Abstract: Described herein are systems and methods for providing a natural language comprehension system that employs a two-stage process for machine comprehension of text. The first stage indicates words in one or more text passages that potentially answer a question. The first stage outputs a set of candidate answers for the question, along with a first probability of correctness for each candidate answer. The second stage forms one or more hypotheses by inserting each candidate answer into the question and determines whether a sematic relationship exists between each hypothesis and each sentence in the text. The second processing circuitry generates a second probability of correctness for each candidate answer and combines the first probability with the second probability to produce a score that is used to rank the candidate answers. The candidate answer with the highest score is selected as a predicted answer.

Abstract: Methods and systems for compressing a neural network (NN) which performs an inference task and for performing computations of a Kronecker layer of a Kronecker NN are described. Data samples are obtained from a training dataset. The input data of the data samples are inputted into a trained NN to generate NN predictions for the input data. Further, the input data are inputted into a Kronecker NN to generate Kronecker NN predictions for the input data. Two losses are computed: a knowledge distillation loss, based on outputs generated by a layer of the NN and a corresponding Kronecker layer of the Kronecker NN and a loss for Kronecker layer, based on the Kronecker NN predictions and ground-truth labels of the data samples. The two losses are combined into a total loss, which is propagated through the Kronecker NN to adjust values of learnable parameters of the Kronecker NN.

Abstract: A method and an apparatus for generating a shared encoder are provided, which belongs to a field of computer technology and deep learning. The method includes: sending by a master node a shared encoder training instruction to child nodes, so that each child node obtains training samples based on a type of a target shared encoder included in the training instruction; sending an initial parameter set of the target shared encoder to be trained to each child node after obtaining a confirmation message returned by each child node; obtaining an updated parameter set of the target shared encoder returned by each child node; determining a target parameter set corresponding to the target shared encoder based on a first preset rule and the updated parameter set of the target shared encoder returned by each child node.

Abstract: A method for integrating a machine learning (ML) model that impacts different factor groups for generating a dynamic recommendation to collectively optimize a parameter is provided. The method includes (i) processing a specification information and operational data associated with a demand management service obtained from client devices (116A-N), (ii) training the ML models with processed specification information and the operational data to obtain a trained ML model that includes an anticipation ML model that optimizes demand parameter or recommendation ML model that generates recommendation for optimizing a factor group, (iii) integrating the trained ML model with the ML models by setting an output of a first ML model as a feature of a second ML model and (iv) determining a demand of a product using the trained ML models and quantifying probabilistic values that signify prediction of the demand.

Abstract: Apparatus and methods for refactoring infrastructure. The methods may include (a) defining parameters of an application landscape. The methods may include (b) stress-testing an application in a simulated environment based on: the parameters; and a simulated input to the application. The methods may include (c) identifying a state of stress of the application based on output of the stress-test. The methods may include (d) repeating (b)-(c) with a different simulated input until the state of stress satisfies a predetermined stochastic threshold. The methods may include (e) providing the state of stress to an upside down reinforcement learning (“UDRL”) engine. The methods may include (f) comparing a throughput corresponding to the state of stress to a benchmark throughput. The methods may include (g) redefining the parameters. The methods may include (h) repeating (a)-(f) until a threshold proximity to the benchmark throughput is reached.

Abstract: A data processing system receives a graph that includes a sequence of layers and executes graph cuts between a preceding layer in the graph and a succeeding layer in the graph that succeeds the preceding layer. The preceding layer generates a set of tiles on a tile-by-tile basis and the succeeding layer processes a tensor that includes multiple tiles in the set of tiles. Thus the graph is partitioned into a sequence of subgraphs, and a subgraph in the sequence of subgraphs including a sub-sequence of layers in the sequence of layers. One or more configuration files is generated to configure runtime logic to execute the sequence of subgraphs and the one or more configuration files are stored on a computer-readable media.

Abstract: The present disclosure provides a joint training method and apparatus for models, a device and a storage medium. The method may include: training a first-party model to be trained using a first sample quantity of first-party training samples to obtain first-party feature gradient information; acquiring second-party feature gradient information and second sample quantity information from a second party, where the second-party feature gradient information is obtained by training, by the second party, a second-party model to be trained using a second sample quantity of second-party training samples; and determining model joint gradient information according to the first-party feature gradient information, the second-party feature gradient information, first sample quantity information and the second sample quantity information, and updating the first-party model and the second-party model according to the model joint gradient information.

Abstract: A system and method for designing a physical system using a genetic algorithm includes building a plurality of data structures necessary to build, heal, and verify a plurality of dependency chains; ensuring that multiple dependencies in a respective one of the plurality of dependency chains are represented correctly; removing any dependencies that will be trivially satisfied at random; in response to determining that one or more dependencies is consistent with another dependency, considering one or more combinations of dependencies; and building configurations that satisfy the dependencies and combinations of dependencies by associating the dependencies and combinations of dependencies with selected technology options and recursively specifying and/or revising additional technology options that are consistent with the dependencies or combinations of dependencies, until a configuration is fully specified.

Abstract: Use a computerized trained graph neural network model to classify an input instance with a predicted label. With a computerized graph neural network interpretation module, compute a gradient-based saliency matrix based on the input instance and the predicted label, by taking a partial derivative of class prediction with respect to an adjacency matrix of the model. With a computerized user interface, obtain user input responsive to the gradient-based saliency matrix. Optionally, modify the trained graph neural network model based on the user input; and re-classify the input instance with a new predicted label based on the modified trained graph neural network model.

Abstract: A method optimizes machine learning systems. A computing device accesses a committee of classifiers that have been trained using an initial labeled instance of data from an annotator. The initial labeled instance of data includes annotator-ranked attributes of the data, initial values of the attributes, and an initial prediction label that describes an initial predicted state based on the values. The computing system compares the attributes ranking from the annotator to attributes rankings that are generated by and used by each of the machine learning systems when evaluating one or more instances of unlabeled data that include the attributes, and weights the machine learning systems according to how closely each of the attributes rankings generated by and used by each of the machine learning systems match the attributes ranking from the annotator. The machine learning systems are then optimized based on this matching.

Abstract: Provided are architectures, system, methods, and computer program products that provide a user with the ability to define an association of data and/or information from known reference sets perceived by the user as relevant to a subject matter domain, thereby imparting and formalizing some of the user's knowledge about the domain. An associative relevancy knowledge profiler may also allow a user to create a profile by modifying or restricting the known reference sets and windowing the results from the association as a user might refine any other analysis algorithms. An associative relevancy knowledge profiler may also be used to define a user profile used by the user and others. A user profile may be usable in various manners depending upon, for example, rights management permissions and restrictions for a user.

Abstract: A method using the Sifr optimizer for training a neural network model having layers and parameters comprises providing an input corresponding to each of samples comprised in a batch from a training dataset to an input layer, obtaining outputs from the neural network model, calculating a loss function for each of the samples based on the outputs and corresponding desired values, and determining values of the parameters for minimizing a mismatch between the outputs and the corresponding desired values across the samples for the parameters based on the loss function. Further, the determining of the values for the parameters comprises executing at least one of forward passes and backward passes through the neural network model, obtaining a curvature data based on the executing, obtaining a Sifr update based on the data. The determining of the values for the parameters is based on the Sifr update.

Abstract: According to one embodiment, a learning apparatus includes a setting unit, a training unit, and a display. The setting unit sets one or more second training conditions based on a first training condition relating to a first trained model. The training unit trains one or more neural networks in accordance with the one or more second training conditions and generates one or more second trained models which execute a task identical to a task executed by the first trained model. The display displays a graph showing an inference performance and calculation cost of each of the one or more second trained models.

Abstract: Techniques for responding to a healthcare inquiry from a user are disclosed. In one particular embodiment, the techniques may be realized as a method for responding to a healthcare inquiry from a user, according to a set of instructions stored on a memory of a computing device and executed by a processor of the computing device, the method comprising the steps of: classifying an intent of the user based on the healthcare inquiry; instantiating a conversational engine based on the intent; eliciting, by the conversational engine, information from the user; and presenting one or more medical recommendations to the user based at least in part on the information.

Abstract: A treatment model trained to compute an estimated treatment variable value for each observation vector of a plurality of observation vectors is executed. Each observation vector includes covariate variable values, a treatment variable value, and an outcome variable value. An outcome model trained to compute an estimated outcome value for each observation vector using the treatment variable value for each observation vector is executed. A standard error value associated with the outcome model is computed using a first variance value computed using the treatment variable value of the plurality of observation vectors, using a second variance value computed using the treatment variable value and the estimated treatment variable value of the plurality of observation vectors, and using a third variance value computed using the estimated outcome value of the plurality of observation vectors. The standard error value is output.

Abstract: The present disclosure provides systems and methods for distributed training of machine learning models. In one example, a computer-implemented method is provided for training machine-learned models. The method includes obtaining, by one or more computing devices, a plurality of regions based at least in part on temporal availability of user devices; selecting a plurality of available user devices within a region; and providing a current version of a machine-learned model associated with the region to the plurality of selected user devices within the region. The method includes obtaining, from the plurality of selected user devices, updated machine-learned model data generated by the plurality of selected user devices through training of the current version of the machine-learned model associated with the region using data local to each of the plurality of selected user devices and generating an updated machine-learned model associated with the region based on the updated machine-learned model data.

Abstract: A method, apparatus, device, and storage medium for expanding data are disclosed. The method includes: acquiring a triplet from a knowledge graph; mining a relationship path equivalent to a relationship in the triplet from the knowledge graph, a subject in the triplet being used as a start point of the relationship path, and an object in the triplet being used as an end point of the relationship path; and expanding the triplet based on the relationship path to generate an expanded triplet. This implementation expands the triplet in the knowledge graph, and strengthens the association between the subject and the object in the triplet in a larger context, such that the association between the subject and the object in the triplet is more global.

Abstract: Techniques are disclosed for controlling a device's operation based on an inferred state. More specifically, at each of a set of time points, execution of an application at an electronic device is detected. For each detected execution, an application-usage variable is determined. One or more aggregated metrics are generated based on aggregation of at least some of the application-usage variables. Based on the one or more aggregated metrics, a state identifier is identified that corresponds to an inferred state of a user of the electronic device. A device-operation identifier is retrieved that is associated with the state identifier. A device operation is performed associated with the device-operation identifier.

Abstract: There is a need for more accurate and more efficient predictive data analysis steps/operations. This need can be addressed by, for example, techniques for efficient predictive data analysis steps/operations. In one example, a method includes mapping a primary event having a primary event code to a related subset of a plurality of candidate secondary events by at least processing one or more lifecycle-related attributes for the primary event code using a lifecycle inference machine learning model to detect an inferred lifecycle for the primary event.

Abstract: Mechanisms for performing intelligent federated machine learning (ML) model updates are provided. A plurality of ML model updates, and a plurality of dataset sketch commitment data structures (sketches), are received from a plurality of participant computing systems. Each sketch provides statistical characteristics of a corresponding local dataset used by a corresponding participant to train a local ML model. A potentially malicious participant identification operation is performed based on an analysis of the plurality of sketches to identify one or more potentially malicious participants based on differences in sketches. ML model updates received from participant computing systems identified as potentially malicious participants are discarded to thereby generate a modified set of updates. The federated ML computer model is updated based on the modified set of updates.