515 results about "Softmax function" patented technology

The invention discloses a wind turbine generator gearbox fault diagnosis method and system based on time-frequency feature cross complementation and multi-domain feature fusion. Collected original vibration signals are divided into a training set, a verification set and a test set; respectively utilizing FFT and CWT to obtain a one-dimensional frequency domain signal and a two-dimensional time-frequency graph of the data set; inputting the training set of the time domain signal and the frequency domain signal into a time-frequency cross attention mechanism network to obtain cross complementation characteristics of the time domain and the frequency domain; inputting the training set of the two-dimensional time-frequency graph into a two-dimensional residual CNN into which a complete attention mechanism is introduced to extract time-frequency domain features; fusing the cross complementation features of the time domain and the frequency domain and the time-frequency domain features to obtain multi-domain features, and performing fault classification by using a softmax function; and inputting test set data into the trained multi-domain feature fusion network model for testing to obtain a final fault diagnosis result. According to the method, the problem of incomplete feature extraction of deep learning fault diagnosis based on a single time domain signal and a frequency domain signal is solved, and the fault diagnosis precision is improved.

The invention discloses a photovoltaic module fault prediction method and system based on deep learning, and relates to the technical field of photovoltaic power generation, and the method comprises the steps: building a battery parameter data set through a photovoltaic system simulation model, and analyzing the changes of a photovoltaic characteristic curve under different fault types; based on the change of the photovoltaic characteristic curve, constructing a fault prediction model based on Transform, and performing correlation analysis on extracted change parameters in combination with an attention mechanism; performing parameter optimization and learning rate control by adopting a U-Net decoder, a ReLU activation function and an Adam optimizer on the basis of correlation analysis of variable parameters; a SoftMax function is introduced to classify the severity of the faults; the system comprises a multi-source heterogeneous data acquisition module, a dynamic feature extraction module, a space-time double-flow Transform prediction model, an edge calculation deployment module and an online incremental learning module. The method has better adaptability and generalization ability when facing diversified fault conditions of an actual photovoltaic module, and can identify and predict faults more accurately.

The invention discloses an automatic textile color fastness rating method and system based on a CNN-Transform fusion model, and relates to the technical field of textile digital detection. And then, feature extraction and rating are carried out by using a CNN-Transform fusion model. After the features of the two are fused, weighting is carried out through an attention mechanism, classification is carried out through a full-connection layer, and the color fastness grade probability is determined through a Softmax function. Meanwhile, a cross entropy loss function is combined with a label smoothing technology, an Adam optimizer is utilized, and a learning rate attenuation strategy is matched to train the model. Finally, confidence evaluation and result correction are carried out on the rating result, automatic and high-precision textile color fastness rating is achieved, and the problems that an existing rating method is poor in accuracy, low in consistency, difficult to adapt to fine rating standards and the like are effectively solved.

The invention discloses a multi-modal modulation signal identification method based on gating attention fusion and weighted loss. The method comprises the following steps: acquiring a multi-modal modulation signal and preprocessing the multi-modal modulation signal; performing feature extraction on the multi-modal modulation signal by using a multi-branch feature extraction network; fusing the extracted features by using a gating attention mechanism; constructing an adaptive triple weighted loss function driven by deep reinforcement learning; parameters in the multi-branch feature extraction network and the gating attention mechanism are updated based on the loss function, and loop iteration is carried out until the loss convergence state reaches a preset standard; and inputting signal data to be identified into the trained network, obtaining fused features through multi-branch feature extraction and gating attention feature enhancement, and outputting a modulation signal identification result through a full-connection classification layer and a Softmax function. According to the method, the signal identification precision under the condition of low signal-to-noise ratio can be improved, and meanwhile, the debugging modes capable of being identified by the method are wide in variety.

The invention provides an HPLC (High Performance Liquid Chromatography) dual-mode communication stability method, medium and system, and belongs to the technical field of dual-channel multiplexing communication, and the method comprises the following steps: firstly, collecting HPLC low-frequency / high-frequency carrier signals, working voltage / current, temperature and electromagnetic interference data in real time; stable, variable and noise components of a carrier signal are respectively obtained through empirical mode decomposition, and a correlation coefficient matrix of the variable components and interference factors is established by using multiple linear regression and singular value decomposition. And calculating a signal intensity, a signal-to-noise ratio and an error rate evaluation matrix based on the stability coefficient matrix, and constructing an abnormal feature matrix in combination with a communication protocol specification. And finally, inputting the features into a five-layer full-connection neural network, outputting fault type probability distribution through a softmax function, and taking the type corresponding to the maximum probability value as a diagnosis result. According to the invention, the technical problem of rapid and accurate diagnosis of HPLC communication stability in a complex and changeable low-voltage distribution network environment is solved.

The invention relates to a user psychological state monitoring method and system based on voice and semantic recognition, and the method comprises the steps: firstly obtaining voice data and corresponding text data of a user, and respectively extracting a voice spectrum feature vector and a text feature vector; then, inputting the two features into a cross-modal joint coding network model, and aligning a speech spectrum with a text embedding space by using a multi-head attention mechanism to obtain a joint feature; thirdly, confidence coefficients of the joint features are extracted through a Bayesian network model, weighted features are obtained, a Gaussian mixture model is used for fitting an emotion fluctuation trend, and an emotion distribution probability is calculated; the confidence coefficient weight is dynamically adjusted according to the emotion distribution probability, new weighted features are obtained, and after time sequence alignment processing, local information is extracted by using a convolutional neural network and the features are fused; and finally, calculating a classification probability through a softmax function, and determining a psychological state classification result of the user.

The invention discloses a vector matching knowledge distillation method and system based on adaptive temperature, and the method comprises the following steps: S1, inputting an image into a given teacher model and a given student model, carrying out the convolution processing of the image through the neural networks in the teacher model and the student model, and carrying out the processing through a full-link layer, a logistic regression logit output by the model is obtained, probability distribution characteristics of logits of the student model and logits of the teacher model are extracted through a softmax function, a self-adaptive temperature mechanism is introduced after the probability distribution characteristics are extracted, and temperature distillation is carried out; s2, vector features of probability distribution of the teacher model and the student model are extracted, and cosine distance loss, cross entropy loss and KL divergence loss are introduced to calculate total loss. The probability distribution characteristics of the student model and the teacher model are fully considered, the vector characteristics of the student model and the teacher model are extracted, a knowledge distillation characteristic system is expanded, the problem of non-uniform categories caused by knowledge distillation global pooling is solved, and the problem of distillation error increase caused by excessive mismatching of single data is greatly reduced.

The invention discloses an emotion recognition method based on a multi-domain attention fusion network, and the method comprises the steps: firstly obtaining an EEG signal, carrying out the preprocessing of the EEG signal, carrying out the channel feature extraction through a space-frequency domain mixed channel attention module, and capturing a weighted attention feature; secondly, weighting attention features, and extracting space and frequency domain features of EEG signals through a CNN module; and combining the improved multi-dimensional self-attention with the BiSRU, and capturing comprehensive features of space and frequency domain features. And finally, classifying the comprehensive features through a Softmax function, and outputting an emotion recognition result. According to the method, the capability of capturing the long-term dependency relationship of the electroencephalogram signals is enhanced, and the emotion recognition performance and efficiency are effectively improved.

The invention discloses an electroencephalogram attention classification model, an electroencephalogram attention classification method and a computer system, and solves the problems of low electroencephalogram signal recognition accuracy, insufficient feature extraction and difficulty in processing complex electroencephalogram signals in the prior art. The model comprises a multi-scale separable convolutional network which performs convolution operation on electroencephalogram signals of each input channel and extracts a low-level feature map; the Squeeze-and-Excitation module is used for carrying out inter-channel dynamic adjustment on the low-level feature map so as to generate a re-weighted feature map, and the Squeeze-and-Excitation module is used for carrying out inter-channel dynamic adjustment on the low-level feature map; the lightweight Transform is used for calculating attention weights between time steps corresponding to feature vectors in the re-weighted feature map from a time dimension based on an attention mechanism so as to capture a dependency relationship between the time steps; the MobileNet is used for extracting an advanced feature map from the feature map in which the dependency relationship between the time steps is determined; and the classification module maps the feature vectors in the advanced feature map to vectors of the category number through a full connection layer, and calculates the category probability through a Softmax function to obtain an attention classification result.

The invention discloses a pre-training large model traffic flow prediction method based on double-activation domain bridging and space-time self-attention, and the method comprises the steps: designing a space-time feature multi-embedding module, carrying out the fusion of time, space and feature embedding, and constructing the multi-granularity representation of traffic flow data; designing a double-activation field bridging module, and aligning traffic flow data with the pre-training model space representation through a double-path activation mechanism; designing a dynamic gating mechanism, and adjusting the weight of the double activation branches through a softmax function; a LoRA strategy is combined with a partial attention freezing method to carry out fine tuning on the large language model; a pre-training multi-granularity space-time Transform module is designed, a space-time self-attention mechanism is used, and the modeling capability of the model for multi-granularity space-time features in traffic flow data is enhanced; and designing an output regression layer, and outputting a final prediction result by using a DADB module in combination with the convolutional layer. According to the method, the technical gap problem of the pre-training model and the traffic flow data can be effectively solved, and scientific decision support is provided for optimization of a smart city intelligent traffic system.

The invention discloses a rotating machine fault diagnosis method based on a multi-view space-time diagram attention network, and belongs to the technical field of rotating machine intelligent monitoring and fault diagnosis. A multi-view space-time diagram attention network sequentially comprises a topological graph construction layer, a two-channel parallel diagram attention network, a multi-view feature fusion module, a gating recursive unit time feature extraction module, a full connection layer and a Softmax function layer, and during training, firstly, a distance topological graph and a similarity topological graph of input data are constructed; through a two-channel parallel graph attention network, spatial feature extraction is carried out from two perspectives of geometric distance and feature similarity; through a multi-view feature fusion module, dynamic fusion of dual-channel features is realized through a learnable attention weight matrix; and finally, the fused spatial features are input into the gated loop unit network, fault feature information in the time dimension is deeply mined, and the accuracy and robustness of fault diagnosis of the rotating machine can be effectively improved.

The invention provides a fault detection method based on the combination of a Transform model of VMD and a BiGRU, and the method comprises the steps: firstly collecting voltage and current signals, and carrying out the preprocessing of the voltage and current signals; then, decomposing the preprocessed signal into a plurality of modal components by using VMD, and performing normalized coding to form a feature vector; the feature vectors are then input into a Transform model, deep features are extracted, and residual connection and layer normalization processing are carried out; the output of the Transform model is input into the BiGRU network, time sequence change characteristics are captured, and a characteristic vector with a fixed length is output; and finally, inputting the output of the BiGRU network into a softmax classifier, outputting probability distribution through linear transformation and a softmax function, and selecting a fault type with the maximum probability as a final result. And if the fault type is not a normal state, generating early warning information and giving an alarm. The method has the advantages of high-precision fault detection, strong robustness, strong adaptability and good real-time performance, and can effectively improve the accuracy and efficiency of fault diagnosis of the power system.

The invention discloses a lightweight real-time radio frequency fingerprint identification method based on streaming jump connection, and belongs to the field of communication signal processing. The implementation method comprises the following steps: adopting a WiSig data set as a training set and a test set of a neural network; and replacing a two-dimensional convolutional layer in the ResNet network with a one-dimensional convolutional layer. The ADC information sampling throughput in the edge device is greater than the reasoning throughput of the radio frequency fingerprint model, a plurality of input branches are added at different depths of the radio frequency fingerprint model, and the multi-input branch structure enables the generation time of the feature patterns needing to be fused to be consistent, thereby avoiding the distribution of the storage space. And receiver distortion features and channel noise features are removed from the same kind of radio frequency fingerprint information in different time periods within the preset time. Fusion feature fingerprint identification information is used as input of a classifier to obtain a more accurate prediction vector, a vector output by a full connection layer is mapped into a probability value through a Softmax function, a neural network is trained through a cross entropy loss function and an SGD optimizer, and radio frequency fingerprints are identified through the trained neural network.

The invention provides a social robot detection method based on a KAN graph neural network model, and the method comprises the steps: carrying out the data preprocessing of an identified social account data set, carrying out the adaptive synthesis sampling through employing an ADASYN method, and carrying out the balance processing of the data set; after the balanced data set is input into the model, graph structure data is processed through two layers of convolution, the first layer extracts preliminary features, the second layer further learns multi-level graph features, in each layer of convolution, a KAN linear layer enhances the nonlinear expression ability of the features by using B spline interpolation, and the feature extraction efficiency is improved; the model is helped to better capture various social interaction type relationships in the graph data; neighborhood between nodes is used for learning representation of the nodes, the nodes with higher representation capability can be learned through two convolutional layers for training, classification prediction is further performed through a softmax function, and a final prediction result is obtained, so that social robot detection is realized.

The invention provides an osteoporosis prediction method and system based on a centrum CT image, and the method comprises the steps: obtaining a patient centrum CT image, and carrying out the preprocessing of the image, and obtaining standardized three-dimensional voxel data; performing spatial resampling on the original data to a uniform resolution; constructing a hierarchical feature extraction network for centrum bone structure perception, and designing a non-uniform sampling mechanism for centrum density distribution; constructing an intervertebral biomechanical conduction diagram network; designing an osteoporosis specific loss function; and outputting a grading prediction result containing confidence, generating probability distribution of each grade through a softmax function, and generating a visual thermodynamic diagram of the lesion area. Through time sequence consistency constraint, the system can analyze image changes of the same patient at different time points and evaluate the treatment effect and the disease progress. The dynamic monitoring ability provides a powerful tool for long-term management of chronic osteoporosis, and is helpful for timely adjustment of treatment schemes and improvement of prognosis of patients.

According to one embodiment, a method, computer system, and computer program product for routing acceleration in mixture of experts ensembles is provided. The present invention may include receiving input data at a router; generating a plurality of output vectors by applying a routing function to the input data, wherein each output vector is associated with one or more respective tiles or pairs of tiles in a plurality of MVM tiles; determining a plurality of sub-vectors in the output vectors, wherein each sub-vector in the plurality of sub-vectors is associated with a respective output vector in the plurality of output vectors, and merging the sub-vectors into an element vector; generating a probability distribution vector by applying a Softmax function to the element vector and determining the largest elements of the probability distribution; and configuring the router based on the one or more largest elements of the probability distribution.

The invention discloses a three-dimensional positron emission tomography data image classification method based on multi-stage progressive feature extraction, and is applied to the technical field of Alzheimer's disease auxiliary diagnosis. The auxiliary diagnosis method comprises the following steps: acquiring and preprocessing PET image data of an Alzheimer's disease patient; improving the reliability of the data set by using data enhancement; performing long-range dynamic modeling on the three-dimensional voxel sequence through a stacked Lmamba block; global context semantic adaptive fusion is realized through a layer-by-layer cross-scale channel attention fusion module (CSACF), and a channel spatial perception module (CSPM) is constructed to optimize spatial feature fusion; an inverted bottleneck module is mixed with long-distance space and position information to enhance the capturing capability of the model on detail features; and finally, predicting the disease category probability through global average pooling, full connection and softmax functions. According to the method, the precision of AD early diagnosis and MCI conversion risk prediction can be greatly improved, the defects of a medical image diagnosis method of a convolutional neural network (CNN) and Transform in long-range dependence on modeling and calculation complexity are overcome, and the method has good application prospects and is suitable for AD early detection and MCI conversion risk assessment.

The invention relates to a motor imagery electroencephalogram signal classification method based on multi-scale time sequence fusion. The method comprises the following steps: S1, standardizing electroencephalogram signals before inputting a network model; s2, inputting the processed data into a multi-scale channel attention convolution module, wherein the multi-scale channel attention convolution module focuses on capturing low-level spatial-temporal characteristics with discrimination from original motor imagery electroencephalogram signals (MI-EEG); s3, segmenting the electroencephalogram signal into a plurality of local time sequence segments by sliding the window; s4, inputting the data output by the sliding window into the time fusion residual network in parallel so as to further extract advanced time features from the time sequence; and S5, fusing the high-order time sequence characteristics of all windows through a full connection layer, and outputting probability prediction of a motor imagery task through a Softmax function. The method can effectively solve the problems that convolution scale division of the motor imagery network is limited, and features of all channels and advanced time features are ignored, and the average classification accuracy of the motor imagery electroencephalogram signals is improved.

The invention discloses a model training method and device based on a reattention mechanism. In the scheme, an embedded vector sequence is input into a multi-head attention layer of a pre-training language model to extract global semantic features, and a global context vector is generated based on the global semantic features; based on a sparse Softmax function and the global context vector, a sparse gating network calculates dynamic weights corresponding to multiple attention heads, and constructs an attention weight vector; performing weighted fusion on the context vectors output by the attention heads and the corresponding attention weight vectors by a multi-head attention layer, and performing linear transformation on the fused context representation based on a fixed weight matrix to obtain a self-attention value; and adjusting parameters of the pre-training language model according to the self-attention value through a preset loss function to obtain a target pre-training language model. According to the method, the sparse gating mechanism is introduced to dynamically adjust the weight of the multiple attention heads, and the model prediction precision is remarkably improved.

The invention relates to the field of heterogeneous parallel computing, in particular to an attention operator head dimension block computing method applied to a sea light DCU, which comprises the following steps of: expanding a sequence length block computing range in charge of each thread block in an attention operator and a matrix computing range processed by a single thread working group; a blocking parameter splitqk and a blocking parameter splitv are selected, wherein the blocking parameter splitqk and the blocking parameter splitv are selected; selecting a proper block cutting parameter splitqk for the head dimensions of the query tensor Q and the key tensor K for cutting, accumulating block calculation results, and calculating S block calculation results through a softmax function to obtain P block calculation results; selecting a proper block cutting parameter splitv for the head dimensions of the value tensor V and the output tensor O for cutting, wherein a block calculation result corresponds to a corresponding block of the output tensor O; and finally, segmenting the tensor O according to the splitv, and writing each partitioning result back to the global memory. The method is suitable for a heterogeneous parallel computing system composed of a CPU and a DCU, computing resources of a DCU computing unit can be saved, performance loss caused by resource overflow is reduced, and computing efficiency is improved.

The invention belongs to the technical field of photovoltaic system fault diagnosis and prediction, and particularly relates to a photovoltaic system fault classification method and early warning system based on deep learning. The method comprises the following steps: firstly, establishing a battery parameter data set through a photovoltaic system simulation model, and analyzing photovoltaic characteristic curve changes under different fault types; secondly, constructing a fault prediction model based on Transform, performing correlation analysis on extracted change parameters in combination with an attention mechanism, and capturing a complex relationship through linear embedding and position coding; then, a U-Net decoder, a ReLU activation function and an Adam optimizer are adopted to carry out parameter optimization and learning rate control; and finally, a SoftMax function is introduced to classify the severity of the fault, so that the accuracy of fault prediction and the beforehand intervention capability are improved. The method has high fault identification precision and stability in a complex scene.

The invention discloses a strip steel surface defect detection and classification method based on a graph neural network, and the method comprises the steps: segmenting a strip steel surface image into different scale sub-blocks, and extracting the node features of each scale sub-block; constructing graph structures with different scales according to the similarity of the node features; constructing a stacked multi-scale image attention network model, taking the image structures of different scales as input for training, inputting the image structures of different scales into the trained stacked multi-scale image attention network model, performing feature extraction on nodes in the image structures of different scales, and performing feature extraction on the nodes of the image structures of different scales; aggregating and updating node features in the different scale graph structures through a graph attention mechanism, and obtaining a feature matrix according to the updated node features in the different scale graph structures; and paving the obtained feature matrix into a one-dimensional vector, inputting the one-dimensional vector into a full connection layer, carrying out nonlinear operation, outputting the probability of each category through a softmax function, and taking the category with the maximum output probability as the category of the strip steel surface defects.

Apparatus for assigning relevance scores to portions to be evaluated of a machine learning (ML) predictor including one or more attention modules, the apparatus configured to determining the relevance scores by back-propagating an initial relevance score at an output of the ML predictor or at internal portion of the ML predictor, on the basis of activations of the ML predictor, which manifest itself in an performed by the ML predictor in portions of the ML predictor, which lie upstream relative to the output or the internal portion along an activation direction of the ML predictor, and include the one or more attention modules. Additional apparatuses utilizing the proposed apparatus as well as a corresponding method and computer program are also disclosed.

The invention discloses a multi-modal dual-path fusion network structure based on data level fusion, and relates to the technical field of dual-path fusion network structures, comprising the steps of data set construction and preprocessing, selection and processing of a plurality of data sets, and image enhancement and normalization; constructing a network structure, constructing a dual-path network, pre-training on different data sets by using a mainstream backbone network, optimizing training, splicing feature layers, and performing classified training; and performing data fusion and classification, performing horizontal fusion on the double-path features, determining the weight by adopting a weighting strategy, and inputting the fused features into a full connection layer for classification according to a softmax function. According to the method, the ophthalmic disease diagnosis accuracy and model robustness are improved, the generalization ability is improved, network training is optimized, and medical intelligent diagnosis development is promoted.

The invention discloses a reservoir level prediction method based on a neural network and uncertainty perception and a medium. The method comprises the following steps: performing multi-scale characteristic decomposition on multivariable hydrological time series data; then, the macroscopic state and the microcosmic anomaly degree of system operation are subjected to combined recognition in parallel; then, constructing a hybrid expert (MoE) neural network architecture comprising a plurality of expert sub-networks and a gating network; a dynamic weight distribution mechanism of uncertainty perception is designed, and a temperature parameter of a Softmax function in a gated network is regulated and controlled in real time by using previously identified abnormality, so that a more conservative decision is forcibly made when a system enters an unknown mode; and finally, performing weighted fusion on a model prediction result, and performing end-to-end optimization on the whole system. According to the method, the problem that the generalization ability of a traditional model is insufficient when the traditional model faces a data distribution external mode is effectively solved, and the prediction robustness and precision are remarkably improved.

Techniques are disclosed for semantic moderation of conversational agents. An example system includes a memory having instructions, and a processor communicatively coupled to the memory and configured to execute the instructions. Example instructions include: scoring a received input string against a plurality of target classes to derive two or more scores, wherein each target class is associated with an independent score, and the scoring is performed without applying a softmax function to the independent scores; using the scores to generate a rule-based determination that indicates whether to pass or filter the input string; upon a determination to filter the input string, performing filtering processing on the input string; and otherwise, performing output processing on the input string.

The invention relates to the field of building structure data processing, in particular to a building structure health monitoring method based on machine learning. The method comprises the steps that a distributed federated learning architecture is adopted, local models generate synthetic sensing data by adopting a dynamic self-learning-based SMOTE algorithm, and a self-encoding neural network based on accelerated convergence in each local model is trained; the central server carries out parameter aggregation and then updates the global model; acquiring new building sensing data, performing feature dimension reduction on the new building sensing data by adopting the trained self-encoding neural network based on accelerated convergence, and then classifying the new building sensing data after feature dimension reduction by adopting a Softmax function. The existing building structure health monitoring method has the problems of low classification accuracy and high possibility of large-batch data leakage. The building structure health monitoring method based on machine learning provided by the invention is high in classification accuracy, and is not liable to cause large-batch data leakage.

The invention relates to a garbage classification multi-stage verification method based on a dynamic confidence threshold. The method comprises the following steps: acquiring image information to be identified; performing image preprocessing on the to-be-identified image information; performing garbage classification identification on the processed to-be-identified image through the trained classification model, and outputting an identification result; after an identification result output by the classification model is converted into probability distribution through a Softmax function, the identified garbage type confidence coefficient and the probability thereof are obtained; judging whether the highest confidence degree in the obtained garbage types exceeds a preset threshold value or not; if so, outputting the garbage type corresponding to the highest confidence coefficient; and if not, outputting a preset number of garbage types with the confidence ranked before and the probability of the garbage types. The result is directly output when the confidence coefficient is high, the user experience is improved, when the confidence coefficient is low, the preset number of garbage types ranked before and the probability of the garbage types are displayed, and the misjudgment risk is avoided.

The invention discloses a Parkinson's disease detection model construction method and platform based on voice features, and the method comprises the steps: carrying out the normalization processing of an encoded questionnaire result and extracted voice features, and obtaining a normalized questionnaire result and voice features; respectively converting the normalized questionnaire result and the voice features into high-dimensional representation vectors through an embedded layer; obtaining an attention score between each embedding vector and questionnaire embedding, and normalizing the score through a softmax function; according to the attention score, performing weighted fusion on the embedded vector of the questionnaire result and the embedded vector of the voice feature to obtain a fused feature vector; and inputting the obtained fusion feature vector into a risk assessment model to obtain the risk prediction of the Parkinson's disease. Through the technical scheme provided by the invention, a convenient and accurate prediction service can be provided for the Parkinson's disease patient.

An AI accelerator apparatus using in-memory compute chiplet devices. The apparatus includes one or more chiplets, each of which includes a plurality of tiles. Each tile includes a plurality of slices, a central processing unit (CPU), and a hardware dispatch device. Each slice can include a digital in-memory compute (DIMC) device configured to perform high throughput computations. In particular, the DIMC device can be configured to accelerate the computations of attention functions for transformer-based models (a.k.a. transformers) applied to machine learning applications. A single input multiple data (SIMD) device configured to further process the DIMC output and compute softmax functions for the attention functions. The chiplet can also include die-to-die (D2D) interconnects, a peripheral component interconnect express (PCIe) bus, a dynamic random access memory (DRAM) interface, and a global CPU interface to facilitate communication between the chiplets, memory and a server or host system.