39 results about "Feed forward network" patented technology

This invention belongs to the field of medical image processing technology, specifically relating to a method for abdominal multi-organ medical image segmentation based on a state-space model. Addressing the shortcomings of existing technologies such as the difficulty of CNNs in modeling long-distance dependencies, the high computational complexity of Transformers, and the large semantic gaps, incomplete segmentation contours, and easy organ encroachment issues in VM-UNet skip connections, this solution makes key improvements: It constructs an improved VM-UNet-Skip architecture, placing skip connections before downsampling to reduce the semantic gap between the small decoder features; it introduces a multi-scale global-local information aggregation module, capturing global anatomical dependencies through multi-head Mamba units and enhancing local detail representations with a convolutional gated feedforward network; and it relies on the VMamba encoder-decoder to achieve efficient feature extraction and reconstruction. The method flow includes: constructing initial feature representations through patch embedding layers, extracting multi-scale hierarchical features through the encoder, enhancing features through an information fusion module, and fusing and mapping the enhanced features to obtain the segmentation result through the decoder. This invention effectively improves the segmentation accuracy of abdominal multi-organs, solves the problems of insufficient contour integrity and organ encroachment, while reducing the number of model parameters and computational complexity, providing reliable technical support for clinical diagnosis and surgical planning.

This invention provides a flexible production line control method and system based on a large vertical domain model. The method includes: constructing a pruning index combining temporal stability and spatial similarity for the original large vertical domain model to evaluate the redundancy of the activation tensors in the intermediate layers, and performing structured pruning accordingly to obtain a sparse structure model; performing mixed-precision quantization on the pruned model: allocating quantization bits according to the sensitivity of each network layer weight to task loss, and calculating scaling factors and zeros based on the range of non-zero weights to represent sparse weights; calculating the output deviation between the quantized pruned model and the original model under the same input, and training a feedforward network with the former's output as input to fit and compensate for the output deviation; combining the quantized pruned model and the feedforward network to form a composite control model; after deployment, adding the output command of the quantized pruned model to the compensation vector during inference to generate control commands.

The application belongs to the technical field of deep learning, and discloses a time sequence anomaly detection method based on a difference component. First, the original independent samples are changed into dependent samples through permutation and combination of a small sample data set, the number of samples is expanded to obtain combined sequence samples, and then the difference component including a learnable kernel is used to generate a difference feature map of the combined sequence samples, and the difference feature map is input into a feedforward network for training. The application effectively solves the problem of unsatisfactory classification effect between complex data in small sample learning, and proposes a new data expansion method and classification component, so that the classification detection task can be efficiently realized without long time training.

This invention discloses a lightweight remote sensing small target detection method based on edge enhancement and multidimensional attention. First, multi-scale feature extraction based on stacked phantom-guided reparameterization is performed on the remote sensing image. Frequency-domain guided gated edge enhancement is employed, and global context modeling is performed on the enhanced feature map based on a multi-head self-attention mechanism. After residual connection and layer normalization, the enhanced feature map is fed into a feedforward network for nonlinear transformation, outputting globally encoded high-dimensional features. A target query based on self-attention information interaction is then performed, interacting the target query with the high-dimensional features based on cross-attention. Lightweight multidimensional attention is then used to perform channel and spatial weighting on the interacted high-dimensional features, mapping them to target category and bounding box coordinates via a prediction head, outputting the final detection result. This method addresses problems such as complex backgrounds, large target scale differences, low pixel ratio of small targets, and inherent speckle noise in synthetic aperture radar images.

This invention relates to a lightweight model design method and system for real-time personalized AI teaching. The method involves designing a lightweight large language model as the teacher model, training the teacher model using a training set, and employing structured pruning techniques to remove redundant attention heads and feedforward network layers from the teacher model, generating a streamlined student model. Knowledge distillation is used to transfer the attention patterns and hidden layer representations of the teacher model to the student model. Finally, a mixed-precision quantization method is used to compress and optimize the student model to obtain a lightweight teaching model. This reduces the magnitude of model parameters and computational complexity, making it easy to deploy and run, suitable for resource-constrained environments, and simultaneously providing efficient knowledge transfer and personalized tutoring, thereby improving teaching quality and learning efficiency.

The present application relates to the technical field of image inpainting, and particularly relates to an image inpainting method and system based on superpixel guided state space model, which comprises the following steps: obtaining a degraded image, and extracting shallow features of the degraded image through a feature embedding layer; constructing a semantic region-aware M module, which comprises a superpixel Mamba and a feature refinement feedforward network, and the superpixel Mamba is embedded with a superpixel guided selective scanning mechanism and a region-level gating mechanism; constructing a multi-level symmetric encoder-decoder architecture; inputting the shallow feature map into the multi-level symmetric encoder-decoder architecture for multi-scale feature learning; fusing the deep feature map and the shallow feature map to obtain a fused feature map; and performing feature reconstruction on the fused feature map through a reconstruction convolution layer to obtain a restored image. The present application has significant technical progress and practical application value in terms of image inpainting accuracy, computational efficiency, scene adaptability and the like.

The application discloses a multi-modal summarization method and system based on visual information fusion, relates to data processing and artificial intelligence technology, and comprises the following steps: splicing extracted visual features and text features, and inputting the spliced visual-text features into a multi-head attention module and a feedforward network module of an encoder; taking the extracted visual features and the output of the feedforward network module as inputs of a modal fusion layer of the encoder to obtain text features combined with visual information; inputting the obtained text features combined with visual information and summary text features into a decoder; taking the extracted visual features and the output of a decoder feedforward network module as inputs of a decoder modal fusion layer to obtain text features output by the decoder; inputting the output vector of the text features output by the decoder into a Softmax layer through a linear layer, and obtaining required summary text according to a probability distribution and a vocabulary table. The method disclosed by the application can improve the quality and accuracy of generated summaries.

The application is an underwater image super-resolution reconstruction method based on dynamic window and state space cooperation. The method comprises a shallow feature extraction module, a deep feature extraction module and an image reconstruction module, through the construction of a double-path feature interaction mechanism, the alternative learning of a dynamic window attention module and a visual state space module is adopted, the cooperative optimization of local detail enhancement and global context modeling is realized. A dynamic channel gate feedforward network is innovatively designed, through a channel weight dynamic adjustment mechanism guided by a global average pooling, adaptive fusion of multi-scale features is realized. The method effectively solves the problems of blur and contrast reduction of underwater images caused by scattering effect, the PSNR is 28.21dB in the x3 task of the USR-248 dataset, which is 0.35dB higher than that of SRFormer, the experimental results show that the model has excellent performance in objective evaluation index and subjective evaluation, and remarkable effect is achieved in detail recovery, which has significant technical effect and industrial application value.

The method comprises the following steps: inputting an IQ signal into three parallel 1D convolution branches to extract local and global features of the IQ signal; performing self-adaptive enhancement of the features through a multi-head attention mechanism after layer normalization; performing nonlinear transformation and feature enhancement through a feedforward network after layer normalization; performing global average pooling and global maximum pooling based on a double-path pooling strategy to output main features; performing self-adaptive modulation of the main features based on a FiLM-style fusion mechanism using auxiliary information; performing signal classification on the modulated features; and training a communication signal classification network according to a signal classification result; the system, the device and the medium are used for implementing the method; the original IQ waveform and physical layer metadata are fused by using auxiliary information to recalibrate the main features through the FiLM-style fusion mechanism, so that the cross-scene generalization capability is improved.

This application discloses a method, apparatus, equipment, and medium for equivalent modeling of wind power clusters based on Transformer-BiGRU parallel networks, relating to the field of wind power cluster modeling technology. The method includes: building a cluster simulation model based on a mathematical model of a wind turbine generator and dividing the time-series dataset; constructing a bi-branch parallel network containing Transformer and bidirectional gated recurrent units; extracting global dependency features through a multi-head self-attention mechanism and a feedforward network; capturing local bidirectional dynamic features using the forward and reverse layers of the bidirectional gated recurrent units; fusing the features and then performing fusion and dimensionality reduction through a fully connected layer to obtain a predicted sequence; using the Grey Wolf optimization algorithm to encode and iteratively optimize the network hyperparameters to obtain a target hyperparameter combination; and training the target model using the dataset based on the optimized hyperparameters and network structure to perform equivalent modeling of wind power clusters, thereby improving the efficiency of equivalent modeling of wind power clusters.

This disclosure provides a mixed-precision quantization method, apparatus, device, medium, and program product, relating to the field of deep learning technology. The method is applied to a large language model, where the feedforward network of the large language model includes a multiplicative structure configured with a first weight matrix and a second weight matrix. The method includes: determining the statistical characteristics of the multiplicative structure on each output channel of the feedforward network based on the first and second weight matrices; determining the numerical amplification risk index of the multiplicative structure on each output channel based on the statistical characteristics; and configuring a corresponding computational precision for each output channel according to the numerical amplification risk index; wherein output channels with different numerical amplification risk indices are configured with different computational precisions. This method eliminates the need for real-time scanning of input data, significantly reducing runtime overhead.

This invention discloses a cross-language compiler optimization method based on graph neural networks. The method uses a UAST module to map the abstract syntax trees (ASTs) of different programming languages ​​to a shared semantic space. The FUAST module then fuses control flow and data flow information to form a flow-enhanced AST representation. The Sub2Vec module extracts key optimization subgraphs such as control flow, function inlining, and loops, and models their local structure and global relationships. The IRGAT module uses a sparse attention mechanism after structural pruning to aggregate features from the subgraphs, generating graph-enhanced sequence vectors. Finally, the SALT module, based on a lightweight Transformer architecture, embeds IRGAT as a structure-aware component, which processes and outputs optimization decisions through a feedforward network. This invention achieves unified semantic understanding, dynamic information fusion, and key region focusing for multi-language programs, and possesses compiler platform versatility.

A graph attention network-bidirectional recurrent interpolation network (GAI-BRI)-based air pollutant concentration interpolation prediction method, a device, and a medium. The method comprises: on the basis of the extracted air pollutant concentrations, meteorological data and satellite data from multiple distributed sites within a given time range and region, constructing an air pollutant concentration map for describing pollutant conditions within the given time range and region; and inputting numerical values in the air pollutant concentration map into a GAT-BRI prediction model to obtain an air pollutant concentration interpolation prediction result for a future time period, wherein the GAT-BRI prediction model comprises a GAT that is arranged at the bottom layer and used for extracting spatial features from the air pollutant concentration map, a BRI that is arranged at the middle layer and used for extracting temporal features from the spatial features outputted by the GAT, and a feedforward network that is arranged at the top layer and used for outputting an interpolation prediction result. The method fully exploits the spatial and temporal dependencies among multiple cities, achieving high prediction accuracy.

The application discloses a lightweight visual global perception method based on RSFVIT, and relates to the technical field of visual perception models, and comprises the following steps: constructing an RSFVIT structure comprising a three-level network, and completing a training stage of the constructed RSFVIT structure based on reparameterization position coding; performing weight conversion and reparameterization on a frequency domain branch and a space domain branch to obtain a standard space domain convolution kernel; performing channel optimization on the RSFVIT structure through split mixed self-attention; introducing a learnable idle channel reparameterization feedforward network, dynamically weighting and distributing feature channels through a learnable mask vector in the training stage, and thresholding the mask vector to convert it into a binary mask in the inference stage; through split mixed self-attention, through the design of retaining full-resolution Query, and through more fine over-parameterization control on idle channels, the precision upper limit of the model is improved under the premise of not increasing the inference cost.

The application discloses a spectrum prediction method based on a dynamic correlation matrix, comprising the following steps: constructing and training a spectrum prediction network based on a Transformer and a dynamic correlation matrix, wherein the spectrum prediction network comprises an encoder module and a decoder module; the encoder module comprises four encoders, each of which comprises a dynamic-static correlation combined multi-head attention module and a feedforward network module, and each module is followed by a residual connection normalization layer; the dynamic-static correlation combined multi-head attention module receives incoming sequence data, a static correlation matrix and a dynamic correlation matrix generated by the dynamic correlation module at different time points; the decoder module comprises four decoders, each of which comprises a masked multi-head attention module, an encoder-decoder multi-head attention module and a feedforward network module, and each module is followed by a residual connection normalization layer; and the trained spectrum prediction network is used for spectrum prediction.

The application discloses a kind of route perception and space-frequency cooperation Mamba driven image super-resolution method and device, it is related to image processing field, in the application, the input feature of route enhancement space-frequency coordination module is input to local feature extraction module, obtain first intermediate feature, again in turn after first layer normalization layer and space-frequency cooperation Mamba module, again with first intermediate feature residual connection, obtain second intermediate feature, again in turn after second layer normalization layer and first gated convolution feedforward network, again with second intermediate feature residual connection, obtain third intermediate feature, again in turn after third layer normalization layer and difficulty perception dynamic routing module, again with third intermediate feature residual connection, obtain fourth intermediate feature, again in turn after fourth layer normalization layer and second gated convolution feedforward network, again with fourth intermediate feature residual connection, obtain the output feature of route enhancement space-frequency coordination module, solve the problem of weak feature representation ability, low reconstruction quality.

The application discloses an image super-resolution reconstruction method based on space-frequency dual-domain representation learning, and relates to the technical field of image super-resolution reconstruction. The method constructs a multi-scale receptive field through four-way differentiated convolution, synchronously captures coarse-grained structure and fine-grained texture information, and combines a convolution feedforward network to complete feature refinement; the method introduces a window multi-head self-attention mechanism to capture spatial long-range dependencies to optimize low-frequency global representation, and then realizes high-frequency information refinement from global to local through an efficient distillation structure; the method modulates spectral space information adaptively by means of a learnable frequency domain filter, simultaneously generates spatial attention guided by high-frequency prior, and promotes deep interaction of dual-domain features; the method realizes deep feature mapping through six recursively connected feature representation groups, and completes high-resolution image reconstruction in combination with a pixel shuffling layer. While keeping the model lightweight, the application effectively breaks through the inherent limitations of single architecture and single-domain representation, and significantly improves the comprehensive performance of super-resolution reconstruction.

This invention relates to the field of image communication technology, specifically disclosing an image semantic communication method and system based on a semantic representation recalibration network. To improve the effectiveness of intermediate semantic representations under low bit rate and complex channel conditions, this invention introduces a semantic representation recalibration network during the deep feature transformation process at the encoding end. This network constructs two semantic representation recalibration processes within the feedforward network. It extracts local context information under different receptive fields through a multi-scale local context extraction branch, and combines channel attention and spatial attention branches to adaptively enhance key features in the intermediate semantic representation and suppress redundant features, thereby improving the effective semantic carrying capacity under a limited transmission budget. This effectively enhances the effective semantic carrying capacity within a limited transmission budget under bandwidth-constrained and noisy channel conditions, alleviates the information loss problem in low-dimensional transmission representation, and ultimately achieves an effective improvement in image reconstruction quality.

The application relates to the computer technical field, in particular to a distributed reasoning method and an electronic device, the method comprises the following steps: performing parallel processing on an input sequence of a current reasoning request to obtain an initial key-value cache and an initial hidden state, then obtaining a first hidden state and expert distribution metadata, and performing matrix multiplication on the first hidden state based on the expert distribution metadata to generate a second hidden state; updating the initial key-value cache based on the second hidden state to obtain an updated key-value cache, taking the updated key-value cache as the initial key-value cache, taking the second hidden state as the initial hidden state, and re-executing the above steps until a preset reasoning end condition is reached to obtain a final reasoning result. Therefore, the problem of resource mismatch and low utilization caused by the isomorphism deployment of the attention module and the feedforward network module in the related art is solved, the hardware utilization is improved, and the reasoning cost is reduced.

The present application relates to the technical field of computer vision and deep learning, and discloses a continuous learning visual perception system based on Kolmogorov-Arnold network. The present application aims to solve the technical problem that traditional Transformer is prone to serious catastrophic forgetting when dealing with sequential incremental tasks. The system mainly comprises an image embedding module, a hybrid encoder composed of KAN-Transformer layers and a continuous learning control module. The core point is to introduce KAN representation into the feedforward network, and replace the fixed activation function with a learnable B-spline basis function. When processing incremental learning tasks, the system uses the local support characteristics of the B-spline basis function, selectively freezes the learned spline coefficients, and absorbs new task characteristics by expanding the node grid, realizing parameter-level task decoupling at the bottom. The present application is mainly used in complex and variable continuous visual perception classification scenarios, greatly reducing the forgetting rate and saving the parameter increment overhead.

The embodiment of the specification discloses a knowledge question and answer method and device based on a knowledge graph and a medium, relates to the technical field of knowledge question and answer, and the method comprises the following steps: acquiring a medical question text input by a user, extracting at least one triple from the medical question text, splicing the triple and the medical question text to determine an input sequence; based on a pre-constructed medical knowledge graph, performing double position coding on each element in the input sequence to determine the element position coding corresponding to each element, so as to determine a corresponding first combined representation, wherein the first combined representation comprises a text embedding vector, a sequence position coding embedding vector and a graph position coding embedding vector; inputting the first combined representation into a graph hybrid expert architecture large model, generating a graph expert output representation and a feedforward network expert output representation through the graph hybrid expert architecture large model, splicing the graph expert output representation and the feedforward network expert output representation, obtaining a model splicing representation, and generating a medical answer sequence.

The application discloses a medical image segmentation method combining selective edge aggregation and deep neural networks, first constructs a Transformer-based encoder, and replaces MSA and MLP in a standard Transformer block with a selective edge aggregation module and a densely connected feedforward network to realize feature fusion and complementation; then constructs an encoder and a decoder based on densely connected CNN, connects the two encoders in parallel, enables the network to interact information at multiple levels, and fuses multi-scale features from the double encoders and the low-to-high up-sampling path based on the decoder of the densely connected CNN to restore the spatial resolution of the feature map in a fine-grained and deep-level manner; finally, a loss function combining target edges and regions is designed to simultaneously optimize the encoder and the decoder with a multi-level optimization strategy, so that the network further learns more semantic information and boundary details to refine the segmentation result. The application can solve the medical image segmentation problem in a real scene.

A data processing service schedules execution of operations for shared experts for a MoE-based feed forward network (FFN) of a machine-learning model (e.g., transformer architecture) while all-to-all (A2A) operations for a set of experts are performed for a set of devices (e.g., graphic processor unit (GPU) devices). By scheduling operations of shared experts with the A2A operations, the data processing service may incorporate shared experts without having to schedule additional time and / or resources, leading to shorter processing times and increased computational efficiency.

This invention relates to a low-power edge-side sleep staging method and system based on a liquid state machine in the field of sensing and computing technology. The method includes the following steps: inputting a single-modal target inference signal into a solidified liquid state machine training model; using the single-modal target inference signal to drive the LSM reservoir of the solidified liquid state machine training model to update the internal state of neurons and generate a membrane potential state vector; based on the membrane potential state vector, generating dynamic gating coefficients for adjusting the internal recurrent feedback loop of the LSM reservoir through the feedforward network calculation of the cross-modal dynamic gating module of the solidified liquid state machine training model; and performing weighted adjustment based on the dynamic gating coefficients and the recurrent feedback current of the LSM reservoir, outputting the current sleep staging result through the readout layer of the solidified liquid state machine training model. This solves the problem of high computational load and high power consumption when performing signal processing on resource-constrained edge devices in existing sleep monitoring.

This application provides a method and apparatus for disease identification in livestock farms based on multimodal data fusion, relating to the field of image recognition. The method includes: acquiring RGB image data and infrared image data from the livestock farm; extracting visual features from the RGB image data and infrared features from the infrared image data using an improved Transformer model; the improved Transformer model refers to a Transformer model used for image feature extraction after removing the Dropout layer, including a multi-head self-attention layer and a feedforward network layer; using a multi-scale fusion strategy of wavelet transform and attention mechanism, fusing visual features and infrared image features stepwise with fine-grained features and high-level semantic features to obtain fused features; based on the fused features, using a disease identification model, outputting the predicted probability of each disease category, and determining the predicted disease category. This application, used in the process of disease identification in livestock farms, solves the technical problem that existing technologies cannot meet the needs of efficient and rapid disease identification in large-scale livestock farms.

The application relates to the technical field of load prediction, and discloses a pre-training time sequence prediction method for building short-term load prediction, which comprises the following steps: obtaining historical energy consumption time sequence data of multiple buildings, standardizing and pretreating the historical energy consumption time sequence data, cutting the historical energy consumption time sequence data into input sequences and target prediction sequences, and constructing a Building-MoE model based on a Transformer architecture, wherein a feedforward network layer of the Building-MoE model adopts a sparse expert hybrid MoE module, and the MoE module comprises a plurality of parallel expert sub-networks and a router. The application supports quick adaptation of a target building under the condition of a small amount of historical data, also supports direct migration to an unseen building or new climate conditions without retraining, can effectively reduce the data collection and building-by-building training cost in engineering deployment, and improves the generalizability and practicality of the model in a cross-building scenario.

The application discloses an MTLA-Transformer hardware accelerator based on FPGA and relates to the technical field of FPGA acceleration and deep learning inference optimization. The accelerator comprises a controller module, an attention calculation module and a feedforward network module. The controller module is used for completing data scheduling of on-chip cache and off-chip memory and KV cache update management. The attention calculation module comprises a reusable matrix multiplication and addition calculation array, a position coding submodule, a HyperNet calculation submodule, a KV cache management submodule, a Softmax submodule and a residual normalization submodule. Projection calculation, fractional calculation and output projection and other matrix multiplication and addition operation multiplexing are realized through a systolic array. Position coding rotation factor precalculation lookup table, HyperNet position correlation result offline precalculation and Softmax scaling coefficient fusion are adopted to reduce online calculation amount and memory access overhead. The feedforward network module completes two-layer linear transformation and activation operation and realizes residual connection normalization processing. The application can improve incremental inference throughput and reduce hardware resource overhead under the premise of ensuring inference correctness and is suitable for edge end low-power real-time inference scenarios.