Abstract: The disclosure discloses a graph optimization method and apparatus for neural network computation. The graph optimization method includes the following steps: S1: converting a computation graph; S2: allocating a register; S3: defining a route selector for a redefined variable; S4: solving the route selector for the redefined variable; S5: defining a criterion of inserting the route selector for the redefined variable into a node; S6: analyzing a dominating edge set of the node for the redefined variable; S7: inserting the route selector for the redefined variable; and S8: renaming the redefined variable. The disclosure solves the problem of the corresponding route selection on a correct definition of the redefined variable when a node including the redefined variable in a computation graph in the compiling period flows through multiple paths of computation flow, reduces the memory cost and promotes the development of implementation application of a deep neural network model.

Abstract: Discloses a method and an apparatus for visual construction of a knowledge graph system. In the present disclosure, data permission of a distributed client is determined through a central server. The central server obtains a master template of a knowledge graph system and sends it to the distributed client. The distributed client receives a natural language inputted by a user and parses to generate an abstract syntax tree. The user completes customization of a subtemplate of the knowledge graph system through visual operation. The distributed client encrypts the subtemplate and then sends it to the central server. When the knowledge graph system is to be used, any knowledge concept is inputted, the central server calls and decrypts the subtemplate and then searches a database, and a tree structure knowledge graph is generated and sent to the distributed client.

Abstract: A job decomposition processing method for distributed computing, which comprises: analyzing a source program to be run by program static analysis to determine a function call graph contained in the source program; determining feature information of functions contained in the source program by program dynamics analysis or/and a program intelligent decomposition algorithm, wherein the feature information of the functions is used to characterize relevant information when each function is being running; decomposing the source program based on the feature information of the functions, a function relationship and available resource information of a computing platform to form an execution recommendation for each function on the computing platform, i.e., which hardware resources are used for computing each function; finally inserting a modifier in the source program and starting computation on the computing platform.

Abstract: Disclosed is a method and system for identifying a protein domain based on a protein three-dimensional structure image. According to the present application, the protein domain is identified based on a structure similarity, the identification errors and omissions of the protein domain caused by protein multi-sequence alignment errors when sequence consistency is not high can be effectively solved. According to the present application, the point cloud segmentation model based on the dynamic graph convolutional neural network is constructed, and by integrating global structural features and local structural features, segmentation of the protein domain and acquisition of semantic labels of the protein domain can be completed at the same time.

Abstract: A system and a method for detecting a moving target based on multi-frame point clouds. The system comprises a voxel feature extraction module; a transformer module used for matching and fusing the feature tensor sequence, fusing a first feature tensor with a second feature tensor, fusing the fused result with a third feature tensor, fusing the fused result with a fourth feature tensor, and repeating the fusing steps with a next feature tensor to obtain a final fused feature tensor; and an identification module used for extracting features from the final fused feature tensor and outputting detection information of a target. The method comprises the following steps: S1, constructing each system module; S2, training the model by the data in a training set; S3, predicting by the trained model.

Abstract: A wafer-level heterogeneous dies integration structure and method are provided. The integration structure includes a wafer substrate, a silicon interposer, heterogeneous dies, and a configuration substrate. A standard integration module is defined by the heterogeneous dies connected to the silicon interposer. The standard integration module is connected to an upper surface of the wafer substrate, and the configuration substrate is connected to a lower surface of the wafer substrate. The wafer substrate is connected to the configuration substrate via Through Silicon Vias on lower surface of the wafer substrate. And the upper surface of the wafer substrate is provided with Re-distributed Layers and a standardized micro bump array to form standard integration zone connected to the standard integration module.

Abstract: The present disclosure relates to deep learning image classification oriented to heterogeneous computing devices. According to embodiments of the present disclosure, the deep learning model can be modeled as an original directed acyclic graph, with nodes representing operators of the deep learning model and directed edges representing data transmission between the operators. Then, a new directed acyclic graph is generated by replacing the directed edges in the original directed acyclic graph with new nodes and adding two directed edges to maintain a topological structure.

Abstract: The present invention discloses a two-step anti-fraud vehicle insurance image collecting and quality testing method, system and device, the method comprises: step 1, collecting vehicle insurance scene images and marking vehicle orientation; step 2, performing object detection on the collected vehicle insurance scene images and screening to obtain object coordinates; step 3, according to the vehicle orientation and the object coordinates, obtaining the specific position of the object coordinates located in the whole vehicle; step 4, according to the object coordinates screened in step 2, performing vehicle component detection on the vehicle insurance scene images, obtaining the component coordinates of the vehicle components, and screening to obtain the vehicle component closest to the object coordinates; step 5, according to the specific position of the object coordinates located in the whole vehicle and the vehicle components closest to the object coordinates, obtaining the position of the vehicle components

Abstract: The present disclosure discloses a UDF-based traffic offloading method and a system, the method includes: step 1, performing P4 language programming: performing P4 language programming description on a protocol packet format of the target traffic in a network and a corresponding flow table description on a protocol packet format of the target packet and parsing, matching and offloading rules of the target packet; step 2: compiling and mapping the P4 program by UDF-oriented compilation and mapping method, to an ASIC chip; step 3: in the ASIC chip, matching a traffic offloading rule according to the UDF rule, and executing a traffic offloading action according to the traffic offloading rule, where the traffic offloading action includes offloading the target packet to a corresponding processing node or discarding the target packet.

Abstract: A system-on-wafer structure and a fabrication method. The structure includes a wafer substrate, an integrated chiplet, a system configuration board and a thermal module. The wafer substrate and the integrated chiplet are bonded through a wafer micro bump array and a chiplet micro bump array. The wafer substrate and the system configuration board are bonded through a copper pillar array on wafer substrate topside and a pad on system configuration board backside. A molding layer is provided between the wafer substrate and the system configuration board, and is configured to mold the wafer substrate, the integrated chiplet and the copper pillar array. Integrated chiplet are electrically connected to each other through a re-distributed layer in wafer substrate. The integrated chiplet is electrically connected to the system configuration board through the re-distributed layer and the copper pillar array. The thermal module is attached to the backside of the wafer substrate.

Abstract: Disclosed is an system for predicting end-stage renal disease complication risk based on contrastive learning, including an end-stage renal disease data preparation module, configured to extract structured data of a patient by using a hospital electronic information system and daily monitoring equipment, and process the structured data to obtain augmented structured data; and a complication risk prediction module, configured to construct a complication representation learning model and a complication risk prediction model, perform training and learning on the augmented structured data through the complication representation learning model to obtain a complication representation, and perform end-stage renal disease complication risk prediction by using the complication representation through the complication risk prediction model.

Abstract: A distributed acoustic sensing (DAS) system based on random fiber lasing amplification (RFLA) and a Rayleigh scattering enhanced fiber (RSEF), and relates to the field of distributed optical fiber sensing. The system comprises a DAS demodulation unit, a high-order RFLA unit and a RSEF. The present disclosure adopt the high-order RFLA technology to replace the traditional high-order distributed Raman amplification technology, and it does not need to use a plurality of pumps with different wavelength, and only needs a high-order random fiber laser pump and a broadband reflector to provide feedback for cascaded random fiber lasings to perform distributed amplification on signal light. By using high-order RFLA combined with the RSEF, high-efficiency and low-threshold RFLA can be achieved, and the signal-to-noise ratio and performances of a DAS system can be further improved, which enables realization of long-distance and high-performance distributed acoustic sensing.

Abstract: A geographical identification forwarding method for area-oriented addressing. The geographic location information is used as a transmission identification, and the communication process based on the geographical identification is realized by constructing the SDN-based geographical identification transmission architecture. In this method, a geographical identification is used instead of a traditional IP identification for network transmission, which effectively alleviates the problem of narrow waist of IP single bearing in the current network. At the same time, through a flow table decomposition design, the flow table size of the switch is effectively controlled. The method provided by the present invention can be extended to a plurality of geographical identification areas to realize large-area real-time cross-domain transmission.

Abstract: The present invention relates to the field of smart grids, and provides a non-intrusive load monitoring method and device based on temporal attention mechanism. The method comprises the following steps: obtaining a total load data, an equipment load data, and corresponding sampling time of a building during a certain period of time; integrating the total load data and the equipment load data with the corresponding sampling time to obtain an enhanced total load data and an enhanced equipment load data; using a sliding window method to segment the enhanced total load data and the enhanced equipment load data, and constructing a deep learning training dataset; constructing a neural network model based on a deep learning training framework and training the model using the training dataset. The present invention can effectively extract the working time mode of the load and its inherent dependencies, thereby improving the accuracy of load monitoring.

Abstract: The present disclosure discloses a system and method for testing the spatial distribution uniformity of an alkali metal atom number density of an atom magnetometer. The system includes a detection laser, a laser beam expanding system, a polarizing element, a magnetic shielding system, an alkali metal atom gas chamber, a beam profile camera, a beam splitting prism, etc., which are sequentially arranged in a light advancing direction. In the method, based on an optical absorption principle, light intensity attenuations of linearly polarized light before and after passing through the alkali metal gas chamber are tested by using the beam profile camera with pixels in the level of um, a two-dimensional distribution matrix of an atom number density in space is calculated, and the distribution uniformity of the atom number density is analyzed by using a discrete coefficient.

Abstract: The disclosure discloses a method of executing dynamic graph for neural network computation and the apparatus thereof. The method of executing dynamic graph includes the following steps: S1: constructing and distributing an operator and a tensor; S2: deducing an operator executing process by an operator interpreter; S3: constructing an instruction of a virtual machine at runtime by the operator interpreter; S4: sending the instruction to the virtual machine at runtime by the operator interpreter; S5: scheduling the instruction by the virtual machine; and S6: releasing an executed instruction by the virtual machine. According to the method of executing dynamic graph for neural network computation and the apparatus thereof provided by the disclosure, runtime is abstracted to be the virtual machine, and the virtual machine acquires a sub-graph of each step constructed by a user in real time through the interpreter and schedules, the virtual machines issues, and executes each sub-graph.

Abstract: Disclosed are an input/output proxy method and apparatus for a mimic Redis database. Through a pseudo server module, it is ensured that the interface of the Redis database is consistent with the external interface of the native Redis, so that it is convenient to implant the Redis database into arbitrary Redis application scenarios; the isolation of the modules inside is realized by independent processes, thus facilitating independent development, maintenance and expansion; and the synchronization function is integrated into the input/output proxy to achieve resource reuse; for the synchronization function, the random credit attenuation mechanism is cleverly utilized to ensure the synchronization function while taking into account the saving of resources.

Abstract: A programmable two-dimensional simultaneous multi-beam optically operated phased array receiver chip is manufactured based on silicon-on-insulator (SOI) and indium phosphide (InP) semiconductor manufacturing processes, including the SiN process. The InP-based semiconductor is used for preparing a laser array chip and a semiconductor optical amplifier array chip, the SiN is used for preparing an optical power divider, and the SOI semiconductor is used for preparing a silicon optical modulator, a germanium-silicon detector, an optical wavelength multiplexer, a true delay line, and other passive optical devices. The whole integration of the receiver chip is realized through heterogeneous integration of the InP-based chip and the SOI-based chip. Simultaneous multi-beam scanning can be realized through peripheral circuit programming control.

Abstract: The present invention discloses a preprocessing and imputing method for structural data, comprising: step 1, querying the missing information of an original data, counting missing values, and obtaining a missing rate for the original data; step 2, based on the missing rate, performing listwise deletion on the original data, and then traversing the rows to generate corresponding dichotomous arrays, converting the arrays to the form of histogram, calculating the maximum rectangular area formed by the corresponding histogram, and then sorting all rectangular areas to obtain the maximum complete information matrix; step 3, using multiple imputation by chained equations, auto-encoders, or generative adversarial imputation networks to impute missing values on the original data.

Abstract: Disclosed are a data caching method and apparatus for multiple concurrent deep learning training tasks. The method includes: step 1, executing preheating training for each task, collecting feature parameters of training batch samples, and sorting all the tasks according to the collected feature parameters; step 2, calculating the sample number of each training batch hit in a cache of each task under system pre-allocation, and the expected sample number of each training batch hit in the cache of each task; step 3, concurrently executing deep learning training by using a cache dynamic allocation and management strategy; and step 4, when each task enters a last training epoch, adding no new sample data to the caches of these tasks, gradually releasing the occupied cache, and making the released cache to be used by other tasks that are not finished.