CN116304889A - Receptor classification method based on convolution and transducer - Google Patents

Receptor classification method based on convolution and transducer Download PDF

Info

Publication number
CN116304889A
CN116304889A CN202310578918.3A CN202310578918A CN116304889A CN 116304889 A CN116304889 A CN 116304889A CN 202310578918 A CN202310578918 A CN 202310578918A CN 116304889 A CN116304889 A CN 116304889A
Authority
CN
China
Prior art keywords
sequence
convolution
neural networks
acceptor
function
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.)
Pending
Application number
CN202310578918.3A
Other languages
Chinese (zh)
Inventor
刘峻江
周树森
臧睦君
王庆军
柳婵娟
刘通
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.)
Ludong University
Original Assignee
Ludong University
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 Ludong University filed Critical Ludong University
Priority to CN202310578918.3A priority Critical patent/CN116304889A/en
Publication of CN116304889A publication Critical patent/CN116304889A/en
Pending legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G06—COMPUTING; CALCULATING OR COUNTING
    • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
    • G06N3/00—Computing arrangements based on biological models
    • G06N3/02—Neural networks
    • G06N3/08—Learning methods
    • G06N3/082—Learning methods modifying the architecture, e.g. adding, deleting or silencing nodes or connections
    • 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
    • G06N3/084—Backpropagation, e.g. using gradient descent
    • G—PHYSICS
    • G16—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR SPECIFIC APPLICATION FIELDS
    • G16B—BIOINFORMATICS, i.e. INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR GENETIC OR PROTEIN-RELATED DATA PROCESSING IN COMPUTATIONAL MOLECULAR BIOLOGY
    • G16B40/00—ICT specially adapted for biostatistics; ICT specially adapted for bioinformatics-related machine learning or data mining, e.g. knowledge discovery or pattern finding
    • G16B40/20—Supervised data analysis

Landscapes

  • Engineering & Computer Science (AREA)
  • Physics & Mathematics (AREA)
  • Theoretical Computer Science (AREA)
  • Health & Medical Sciences (AREA)
  • Life Sciences & Earth Sciences (AREA)
  • Data Mining & Analysis (AREA)
  • Artificial Intelligence (AREA)
  • Biophysics (AREA)
  • Software Systems (AREA)
  • Evolutionary Computation (AREA)
  • General Health & Medical Sciences (AREA)
  • General Physics & Mathematics (AREA)
  • Computing Systems (AREA)
  • General Engineering & Computer Science (AREA)
  • Molecular Biology (AREA)
  • Mathematical Physics (AREA)
  • Computational Linguistics (AREA)
  • Biomedical Technology (AREA)
  • Medical Informatics (AREA)
  • Databases & Information Systems (AREA)
  • Computer Vision & Pattern Recognition (AREA)
  • Epidemiology (AREA)
  • Bioethics (AREA)
  • Public Health (AREA)
  • Bioinformatics & Cheminformatics (AREA)
  • Bioinformatics & Computational Biology (AREA)
  • Biotechnology (AREA)
  • Evolutionary Biology (AREA)
  • Spectroscopy & Molecular Physics (AREA)
  • Peptides Or Proteins (AREA)

Abstract

The invention belongs to the field of bioinformatics, relates to a receptor classification method based on convolution and Transformer, and aims to improve the performance of extracting the characteristics of a cell receptor sequence and predicting the type of the receptor sequence. The method adopts a word vector matrix to process all acceptor sequences, uses two layers of convolutional neural networks to extract primary characteristics, uses two Transformer encoders with different attention head numbers to extract further characteristics, and finally uses two convolutional neural networks with different output channel numbers to extract deep characteristics and obtain classification results. The method comprises the following steps: four steps of acceptor sequence pretreatment, primary feature extraction, time sequence feature extraction, advanced feature extraction and classification. The receptor classification method based on convolution and transform can improve the data utilization rate, effectively extract the time sequence characteristics of the receptor sequence, obtain a better classification effect and have wide application value.

Description

Receptor classification method based on convolution and transducer
Technical Field
The invention belongs to the field of bioinformatics, and relates to a receptor classification method based on convolution and a transducer.
Background
Cellular receptors play an important role in human vital activities, and recent studies have shown that human physical states can be judged by the receptor state of cells, for example, the receptor sequence of T cells can be used for early detection of cancer.
There are many advanced neural network approaches that have been successfully used for cell receptor sequence classification. However, most neural network approaches have the following disadvantages due to the variable length of the cellular receptor sequences: (1) The need to process cell receptor sequences of different lengths using multiple different models greatly reduces the utilization of the data. (2) The time sequence features of the receptor sequences cannot be extracted effectively, and the classification effect cannot reach the expected target.
Disclosure of Invention
The invention provides a method for solving the problem of poor classification effect of a receptor sequence, namely a method for classifying the receptor sequence based on convolution and a transducer, wherein the transducer is a deep learning model framework using an attention mechanism, and the RNN in a traditional neural network is replaced by using a multi-head self-attention mechanism, so that parallel calculation can be performed to obtain a faster training speed and a better effect. First, a corresponding word vector is set for each amino acid in the acceptor sequence, and shorter cellular acceptor sequences are filled with nonsensical word vectors so that all acceptor sequences are the same length and acceptor sequence features can be extracted. And extracting primary features of the word vector matrix by using two layers of convolutional neural networks, extracting further features by using two converger encoders with different attention head numbers, and finally extracting deep features by using two convolutional neural networks with different output channel numbers to obtain a classification result.
In order to achieve the above purpose, the technical scheme adopted by the invention is as follows:
the acceptor classifying method based on convolution and transform includes four steps of acceptor sequence preprocessing, primary feature extraction, time sequence feature extraction, advanced feature extraction and classification, and the specific steps are as follows:
step 1, setting a word vector for each amino acid in a receptor sequence, mapping the receptor sequence into a corresponding word vector matrix, and filling a shorter word vector matrix to the maximum length by using nonsensical word vectors;
step 2, sequentially passing the word vector matrix obtained in the step 1 through two convolutional neural networks with different convolutional kernel sizes, an activation function and a normalization function, so as to obtain primary characteristics of a receptor sequence;
step 3, constructing two converters with different attention heads, respectively transmitting the primary characteristics obtained in the step 2 into the two converters, and splicing the results obtained by the two converters together to obtain the time sequence characteristics of the receptor sequence;
and 4, constructing two convolutional neural networks with different output channel numbers, respectively transmitting the time sequence features obtained in the step 3 into the two convolutional neural networks, splicing the results obtained by the two convolutional neural networks together end to end, so as to obtain the advanced features of the receptor sequence, and transmitting the features into a full-connection layer formed by two linear layers to obtain a final prediction result.
A receptor classification method based on convolution and transform includes the following implementation process in step 1:
each amino acid in the acceptor sequence is mapped to an index value corresponding to the amino acid, and acceptor sequences of less than the maximum length are filled in using indexes other than the amino acid index value so that the lengths of all acceptor sequences are the same. Generating a word vector with the same dimension for each amino acid index and additional indexes used for filling by using an nn.decoding function of a Pytorch framework, and mapping the filled acceptor sequence into a word vector matrix.
The acceptor classifying method based on convolution and transducer includes the following steps:
and (3) randomly dividing the data set, wherein two thirds of the data set is used as a training set, one third of the data set is used as a test set, the acceptor sequence word vector matrix generated in the step (1) is used as input, two convolution neural networks with different convolution kernels are constructed by using an nn.Conv1d function of a Pytorch framework, the acceptor sequence word vector matrix is used as input to the two convolution neural networks, an activation function and a normalization function in sequence, and therefore primary characteristics of an acceptor sequence are extracted.
A receptor classification method based on convolution and transform comprises the following implementation process in step 3:
constructing two converterlers with different attention header numbers by using nn.converterler encorber layer and nn.converterler encorber function of Pytorch framework, respectively transmitting the primary features obtained in the step 2 into the two converters, and splicing the results obtained by the two converters together to obtain the time sequence features of the acceptor sequence.
A receptor classification method based on convolution and transform comprises the following implementation process in step 4:
and (3) constructing two convolutional neural networks with different output channel numbers by using an nn.Conv1d function of a Pytorch framework, and respectively transmitting the time sequence characteristics obtained in the step (2) into the two convolutional neural networks, an activation function, a normalization function and a random inactivation function. And constructing a full-connection layer comprising two linear layers by using an nn.linear function, splicing the results obtained by the two convolutional neural networks together end to end, and transmitting the results into the full-connection layer to obtain a final prediction result.
Drawings
FIG. 1 is a flow chart of a method for classifying receptors based on convolution and transducer.
FIG. 2 is a flow chart of a pretreatment of a receptor sequence.
Fig. 3 is a flow chart for extracting primary features.
Fig. 4 is a flow chart for extracting time series features.
Fig. 5 is a flow chart for extracting high-level features.
Fig. 6 is a full connection flow chart.
Detailed Description
The invention is described in detail below with reference to the drawings and examples.
The invention provides a receptor classification method based on convolution and a transducer, which is particularly used for classifying receptor sequences.
Fig. 1 is a flow chart of a receptor classification method based on convolution and transform, which comprises four steps of receptor sequence pretreatment, primary feature extraction, time sequence feature extraction, and advanced feature extraction and classification, and the specific implementation modes are as follows:
step 1: preprocessing acceptor sequence data fig. 2 is a flowchart of preprocessing acceptor sequence data, including the following:
each receptor sequence is a one-dimensional matrix of sizetcr_lenX 1, as shown in equation 1:
Figure SMS_1
(1)
wherein the method comprises the steps of
Figure SMS_2
In order to constitute one of the 20 amino acids of the acceptor sequence, the 20 amino acids are glycine, alanine, valine, leucine, isoleucine, methionine (methionine), proline, tryptophan, serine, tyrosine, cysteine, phenylalanine, asparagine, glutamine, threonine, aspartic acid, glutamic acid, lysine, arginine, respectivelyAnd a histidine group, which is a group of amino acids,tcr_lenthe acceptor sequences were 12, 13, 14, 15, 16, 17, respectively, for the initial length of the acceptor sequence, which was 6 in total. An index table is established for 20 amino acids, each index corresponds to a 15-dimensional word vector, and 1 index and the corresponding word vector are additionally added into the index table to serve as nonsensical word vectors for filling. In the index table, proteins W correspond to 0, F correspond to 1, G correspond to 2, a correspond to 3, V correspond to 4, I correspond to 5, L correspond to 6, M correspond to 7, P correspond to 8, Y correspond to 9, S correspond to 10, T correspond to 11, N correspond to 12, Q correspond to 13, C correspond to 14, K correspond to 15, R correspond to 16, H correspond to 17, D correspond to 18, E correspond to 19, 20 are index values used for filling. Each amino acid in the receptor sequence is replaced by the index value corresponding to the amino acid, the length of the receptor sequence with the length less than 17 is filled to 17, and the T cell receptor sequence is used for->
Figure SMS_3
For example, the index value sequence of the receptor sequence obtained by converting it into an index and filling it is
Figure SMS_4
Where 20 is the index value of the word vector used for filling. Using nn.decoding function of Pytorch frame to generate a word vector with unified dimension for each amino acid index and filling extra index, the word vector dimension is 516, and the cell receptor sequence is mapped into word vector matrix according to index value sequence, as shown in formula 2:
Figure SMS_5
(2)
wherein the method comprises the steps of
Figure SMS_6
Is the first in the acceptor sequencei516-dimensional word vectors of amino acids,tc_max_lenfor the maximum length 17 of all cell receptor sequences, the generated word vector matrix size is 17×516.
Step 2: primary feature extraction fig. 3 is a flowchart for extracting primary features, and the details thereof are as follows:
the data sets are randomly divided, wherein two thirds of the data sets are used as training sets, one third of the data sets are used as test sets, the torch.nn.cross Entropyloss is used as a loss function of training, the torch.optim.AdamW is used as a training optimizer, and the torch.optim.lr_schedule.MultiStepLR is used for learning rate reduction. And (2) taking the acceptor sequence word vector matrix generated in the step (1) as input, and constructing a convolutional neural network of two convolution kernels with different sizes by using an nn.Conv1d function of a Pytorch framework. The convolution kernel of the first convolution neural network is 3 in size, and the step length is 1; the convolution kernel size of the second convolution neural network is 5, and the step size is 1. And sequentially inputting the acceptor sequence word vector matrix as input into two convolutional neural networks, an activation function generated by nn. Relu and a normalization function generated by nn. Batchnormal, and performing random inactivation operation with 30% probability by using the nn. Dropout function after the last normalization function is finished, so as to extract the primary characteristics of the acceptor sequence. The primary feature dimension is 17×510, which can accelerate the convergence of the transducer model and improve the final effect, and the feature of the dimension 510 can provide the transducer with a choice of more attention header numbers.
Step 3: time sequence feature extraction, fig. 4 is a flowchart of extracting time sequence features, which is specifically as follows:
two specific different numbers of attention header Transformer encoders are constructed using nn.transformerlerderler layer and nn.transformerler encoder functions of the pytorrch framework. The first encoder had a number of attention heads of 10, the second encoder had a number of attention heads of 5, the number of hidden layers of both encoders was 3, the input dimension 510, and the random deactivation rate was 0.3. And (3) respectively transmitting the primary characteristics obtained in the step (2) into two transducer encoders, and splicing the results obtained by the two encoders together to obtain the time sequence characteristics of the receptor sequence. The dimensions of the feature are 34×510, and this operation can extract time series features of different dimensions of the acceptor sequence.
Step 4: advanced feature extraction and classification, fig. 5 is a flowchart of the extraction of advanced features, and fig. 6 is a fully connected flowchart, which is specifically described as follows:
taking the time sequence characteristics generated in the step 3 as input, constructing two one-dimensional convolution modules with the same structure but different output channel numbers by using an nn.Conv1d function of a Pytorch framework. The time series features are input as inputs to the two convolution modules, respectively, with the first dimension 34 of the time series features as the number of input channels of the convolutional neural network. The number of output channels of the first convolutional neural network is 1200, and the number of output channels of the second convolutional neural network is 900. The two convolutional neural networks use the same convolutional kernel, with a size 510 and a step size 1. The two results obtained by the convolutional network are respectively passed into an activation function, which is generated by nn. The results from the activation function are then each passed into a normalization function, which is generated by nn. And then carrying out random inactivation operation on the result of the normalization function with 0.3 as probability, wherein the random inactivation function is nn. And splicing the results obtained by the convolution modules to obtain high-level characteristics, wherein the dimension is 2100. Constructing two linear layers by using an nn.linear function as a full-connection module, wherein the input dimension of the first linear layer is 2100, and the output dimension is 1000; the second linear layer has an input dimension of 1000 and an output dimension of 2. And inputting the advanced features into the fully-connected module to obtain a prediction result.
On the data set provided by deep, using torch.nn. Cross Entropyloss as a loss function of training, using torch.optim.AdamW as a training optimizer, using torch.optim.lr_schedule.MultiStepLR to perform learning rate decrease, and performing gridding search to obtain optimal model parameters as shown in the following table.
TABLE 1 optimal model parameters
Parameter name Parameter value
Number of iterations 2000
Batch size 200
Random rate of deactivation 0.3
The first convolution module outputs the channel number 1200
The second convolution module outputs the channel number 900
Learning rate 1×10 -6
Regularization parameters 5×10 -3
The receptor classification method provided by the invention is applied to classifying whether a detected person has cancer or not according to the T cell receptor sequence. The Accuracy and the AUC obtained by ten times of three-fold cross validation on the data set provided by the DeepCat are respectively 0.80 and 0.86, and are better than the performance of the DeepCat on the data set, wherein the Accuracy and the AUC of the DeepCat are respectively 0.70 and 0.74, so that the invention is higher than other advanced methods on important evaluation standards of the classification effect, namely the Accuracy and the AUC, and therefore, compared with the prior art, the invention effectively improves the classification Accuracy. The invention uses word vector to process and fill the acceptor sequence, which improves the data utilization rate, and can well extract the time sequence characteristics of the acceptor sequence through word vector and neural network, thus the invention can obtain better acceptor sequence classification performance compared with other methods.
The foregoing is a further detailed description of the invention in connection with the preferred embodiments, and it is not intended that the invention be limited to the specific embodiments described. It will be apparent to those skilled in the art that several simple deductions or substitutions may be made without departing from the spirit of the invention, and these should be considered to be within the scope of the invention.

Claims (5)

1. The acceptor classifying method based on convolution and transform includes four steps of acceptor sequence preprocessing, primary feature extraction, time sequence feature extraction, advanced feature extraction and classification, and the specific steps are as follows:
step 1, setting a word vector for each amino acid in a receptor sequence, mapping the receptor sequence into a corresponding word vector matrix, and filling a shorter word vector matrix to the maximum length by using nonsensical word vectors;
step 2, sequentially passing the word vector matrix obtained in the step 1 through two convolutional neural networks with different convolutional kernel sizes, an activation function and a normalization function, so as to obtain primary characteristics of a receptor sequence;
step 3, constructing two converters with different attention heads, respectively transmitting the primary characteristics obtained in the step 2 into the two converters, and splicing the results obtained by the two converters together to obtain the time sequence characteristics of the receptor sequence;
and 4, constructing two convolutional neural networks with different output channel numbers, respectively transmitting the time sequence features obtained in the step 3 into the two convolutional neural networks, splicing the results obtained by the two convolutional neural networks together end to end, so as to obtain the advanced features of the receptor sequence, and transmitting the features into a full-connection layer formed by two linear layers to obtain a final prediction result.
2. The method for classifying receptors based on convolution and transform according to claim 1, wherein the implementation process of step 1 is as follows:
mapping each amino acid in the acceptor sequence to an index value corresponding to the amino acid, and filling acceptor sequences with less than the maximum length by using indexes except the amino acid index value so as to ensure that the lengths of all the acceptor sequences are the same; generating a word vector with the same dimension for each amino acid index and additional indexes used for filling by using an nn.decoding function of a Pytorch framework, and mapping the filled acceptor sequence into a word vector matrix.
3. The method for classifying receptors based on convolution and transform according to claim 1, wherein the implementation process of step 2 is as follows:
and (3) randomly dividing the data set, wherein two thirds of the data set is used as a training set, one third of the data set is used as a test set, the acceptor sequence word vector matrix generated in the step (1) is used as input, two convolution neural networks with different convolution kernels are constructed by using an nn.Conv1d function of a Pytorch framework, the acceptor sequence word vector matrix is used as input to the two convolution neural networks, an activation function and a normalization function in sequence, and therefore primary characteristics of an acceptor sequence are extracted.
4. The method for classifying receptors based on convolution and transform according to claim 1, wherein the implementation process of step 3 is as follows:
constructing two converterlers with different attention header numbers by using nn.converterler encorber layer and nn.converterler encorber function of Pytorch framework, respectively transmitting the primary features obtained in the step 2 into the two converters, and splicing the results obtained by the two converters together to obtain the time sequence features of the acceptor sequence.
5. The method for classifying receptors based on convolution and transform according to claim 1, wherein the implementation process of step 4 is as follows:
constructing two convolutional neural networks with different output channel numbers by using an nn.Conv1d function of a Pytorch framework, and respectively transmitting the time sequence characteristics obtained in the step 2 into the two convolutional neural networks, an activation function, a normalization function and a random inactivation function; and constructing a full-connection layer comprising two linear layers by using an nn.linear function, splicing the results obtained by the two convolutional neural networks together end to end, and transmitting the results into the full-connection layer to obtain a final prediction result.
CN202310578918.3A 2023-05-22 2023-05-22 Receptor classification method based on convolution and transducer Pending CN116304889A (en)

Priority Applications (1)

Application Number Priority Date Filing Date Title
CN202310578918.3A CN116304889A (en) 2023-05-22 2023-05-22 Receptor classification method based on convolution and transducer

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
CN202310578918.3A CN116304889A (en) 2023-05-22 2023-05-22 Receptor classification method based on convolution and transducer

Publications (1)

Publication Number Publication Date
CN116304889A true CN116304889A (en) 2023-06-23

Family

ID=86785342

Family Applications (1)

Application Number Title Priority Date Filing Date
CN202310578918.3A Pending CN116304889A (en) 2023-05-22 2023-05-22 Receptor classification method based on convolution and transducer

Country Status (1)

Country Link
CN (1) CN116304889A (en)

Cited By (3)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116913383A (en) * 2023-09-13 2023-10-20 鲁东大学 T cell receptor sequence classification method based on multiple modes
CN117095825A (en) * 2023-10-20 2023-11-21 鲁东大学 Human immune state prediction method based on multi-instance learning
CN117854601A (en) * 2024-03-04 2024-04-09 鲁东大学 Decisive complementary region dividing method based on gene type and amino acid sequence

Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109817276A (en) * 2019-01-29 2019-05-28 鲁东大学 A kind of secondary protein structure prediction method based on deep neural network
CN111667884A (en) * 2020-06-12 2020-09-15 天津大学 Convolutional neural network model for predicting protein interactions using protein primary sequences based on attention mechanism
CN115062145A (en) * 2022-05-26 2022-09-16 电子科技大学 Cloud ERP community cross-domain problem classification method based on BERT-TextCNN
CN115713970A (en) * 2022-11-16 2023-02-24 武汉大学 Transcription factor identification method based on Transformer-Encoder and multi-scale convolutional neural network

Patent Citations (4)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN109817276A (en) * 2019-01-29 2019-05-28 鲁东大学 A kind of secondary protein structure prediction method based on deep neural network
CN111667884A (en) * 2020-06-12 2020-09-15 天津大学 Convolutional neural network model for predicting protein interactions using protein primary sequences based on attention mechanism
CN115062145A (en) * 2022-05-26 2022-09-16 电子科技大学 Cloud ERP community cross-domain problem classification method based on BERT-TextCNN
CN115713970A (en) * 2022-11-16 2023-02-24 武汉大学 Transcription factor identification method based on Transformer-Encoder and multi-scale convolutional neural network

Cited By (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN116913383A (en) * 2023-09-13 2023-10-20 鲁东大学 T cell receptor sequence classification method based on multiple modes
CN116913383B (en) * 2023-09-13 2023-11-28 鲁东大学 T cell receptor sequence classification method based on multiple modes
CN117095825A (en) * 2023-10-20 2023-11-21 鲁东大学 Human immune state prediction method based on multi-instance learning
CN117095825B (en) * 2023-10-20 2024-01-05 鲁东大学 Human immune state prediction method based on multi-instance learning
CN117854601A (en) * 2024-03-04 2024-04-09 鲁东大学 Decisive complementary region dividing method based on gene type and amino acid sequence
CN117854601B (en) * 2024-03-04 2024-05-14 鲁东大学 Decisive complementary region dividing method based on gene type and amino acid sequence

Similar Documents

Publication Publication Date Title
CN116304889A (en) Receptor classification method based on convolution and transducer
CN113035280A (en) RBP binding site prediction algorithm based on deep learning
CN111832546A (en) Lightweight natural scene text recognition method
CN111291647A (en) Single-stage action positioning method based on multi-scale convolution kernel and superevent module
CN113240683B (en) Attention mechanism-based lightweight semantic segmentation model construction method
Liu et al. RB-Net: Training highly accurate and efficient binary neural networks with reshaped point-wise convolution and balanced activation
CN111814626A (en) Dynamic gesture recognition method and system based on self-attention mechanism
CN113741858B (en) Memory multiply-add computing method, memory multiply-add computing device, chip and computing equipment
CN108629412A (en) A kind of neural metwork training accelerated method based on mesh free maximum mutual information criterion
CN114420211A (en) Attention mechanism-based RNA-protein binding site prediction method
CN110781912A (en) Image classification method based on channel expansion inverse convolution neural network
CN113052264A (en) Method for compressing target detection neural network
CN112860904A (en) External knowledge-integrated biomedical relation extraction method
CN112101487A (en) Compression method and device for fine-grained recognition model
CN116340944A (en) Malicious code classification method and system based on RGB image and lightweight model
CN114880538A (en) Attribute graph community detection method based on self-supervision
Byttebier et al. Small-footprint acoustic scene classification through 8-bit quantization-aware training and pruning of ResNet models
CN111126560A (en) Method for optimizing BP neural network based on cloud genetic algorithm
CN114299305A (en) Salient object detection algorithm for aggregating dense and attention multi-scale features
Su et al. High-Similarity-Pass Attention for Single Image Super-Resolution
Zhang et al. Short-term wind speed forecasting based on adaptive secondary decomposition and robust temporal convolutional network
CN110210412B (en) Hyperspectral image classification method based on deep learning and multi-example learning
CN104794215A (en) Fast recursive clustering method suitable for large-scale data
CN110909027A (en) Hash retrieval method
Wei et al. Joint learning method based on transformer for image retrieval

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