the bayesian neural network stochastic additive decomposition structure as claimed in claim 2, wherein in step S4, the weight parameter scaling factor, the input data parameter scaling factor, the number of inputs of the intermediate multiplexers, and the number of intermediate multiplexers are multiplied to obtain the inner product scaling factor of the inner product operation unit.
6. The image classification method based on the bayesian neural network stochastic additive decomposition structure as claimed in claim 2, wherein the determining of the output scale factor of the bayesian neural network in step S5 specifically comprises the following substeps:
s51, taking the least common multiple of the inner product scaling factor and the bias parameter scaling factor as a common scaling factor;
s52, determining a neuron scaling factor of a single neuron according to the common scaling factor;
s53, determining the neuron scaling factors corresponding to all neurons in different layers of the Bayesian neural network;
and S54, taking the largest neuron scaling factor in different layers of the Bayesian neural network as the layer scaling factor of the corresponding layer, and taking the layer scaling factor of the last layer of the Bayesian neural network as the output scaling factor.
7. The image classification method based on the bayesian neural network stochastic additive decomposition structure as claimed in claim 6, wherein said step S5 of outputting the final output result of the bayesian neural network specifically comprises the following substeps:
s55, inputting the inner product operation output result and the bias parameter bit stream into a bias addition unit for operation, and determining the output bit stream of the bias addition unit;
s56, inputting the output bit stream of the bias addition unit into an activation function for activation to obtain a neuron output result of the neuron;
s57, performing the Bayesian neural network forward reasoning based on the neuron output result and the layer scaling factor, thereby obtaining output bit stream data of the Bayesian neural network;
and S58, converting the output bit stream data of the Bayesian neural network into floating point result data represented by a floating point through a counter, and amplifying the floating point result data according to the output scale factor to obtain the final output result.
8. The image classification method based on the Bayesian neural network stochastic addition decomposition structure as recited in claim 6, further comprising, between step S51 and step S52, determining magnitudes of the inner product scaling factor and the bias parameter scaling factor, if the inner product scaling factor is smaller than the bias parameter scaling factor, updating the inner product scaling factor according to the common scaling factor, rescaling the inner product operation result according to the updated inner product scaling factor, if the inner product scaling factor is larger than the bias parameter scaling factor, updating the bias parameter scaling factor according to the common scaling factor, and rescaling the bias parameter bit stream according to the updated bias parameter scaling factor.
9. The image classification method based on the bayesian neural network stochastic additive decomposition structure as recited in claim 1, wherein the determining of the inner product operation output result of the inner product operation unit in step S4 is to input the random bit stream data to all intermediate multiplexers, determine the output result of each intermediate multiplexer, and accumulate the output results corresponding to all intermediate multiplexers by the accumulation multiplexers to obtain the inner product operation output result, wherein the inner product operation output result is a bit stream reduced by the inner product scaling factor.
10. The image classification method based on the bayesian neural network stochastic additive decomposition structure as recited in claim 1, further comprising, between the step S5 and the step S6:
a1, performing the Bayesian neural network forward reasoning for preset times in a random calculation mode, and simulating to obtain a first neuron output value distribution obtained by random calculation;
a2, performing the Bayesian neural network forward reasoning for preset times in a floating point operation mode, and simulating to obtain a second neuron output value distribution obtained by calculating random sampling values of weight and bias distribution;
a3, judging whether the difference value of the second prediction accuracy rate corresponding to the second neuron output value distribution minus the first prediction accuracy rate corresponding to the first neuron output value distribution is smaller than a preset threshold value, if so, keeping the number of the intermediate multi-path selectors unchanged, and if not, sequentially increasing or decreasing the number of the intermediate multi-path selectors until the difference value of the second prediction accuracy rate minus the first prediction accuracy rate is smaller than the preset threshold value.
CN202110544590.4A 2021-05-19 2021-05-19 Image classification method based on Bayes neural network random addition decomposition structure Active CN113516171B (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202110544590.4A CN113516171B (en) 2021-05-19 2021-05-19 Image classification method based on Bayes neural network random addition decomposition structure

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202110544590.4A CN113516171B (en) 2021-05-19 2021-05-19 Image classification method based on Bayes neural network random addition decomposition structure

Publications (2)

Publication Number Publication Date
CN113516171A true CN113516171A (en) 2021-10-19
CN113516171B CN113516171B (en) 2023-04-07

Family

ID=78064500

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202110544590.4A Active CN113516171B (en) 2021-05-19 2021-05-19 Image classification method based on Bayes neural network random addition decomposition structure

Country Status (1)

Country Link
CN (1) CN113516171B (en)

Cited By (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113240025A (en) * 2021-05-19 2021-08-10 电子科技大学 Image classification method based on Bayesian neural network weight constraint

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107368797A (en) * 2017-07-06 2017-11-21 湖南中云飞华信息技术有限公司 The parallel method for detecting human face of multi-angle, device and terminal device
CN110442980A (en) * 2019-08-09 2019-11-12 重庆大学 Certainty damnification recognition method based on similar bayes method
US20200097762A1 (en) * 2018-09-26 2020-03-26 Samsung Electronics Co., Ltd. Method and apparatus for multi-category image recognition
CN112698811A (en) * 2021-01-11 2021-04-23 湖北大学 Neural network random number generator sharing circuit, sharing method and processor chip

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN107368797A (en) * 2017-07-06 2017-11-21 湖南中云飞华信息技术有限公司 The parallel method for detecting human face of multi-angle, device and terminal device
US20200097762A1 (en) * 2018-09-26 2020-03-26 Samsung Electronics Co., Ltd. Method and apparatus for multi-category image recognition
CN110442980A (en) * 2019-08-09 2019-11-12 重庆大学 Certainty damnification recognition method based on similar bayes method
CN112698811A (en) * 2021-01-11 2021-04-23 湖北大学 Neural network random number generator sharing circuit, sharing method and processor chip

Non-Patent Citations (2)

* Cited by examiner, † Cited by third party
Title
RUIZHE CAI 等: ""Hardware Acceleration of Bayesian Neural Networks Using RAM Based Linear Feedback Gaussian Random Number Generators"", 《2017 IEEE INTERNATIONAL CONFERENCE ON COMPUTER DESIGN (ICCD)》 *
孙召曦: ""低开销贝叶斯神经网络推理加速器研究"", 《万方》 *

Cited By (2)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN113240025A (en) * 2021-05-19 2021-08-10 电子科技大学 Image classification method based on Bayesian neural network weight constraint
CN113240025B (en) * 2021-05-19 2022-08-12 电子科技大学 Image classification method based on Bayesian neural network weight constraint

Also Published As

Publication number Publication date
CN113516171B (en) 2023-04-07

Similar Documents

Publication Publication Date Title
Ma et al. A hybrid attention-based deep learning approach for wind power prediction
CN111797122B (en) Method and device for predicting change trend of high-dimensional reappearance concept drift stream data
Kung et al. A power-aware digital feedforward neural network platform with backpropagation driven approximate synapses
CN111353578A (en) Information processing apparatus, neural network program, and processing method for neural network
WO2019006976A1 (en) Neural network weight discretizing method, system and device, and readable storage medium
CN114283320B (en) Branch-free structure target detection method based on full convolution
CN113516171B (en) Image classification method based on Bayes neural network random addition decomposition structure
CN111353534B (en) Graph data category prediction method based on adaptive fractional order gradient
CN112149896A (en) Attention mechanism-based mechanical equipment multi-working-condition fault prediction method
US10410140B1 (en) Categorical to numeric conversion of features for machine learning models
Qiao et al. A framework for multi-prototype based federated learning: Towards the edge intelligence
US20210256389A1 (en) Method and system for training a neural network
CN114492631A (en) Spatial attention calculation method based on channel attention
Cheng et al. Minimizing power for neural network training with logarithm-approximate floating-point multiplier
Kalyanam et al. Range based hardware optimization of multilayer perceptrons with RELUs
CN113128666A (en) Mo-S-LSTMs model-based time series multi-step prediction method
Panchariya et al. Nonlinear system identification using Takagi-Sugeno type neuro-fuzzy model
CN115619563A (en) Stock price analysis method based on neural network
Temenos et al. A stochastic computing sigma-delta adder architecture for efficient neural network design
CN113516172A (en) Image classification method based on random computation Bayesian neural network error injection
CN113516170B (en) Image classification method based on Bayesian neural network random addition saturated structure
CN115660038A (en) Multi-stage integrated short-term load prediction based on error factors and improved MOEA/D-SAS
Yu et al. Analysis and Application of the Spatio-Temporal Feature in Wind Power Prediction.
CN117348837A (en) Quantization method and device for floating point precision model, electronic equipment and storage medium
Zhang et al. Adaptive truncation technique for constrained multi-objective optimization

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