91 results about "Linear layer" patented technology

A novel and improved method of applying adhesive to one side of a running web of wrapping material for a rod-like filler of filter material for tobacco smoke includes directing toward the one side of the web a stream of adhesive which issues from an open-and-shut orifice of a nozzle, and employing one or more jets of compressed air to vary the direction of flow of the stream so that the adhesive coats one marginal portion as well as the adjacent portion of the major central part of the one side and forms at least one non-linear layer. Once the thus treated web is draped around the filler, the adhesive bonds the two marginal portions of the converted web to each other and simultaneously bonds the central part of the web to the filler.

The invention discloses a method for accelerating a GPU directed at a deep learning super-resolution technology. The method conducts concurrent processing on all the steps of a super-resolution technology which is based on deep learning and a convolutional neural network, and operates on a GPU. The concurrent processing refers to conducting concurrent task dividing on convolutions of the super-resolution technology which is based on deep learning and the convolutional neural network into millions of micro-tasks which are irrelevant to one another and can be concurrently executed in any order so as to fully exhibit the super-strong calculating capability of the GPU. Further, the method uses features of a GPU memory to cache convolution nuclear data and input image data to a shared memory and a register so as to substantially optimize calculating speeds of the convolutions. The method integrates the convolutions and non-linear layers. The method selects the optimal method for the sizes of different convolutions. According to the invention, the method accelerates the high quality super-resolution method to meet velocity requirements for processing videos, and does not cause any image quality loss.

A system and method for semantic extraction using a neural network architecture includes indexing each word in an input sentence into a dictionary and using these indices to map each word to a d-dimensional vector (the features of which are learned). Together with this, position information for a word of interest (the word to labeled) and a verb of interest (the verb that the semantic role is being predicted for) with respect to a given word are also used. These positions are integrated by employing a linear layer that is adapted to the input sentence. Several linear transformations and squashing functions are then applied to output class probabilities for semantic role labels. All the weights for the whole architecture are trained by backpropagation.

A method for secure learning of parameters of a convolution neural network, CNN, for data classification includes the implementation, by data processing of a first server, including receiving from a second server a base of already classified learning data, the learning data being homomorphically encrypted; learning in the encrypted domain, from the learning database, the parameters of a reference CNN including a non-linear layer (POLYNOMIAL) operating an at least two-degree polynomial function approximating an activation function; a batch normalization layer before each non-linear layer (POLYNOMIAL); and transmitting the learnt parameters to the second server, for decryption and use for classification.

The invention discloses a comment entity-based aspect-level emotion classification method and device and model training thereof. The model training comprises the steps of obtaining a training text comprising comment texts, different entities associated with the comment texts, aspect information and emotion information; Converting words, entities and aspects of the training text into word vector representations; combiing and representing comments in the corresponding entities and aspects based on the first interaction layer; Endowing words at different positions with different weights based onthe second position attention layer; extracting Basic words and syntactic features based on the third-layer LSTM network and the fourth-layer linear layer; And based on a fifth attention mechanism anda sixth context memory, extracting semantic features of the whole comment under the entity and aspect. The position-based attention mechanism adopted by the invention can better mine the sentiment internal relations of different words and comments under different entities and aspects, thereby obtaining a more accurate prediction result.

The invention relates to the field of deep learning and the field of image classification, in particular to a fine-grained image classification method based on an attention mechanism and cutting filling, and the method comprises the steps: constructing a convolutional neural network model; inputting an original image into a convolutional neural network model, and combining an attention mechanism to obtain an attention image; cutting a concerned image to obtain a sub-image, filling the sub-image, and performing down-sampling to obtain a filled image; inputting the attention image and the filling image into a convolutional neural network model, and obtaining probability values of corresponding categories through a linear layer and a softmax classifier; selecting a maximum probability value,and judging a classification result according to the maximum probability value; marking a classification label on the original image according to the result. The concerned image of the original imageis segmented and then subjected to filling processing, so that the correlation among all parts is destroyed, the network features local features, high-level semantics are not destroyed, and the use and training time of parameters is greatly reduced.

The invention discloses a vein recognition method based on a reinforcement learning algorithm optimized convolution neural network, which utilizes the reinforcement learning algorithm to optimize theconvolution neural network to construct a vein recognition model. The vein recognition model includes a network exploration model based on reinforcement learning algorithm and a multi-layer convolution network evaluation model. The network exploration model includes an encoder, a linear layer, a nonlinear layer, a Softmax classification layer, a decoder and a reward evaluation module. The convolution network evaluation model includes a data input preprocessing layer, multiple normal convolution module layers and multiple compression convolution module layers spaced apart, a global pooling layer and a full connection layer, and finally outputs vein eigenvector. The vein recognition method based on a reinforcement learning algorithm optimized convolution neural network can get the highest evaluation model, so as to obtain a higher recognition accuracy.

An apparatus for preparing the semi-solid molten body of iron and steel is composed of a special container for molten iron or steel, temp regulator, stirrer and bottom vibrator. The linear layer of said special container contains nucleation promoter. Its method includes pouring the molten iron or steel at the temp which is 5-10 deg.C, higher than the temp of liquid-phase line into said special container, and stirring or vibrating for 1-10 min.

The invention provides an abnormal transaction identification method and device based on financial time series characteristics, and a readable storage medium. A large amount of financial transaction flow information data to be detect which is suspected to be abnormal or relate to certain identified abnormal account numbers can be utilized, features of financial time series are extracted by neuralnetwork model adaptively, and then based on the operation of linear layer and softmax layer in a neural network, whether the trading account is pyramid selling account or not is classified and recognized. The abnormal transaction identification method provided by the invention can adaptively extracts financial time series features based on the SoftSeq2Seq-Attention neural network model, which to some extent reduces the investment of labor-intensive feature engineering. Through a single type of financial transaction flow data and fewer features, the method can achieve a good effect of abnormalfinancial account detection and recognition.

The invention discloses a method for preparing fluid with a layered density. The method for preparing the fluid with the layered density can provide a simple and effective layering technology for ocean layered environmental laboratory simulation. According to the method for preparing the fluid with the layered density, the water outlet flow of a fresh water tank and the water outlet flow of a brine tank are controlled with an automatic control method; a mixer is arranged to enable brine and fresh water to be fully mixed; a booster pump is arranged to enable the brine and the fresh water to flow into a layered water pool at a controllable speed, so that layered fluid is formed; the opening degree of an electrically-operated valve is regulated with the PID technology, the rotating speed of the booster pump is regulated through a frequency converter, a dual-regulation method is formed, the flow of the brine and the flow of the fresh water are accurately controlled and are regulated according to a preset flow change rule, and thus a density layered section containing either a jumping layering curve or a linear layering curve is obtained. According to the method for preparing the fluid with the layered density, the layering time is controllable, equipment can be arranged easily, operation is convenient, and the application flexibility is high.

The invention relates to a method for predicting critically ill death based on an attention mechanism temporal convolution network algorithm, and belongs to the field of artificial intelligence and medical treatment. The method comprises the following steps: 1, acquiring and analyzing intensive care and surgery multi-source monitoring data; 2, inputting the data into temporal convolution network (TCN) and extracting characteristics; 3, calculating attention weight of each characteristic through an attention mechanism based on the characteristics extracted from the temporal convolution network;and 4, calculating death risk coefficients through a linear layer to predict the critically ill death risk. The method disclosed by the invention has the advantages that the method is used as a critically ill death prediction tool, death of critically ill patients can be predicted so that the survival probability of the critically ill patients is increased and an effective method of a medical care plan is objectively worked out.

A system and method for semantic extraction using a neural network architecture includes indexing each word in an input sentence into a dictionary and using these indices to map each word to a d-dimensional vector (the features of which are learned). Together with this, position information for a word of interest (the word to labeled) and a verb of interest (the verb that the semantic role is being predicted for) with respect to a given word are also used. These positions are integrated by employing a linear layer that is adapted to the input sentence. Several linear transformations and squashing functions are then applied to output class probabilities for semantic role labels. All the weights for the whole architecture are trained by backpropagation.

The invention provides a self-adaptive training method for an acoustic model. The self-adaptive training method comprises the steps of: step S1, extracting speech features, and taking the speech feature as input for training and generating a seed model, so as to obtain an objective function; step S2, adjusting a network structure of the seed model, and adding a linear layer; step S3, adding a KL divergence regular term on the basis of the objective function; step S4, training the linear layer, and estimating weight and offset of a hidden linear layer by reusing a back propagation algorithm; step S5, and completing the training, and outputting a self-adaptive model. Since LHT can map scene data, and KL divergence can alleviate the over-fitting phenomenon, the self-adaptive training method can ensure that the occurrence of the over-fitting phenomenon in the neural network training process can be reduced in the case of few self-adaptive data, and the identification rate of the scene datais improved.

The invention provides a depth learning model matrix compression method and device. The last linear layer of a depth learning model is connected with M hidden nodes and N classification nodes. The last linear layer is corresponding to a weight matrix W. The method comprises the steps that S101 according to the absolute values of the elements of the weight matrix W, a K value is calculated; and Step S102 the last linear layer is divided into a first linear layer and a second linear layer, wherein the weight matrix of the first linear layer is a matrix P of M*K, and the weight matrix of the second linear layer is a matrix Q of K*N. The output of the first linear layer is the input of the second linear layer. M*N is greater than K*(M+N), and the weight matrix W is compressed.

The invention relates to a pressure-bearing tower-typed container of a plastic corrosion-proof liner, in particular to a corrosion-proof liner used for chemical equipments which have pressure requirement, belonging to the technical field of chemical equipment and industrial corrosion-proof. The pressure-bearing tower-typed container is characterized in that a pressure resistance tower internal wall and manhole, a handhole, a gas inlet, a gas outlet, a filtrate inlet, a liquid inlet and a discharge hole are internally and respectively provided with a metal net array architecture and a spiral quick-open type structure. The metal net array architecture consists of a stiffener, a metal steel plate net and a plastic liner layer; the stiffener and the metal steel plate net are welded on the pressure resistance tower body; the plastic linear layer is covered on the stiffener and the metal steel plate net so as to form a whole. The pressure-bearing tower-typed container has good corrosion-proof effect, impact resistance, stable and reliable structure, realizes the requirement of high-strength plastic lining, has simple technique and convenient construction, and overcomes the non-harmonic problem of the plastic lining layer and the metal composite physical performance for the application of the plastic lining layer in vessel products.

The present invention relates to micro imaging technology, and is especially member and method of realizing the micro imaging of Rayleigh resolution limitation in nano level. The glass slide / cover glass consists of substrate, protective inner dielectric layer, optical non-linear layer containing nano particle and protective dielectric layer, the nano particle has negative dielectric coefficient real part, and the protective dielectric layer is towards the sample. The combination of the super-resolution glass slide / cover glass and common optical microscope can reach effect similar to that of a near field optical microscope, and the glass slide and the cover glass may be used alone or simultaneously. The observed sample is set between the glass slide and the cover glass and set on the sample stage of a common optical microscope, then the spatial resolution of super-Rayleigh resolution limitation may be obtained in the observation.

The invention relates to the technical field of character recognition, and discloses a handwritten text recognition method and device and electronic equipment. The handwritten text recognition methodcomprises the steps: obtaining a to-be-recognized handwritten text image; performing high-order feature extraction on the handwritten text image by using the convolutional layer of the convolutional neural network model obtained by training to obtain feature data; performing serialization processing on the feature data through a recursion block of the convolutional neural network model, and performing serial processing in the depth direction on the serialized data obtained by the serialization processing to obtain serial serialized data; and mapping the series serialized data to an output label through a linear layer of the convolutional neural network model to obtain an output value. According to the handwritten text recognition method, the target text corresponding to the handwritten text image is acquired according to the output value and the preset semantic library, and the target text corresponding to the handwritten text image is recognized through the convolutional neural network model, and the calculation cost can be greatly reduced on the premise that the detection precision is not affected, and the training speed and the recognition speed of the model are increased.

The present invention relates to the technical field of display panels and discloses a display panel and a display device. The display panel includes a lower liner layer, a first metal trace unit, a second metal trace unit, and a third metal trace unit. The first metal trace unit is on the lower liner layer. The second metal trace unit is on the lower liner layer. The third metal trace unit is located above the first metal trace unit and the second metal trace unit and is not in contact with the first metal trace unit and the second metal trace unit, and the third metal trace unit is connectedto a fixed potential. The orthographic projections of the first metal trace unit and the second metal trace unit on the lower liner layer respectively overlap at least partially with the orthographicprojection of the third metal trace unit on the lower linear layer such that the first metal trace unit and the second metal trace unit respectively form a coupling capacitor with the third metal trace unit. In the above manner, the influence of crosstalk caused by the parasitic capacitance between the first metal trace unit and the second metal trace unit can be reduced.

The invention provides a vehicle-mounted tank made of a carbon fiber reinforced composite. The vehicle-mounted tank is formed by combining three to six tank cavities. Each tank cavity is made of an inner liner layer. A rubber air-inflation inner container serves as a die of each inner linear layer, and each inner linear layer is made through a four-dimensional enveloping and winding formation process. The tank cavities are connected with one another through bolts and bonding. Each tank cavity is externally coated with a structural layer. Each structural layer is a glass fiber reinforced thermosetting composite layer and made through a four-dimensional enveloping and winding formation technology. Reinforcing ribs are arranged at the connection positions of the tank cavities. The reinforcing ribs are made of a glass fiber reinforced thermosetting composite and perform hoop winding reinforcing. Through holes are formed among the tank cavities. Supports are arranged at the bottoms of the tank cavities. The vehicle-mounted tank has the beneficial effects that by means of the vehicle-mounted tank made of the carbon fiber reinforced composite, the problem that a metal vehicle-mounted tank is poor in corrosion resistance can be solved; the wall thickness and design weight of a tank body can be remarkably lowered; fuel consumption of a vehicle is reduced; and the vehicle-mounted tank is small in weight, high in strength, resistant to corrosion and static electricity, capable of saving energy, environmentally friendly, good in comprehensive benefit and the like.

The invention discloses a lightweight cryptographic algorithm SCENERY implementation method, a lightweight cryptographic algorithm SCENERY implementation device and a storage medium. The method comprises the steps that a plaintext to be encrypted is acquired, IP1 initial replacement, a round function, key expansion and IP2 replacement are carried out in sequence, the round function comprises sequentially carrying out round key addition operation, S box replacement and M matrix replacement on data, and the key expansion comprises sequentially carrying out S box replacement, cyclic left shift, round constant addition operation and DP dynamic replacement on a key. An F function of an SPN structure is adopted as a round function, and a binary matrix M is constructed with the purpose of achieving high dependence when an F function linear layer is designed; a round constant and a key expansion intermediate result are selected as control signals for key expansion; DP dynamic replacement is carried out on a current round key expansion intermediate result to obtain a round key, which is a new key expansion mode, the relevance of single key iteration to front wheel input is reduced, the decoding difficulty is increased, the security is improved, and differential and linear attacks and algebraic attacks can be particularly and effectively resisted.

The invention discloses a land utilization data update method based on space aggregation. The method includes reorganizing land utilization data through a space aggregation method, establishing space aggregation relations among planar layers, and generating data of linear layers corresponding to the planar layers according to space topological relations; allocating the update of land utilization data to the update of base target data based on the space aggregation relations among planar layers of the land utilization data; and updating land utilization upper-layer planar data and linear data through the space aggregation relations and the space topological relations after the update of the base target data. The land utilization data update method based on space aggregation simplifies the update of the land utilization data, guarantees the integrity and the cohesiveness of the land utilization data, updates the land utilization data rapidly, and has significant practical significances for establishing land utilization dynamic monitoring and strengthening land management.

The invention provides a Transformer-based code programming language classification method, which comprises the following steps: (1) collecting question and answer posts from Stack Overflow as a data set, and carrying out data preprocessing on data in the original data set; (2) carrying out word embedding on the data subjected to word segmentation by using the BPE to convert words into vectors; (3) on the basis of the constructed data set, performing fine tuning on the RoBERTa model, inputting a generated word vector into the RoBERTa model, performing code semantic learning through a double-layer Transform encoder, and generating a semantic representation vector Xsematic; and (4) mapping the semantic vector Xsemination to a programming language category label through a linear layer, and obtaining a corresponding programming language through a Softmax algorithm. The method has the beneficial effects that the code type can be quickly identified according to the code snippets so as to play a role in assisting developers to quickly find the solution on the question and answer website.

The invention discloses a multi-head attention memory network for short text sentiment classification. The network comprises a multi-hop memory sub-network, the multi-hop memory sub-network comprises a plurality of independent calculation modules which are connected in sequence, and each independent calculation module comprises a first multi-head attention coding layer, a first linear layer and an output layer which are connected in sequence. The input of each multi-head attention coding layer in the multi-hop memory sub-network comprises original memory and historical information memory, and the multi-head attention memory network learns more complex and abstract nonlinear features contained in a text through stacking conversion of independent calculation modules with enough hop counts; the emotion semantic structure in the text is effectively coded. Furthermore, the original memory of the input multi-hop memory sub-network is fully interacted by the recursive calculation process of the multi-head attention coding layer, so that the remote dependency relationship between the text features is modeled with more components, and the context emotion semantic relationship with higher level is mined, thereby improving the classification performance of the model.

The invention discloses a pre-trained language model quantification method and device, and the method comprises the steps of carrying out the first fine tuning of a pre-trained language model on a downstream task; clustering the data in the weight matrixes of all the other embedding layers and all the linear layers except the classification layer of the fine-tuned model by using k-means clustering, and setting the category number as 2n, where n is the bit number occupied by each piece of data of the compressed target model; and carrying out second fine tuning on the quantized model on the downstream task under the condition of maintaining quantization, and finally obtaining a quantized network. According to the scheme provided by the embodiment of the invention, the influence of the improvement of the quantization scheme of the bottom layer on the quantization effect is greatly underestimated and ignored; meanwhile, it is shown that a very good compression effect can be achieved through simple k-means quantization without any skill, and it is shown that the k-means compression method has very large development space and application prospects.

The invention discloses a voice intelligent classification method and system. The method includes: acquiring training set voice data; processing the training set voice data to obtain training set feature data; constructing an initialization model, wherein the initial classification model includes a first convolution layer, a first maximum pooling layer, a bidirectional cyclic neural network model,a second convolutional layer, a second maximum pooling layer, a first fully connected layer, a first dropout layer, a second fully connected layer, a second dropout layer, and a linear layer which are sequentially connected; training the initialization model by using the training set feature data to obtain a classification model; acquiring test set voice data; processing the test set voice data to obtain test set feature data; and classifying the test set feature data by using the classification mode. The method or system of the present invention can speed up the convergence speed in voice classification training, and can improve the judgment accuracy.

A bimetal layer manufacturing method includes the procedure of: forming a first dielectric layer on the surface of a semiconductor substrate which has a first metal layer (conductive layer) of a selected pattern formed thereon; forming a SOG layer on the surface of the first dielectric layer; forming a second dielectric layer; forming required via holes on the foregoing layers until reaching the first metal layer; forming a linear layer from a dielectrics material through PECVD; removing unnecessary linear layer from selected locations through an anisotropic plasma etching process; finally forming a second metal layer on a selected surface of the linear layer where MIM capacitors to be formed, and forming connection plugs in the via openings without generating via hole poison.

The invention provides a code searching and embedding method and device based on global information and local information. According to the method, a shared linear layer and a convolutional layer are designed to be connected behind ON-LSTM, sequence information output by a double-tower ON-LSTM model can be fused and enhanced by using shared network parameters, and utilization of global information is enhanced; structural characteristics of ON-LSTM are ingeniously applied, structural information in the ON-LSTM is selected, CNN enhancement is used, and utilization of local information is enhanced through calculation of a Manhattan distance; interaction vectors, structure vectors and global information vectors output by all the modules are spliced into feature vectors, then similarity is calculated, and global information and local information are balanced.