Abstract: Systems, methods, and software can be used to determine whether to re-label a labeled data. In some aspects, a method includes: obtaining, by an electronic device, a set of labeled data, wherein each of the labeled data comprises a feature vector and a label; for each labeled data in the set of the labeled data: processing the labeled data to obtain a plurality of classification results by using a plurality of machine learning models, wherein each of the plurality of classification results is obtained by using a different machine learning model in the plurality of machine learning models to process the feature vector of the labeled data; and determining a label uncertainty score of the labeled data based on a difference between an average entropy score and an adjustment score; and determining, whether to re-label one or more labeled data in the set of labeled data based on the label uncertainty scores.

Abstract: Content delivery optimization and recommendation is disclosed. A manner of delivering a content object to a mobile device may be determined at least in part by applying a behavior model associated with a user of the mobile device to attributes associated with the content object. The behavior model may be generated based at least in part on observed activities of the user. The content object is provided to the mobile device in the determined manner.

Abstract: Disclosed is an electronic apparatus. The electronic apparatus includes a memory storing at least one instruction, and a processor coupled to the memory and configured to control the electronic apparatus, the processor configured to identify one of a plurality of exit points included in a neural network based on at least one constraint in at least one of processing or the electronic apparatus, process the input data via the neural network and obtain processing results output from the identified exit point as output data.

Abstract: Systems and method for classifying manufacturing defects are disclosed. In one embodiment, a first data sample satisfying a first criterion is identified from a training dataset, and the first data sample is removed from the training dataset. A filtered training dataset including a second data sample is output. A first machine learning model is trained with the filtered training dataset. A second machine learning model is trained based on at least one of the first data sample or the second data sample. Product data associated with a manufactured product is received, and the second machine learning model is invoked for predicting confidence of the product data. In response to predicting the confidence of the product data, the first machine learning model is invoked for generating a classification based the product data.

Abstract: A machine learning model compression system and related techniques are described herein. The machine learning model compression system can intelligently remove certain parameters of a machine learning model, without introducing a loss in performance of the machine learning model. Various parameters of a machine learning model can be removed during compression of the machine learning model, such as one or more channels of a single-branch or multi-branch neural network, one or more branches of a multi-branch neural network, certain weights of a channel of a single-branch or multi-branch neural network, and/or other parameters. In some cases, compression is performed only on certain selected layers or branches of the machine learning model. Candidate filters from the selected layers or branches can be removed from the machine learning model in a way that preserves local features of the machine learning model.

Abstract: Embodiments herein describe optimizing a netlist using machine learning (ML) models that predict which of a plurality of optimization strategies (e.g., a plurality of optimization algorithms) will provide the best results. The netlist can then be optimized using the optimization strategy. Doing so provides significant time and compute resources savings since the netlist can be optimized only once using the selected optimization strategy rather than having to be optimized using each of the plurality of optimization strategies. Moreover, the embodiments herein can permit the addition of more optimization strategies, which could not be considered earlier because of runtime constraints.

Abstract: A method is for providing training data for training a data-based system model for operating a technical system by defining a data point determined from input variables for determining at least one output variable depending on which the technical system is operating. The method includes providing training data that are determined with a scenario other than a real operation of the technical system, the training data are defined for data points determined from the input variables, capturing operational data points determined from the input variables in real-world operation of the technical system, and splitting the training data into training data points and validation data points. The method further includes determining a k-Nearest Neighbor tree from the training data points, and determining a first distribution of distance values of distances between each of the validation data points and a predetermined number of next training data points of the training data points.

Abstract: A taxonomy classification system assigns taxonomy labels to content items of an online system. To assign the taxonomy labels, the taxonomy classification system applies one or more taxonomy model to the content items to determine scores or probabilities that a particular label applies to the content item. Each taxonomy model includes multiple sub-models. Each sub-model corresponds to a different type of information for the content item. For example, a first sub-model corresponds to a description of the content item, a second sub-model corresponds to metrics of the content item in one or more content item publishers, a third sub-model corresponds to similar content items to the content item being evaluated. The taxonomy classification system combines the output from every sub-model to determine a label score for one or more labels in a label class and a taxonomy label from the label class is selected based on the determined label score.

Abstract: Systems, methods, and devices that relate to routing requests to large language models (LLMs) are disclosed. In one example aspect, the system receives session-specific data elements in response to a request to generate an output using LLMs. The system determines a hierarchy of operational constraints including privacy protocols and performance requirements. Weights for a multi-variable optimization are dynamically updated using the session-specific data elements. The system executes the multi-variable optimization across candidate LLMs that satisfy privacy constraints and optimize performance constraints. Based on the optimization, at least one candidate LLM is selected and the request is routed to it. In response to performance feedback, the system automatically selects a different LLM to improve one constraint, resulting in degradation of another constraint.

Abstract: A method for running a multimodal generative artificial intelligence (AI) model includes transforming model parameters of a trained multimodal generative AI network from floating-point precision to selected bit-depth representations including 16-bit, 8-bit or 4-bit integers, wherein quantization comprises minimizing representation error and preserving semantic features across text, image, video, audio, or sensor modalities to reduce memory footprint and computational complexity; packaging the quantized model parameters and network architecture into a compressed deployment bundle; transmitting said bundle to one or more edge devices, wherein model compatibility and runtime configuration for heterogeneous device hardware are validated prior to installation; and conducting AI inference operations on the deployed edge devices with modular lightweight neural network layers, on-device caching of intermediate results, and batched or streaming inference, wherein inference on multimodal inputs is completed without exc

Abstract: Arrangements for intelligent orchestration of quantum programs using smart contracts linked to quantum program non-fungible tokens (NFTs) are provided. A plurality of quantum programs to be executed on target quantum hardware may be received from a digital computing device. NFTs representing each of the plurality of quantum programs may be created and linked to a corresponding quantum program. The NFTs may be stored on a distributed ledger and controlled by smart contracts storing predefined acceptance rules based on which a quantum processing output is deployed to the digital computing device. A quantum program may be validated for respective target quantum hardware by validating the NFT associated with the quantum program. The respective target quantum hardware may ingest the validated quantum program and perform quantum processing. Responsive the output of quantum processing meeting the predefined acceptance criterion, the output of the quantum processing may be deployed in the digital computing device.

Abstract: This disclosure addresses deficiencies in existing methods for analyzing spikes in time-series data, particularly when dealing with vast document repositories. A method includes receiving a user specification of objects of interest, and by subsequently identifying spikes of mentions of these objects in the documents. The method includes retrieving metric data and context data related to mentions of objects of interest in relevant documents from a repository, both from spikes and other time intervals. By analyzing these documents, it is possible to pinpoint the key factors driving the spikes. Finally, use of an LLM provides capabilities to generate comprehensive explanations. This is achieved by submitting one or more prompts to LLM(s), where the prompts incorporate the specification of the objects of interest (or a reformulation thereof), the identified driving factors and a number of representative documents connected to the key driving factors.

Abstract: Heterogenous neural networks are disclosed that have activation functions that hold multi-variable equations. These variables can be passed from one neuron to another. The neurons may be laid out in a topologically similar fashion to a physical system that the heterogenous neural network is modeling. A neural network may have inputs of more than one type. Only a portion of the inputs (a subdomain) may be optimized In such an instance, the neural network may run forward, backpropagate to all inputs, and then perform optimization only on those inputs which will be optimized.

Abstract: Systems and methods are disclosed for processing sparse tensors using a trained neural network model. An input sparse tensor may represent a sparse input point cloud. The input sparse tensor is processed using an encoder stage having a series of one or more encoder blocks, wherein each encoder block includes a sparse convolution layer, a sparse intra-channel attention module, a sparse inter-channel attention module, and a sparse residual tower module. Output from the encoder stage is processed using a decoder stage having a series of one or more decoder blocks, wherein each decoder block includes a sparse transpose convolution layer, a sparse inter-channel attention module, and a sparse residual tower module. The output of the decoder stage is an output sparse tensor representing a sparse labeled output point cloud.

Abstract: Certain aspects of the present disclosure provide techniques for performing operations with probabilistic numeric convolutional neural network, including: defining a Gaussian Process based on a mean and a covariance of input data; applying a linear operator to the Gaussian Process to generate pre-activation data; applying a nonlinear operation to the pre-activation data to form activation data; and applying a pooling operation to the activation data to generate an inference.

Abstract: A method for generating a sparsified convolutional neural network (CNN) is provided that includes training the CNN to generate coefficient values of filters of convolution layers, and performing sparsified fine tuning on the convolution layers to generate the sparsified CNN, wherein the sparsified fine tuning causes selected nonzero coefficient values of the filters to be set to zero.

Abstract: A unified system with a machine learning feature data pipeline that can be shared among various product areas or teams of an electronic platform is described. A set of features can be fetched from multiple feature sources. The set of features can be combined with browsing event data to generate combined data. The combined data can be sampled to generate sampled data. The sampled data can be presented in a format having a structure that is agnostic to a feature source from which the set of features was fetched. The sampled data can be joined with old features by a backfilling process to generate training data designed to train one or more machine learning models. Related methods, apparatuses, articles of manufacture, and computer program products are also described.

Abstract: Efficient use of channel bandwidth response, response timing, along with the ability to acquire the most accurate and up to date response are provided for management of virtual assistant search queries within a communication system. Improved management is obtained using an artificial intelligence (AI) server controlling response activity to a query communication device by adjusting content of a verbose response to the query by varying the verbosity of the response based on channel availability. The channel availability is determined based on channel bandwidth and channel occupancy.

Abstract: A computer-program product storing instructions which, when executed by a computer, cause the computer to, for one or more iterations, update parameters associated with a machine-learning network utilizing perturbations for input data, wherein the perturbations are sampled utilizing Markov chain Monte Carlo, identify a loss value associated with each perturbation in each iteration, and evaluate the machine learning network by identifying an average loss value across each iteration and outputting the average loss value.

Abstract: Analog to digital conversion errors caused by non-linearities or other sources of distortion in an analog-to-digital converter are compensated for by use of a machine learning system, such as a neural network. The machine learning system is trained based on simulation or measurement data, which may utilize a reference ADC or a digital training signal representing a reference ADC that has less distortion errors than the analog-to-digital converter. The effect on the analog to digital conversion errors by Process-Voltage-Temperature parameters may be incorporated into the training of the machine learning system.