Abstract: A circuit and method are provided for performing convolutional neural network computations for a neural network. The circuit includes a transposing buffer configured to receive actuation feature vectors along a first dimension and to output feature component vectors along a second dimension, a weight buffer configured to store kernel weight vectors along a first dimension and further configured to output kernel component vectors along a second dimension, and a systolic array configured to receive the kernel weight vectors along a first dimension and to receive the feature component vectors along a second dimension. The systolic array includes an array of multiply and accumulate (MAC) processing cells. Each processing cell is associated with an output value. The actuation feature vectors may be shifted into the transposing buffer along the first dimension and output feature component vectors may shifted out of the transposing buffer along the second dimension, providing efficient dataflow.

Abstract: A system includes: a recursively architected network of sub-networks organized into a hierarchical layers; the sub-networks including at least a parent feature node, a pool node, a parent-specific child feature (PSCF) node, and a child feature node; the parent feature node of at least one sub-network configured with a selection function actionable on at least two pool nodes connected to the parent feature node of the at least one sub-network; the pool node of the at least one sub-network configured with a selection function actionable on at least two PSCF nodes connected to the pool node of the at least one sub-network; the PSCF node of the at least one sub-network configured to activate a connected child feature node; the child feature node connectable to at least a parent feature node of a sub-network at a lower hierarchical layer.

Abstract: The performance of a neural network (NN) and/or deep neural network (DNN) can be limited by the number of operations being performed as well as management of data among the various memory components of the NN/DNN. By inserting a selected padding in the input data to align the input data in memory, data read/writes can be optimized for processing by the NN/DNN thereby enhancing the overall performance of a NN/DNN. Operatively, an operations controller/iterator can generate one or more instructions that inserts the selected padding into the data. The data padding can be calculated using various characteristics of the input data as well as the NN/DNN as well as characteristics of the cooperating memory components. Padding on the output data can be utilized to support the data alignment at the memory components and the cooperating processing units of the NN/DNN.

Abstract: Techniques for determining and utilizing optimal aggregation schedules are described are described. A deep machine learning model can be trained using multiple processing elements implemented in one or multiple computing devices and that are interconnected using one or multiple types of links. An optimal aggregation schedule for such arbitrary topologies can be determined automatically. The determination may include solving a linear program on the spanning tree polytope. The optimal aggregation schedule can be utilized by the multiple processing elements to train the deep machine learning model.

Abstract: Aspects of the present invention disclose a method, computer program product, and system for pre-training a neural network. The method extracting features of data set received from a source, the data set includes labelled data and unlabeled data. Generating a plurality of data clusters from instances of data in the data set, the data clusters are weighted according to a respective number of similar instances of labeled data and unlabeled data within a respective data cluster. Determining a data label indicating a data class that corresponds to labeled data within a data cluster of the generated plurality of data clusters. Applying the determined data label to unlabeled data within the data cluster of the generated plurality of data clusters. In response to applying the determined data label to unlabeled data within the data cluster of the generated plurality of data clusters, deploying the data cluster to a neural network.

Abstract: A computer-implemented method includes receiving a batch of neural network inputs to be processed using a neural network on a hardware circuit. The neural network has multiple layers arranged in a directed graph and each layer has a respective set of parameters. The method includes determining a partitioning of the neural network layers into a sequence of superlayers. Each superlayer is a partition of the directed graph that includes one or more layers. The method includes processing the batch of inputs using the hardware circuit, which includes, for each superlayer in the sequence: i) loading the respective set of parameters for the layers in the superlayer into memory of the hardware circuit, and ii) for each input in the batch, processing the input through each of the layers in the superlayer using the parameters in the memory of the hardware circuit to generate a superlayer output for the input.

Abstract: A method includes receiving data from a sensor over time. The data comprises a plurality of values that are each indicative of a sensed condition at a unique time. The method also includes determining a real-time value, a mid-term moving average, and a long-term moving average based on the data and determining a most-recent combined average by averaging the real-time value, the mid-term moving average, and the long-term moving average. The method further includes determining an upper setpoint by adding an offset value to the most-recent combined average and determining a lower setpoint by subtracting the offset value to the most-recent combined average. The method also includes transmitting an alert based on a determination that a most recent value of the data is either greater than the upper setpoint or lower than the lower setpoint.

Abstract: Techniques for inferring data to improve the accuracy and completeness of information retrieval are disclosed herein. In some embodiments, a data inference system detects a lack of employment type data for a profile of a user on an online service, with the employment type data identifying at least one type of employment in which the user is interested. In some embodiments, based on the detecting of the lack of employment type data for the profile of the user, the data inference system generates the employment type data based on an inference model and inference data, with the inference data comprising at least one of profile data of the user and a history of the user's interactions with the online service, and the data inference system performs a function of the online service using the generated employment type data.

Abstract: A method and apparatus for generating a training model based on feedback are provided. The method for generating a training model based on feedback, includes calculating an eigenvector of a sample among a plurality of samples; obtaining scores granted by a user for one or more of the plurality of samples in a round, obtaining scores granted by the user for a first number of samples; obtaining scores granted by the user for a second number of samples in response to detecting, based on the eigenvector, an inconsistency between the scores granted by the user for the first number of samples; and generating a training model based on the scores granted by the user for the first and second numbers of samples. A corresponding apparatus is also provided.

Abstract: Aspects described herein may provide an interface and/or search functionality for a dealership to determine vehicles a customer is most likely to purchase. A recommender system may generate vehicle recommendations for a dealership to sell to a customer based on customer information, vehicle information, and dealership information. Machine learning may be used to generate the recommendations. The recommendations may be based on the vehicle preferences of a customer.

Abstract: To provide a numerical controller and a machine learning device that predict an abnormality, based on machine learning with perception of temporal change in data. The numerical controller includes the machine learning device provided with a learning unit that conducts machine learning of trends in operation of a machine on occasions of occurrence of abnormalities in the machine, based on time-series data acquired by a data logger device and relating to the operation of the machine and abnormality information relating to the abnormalities which have occurred in the machine and a prediction unit that predicts an abnormality which will occur in the machine, based on results of the machine learning in the learning unit and time-series data acquired by the data logger device and relating to current operation of the machine.

Abstract: A system and method for performing operations on multi-dimensional functions using a machine learning model, the method including: receiving a problem formulation in input space; mapping the problem formulation from input space to one or more latent vectors or a set in latent feature space using a projection learned using the machine learning model; splitting the one or more latent vectors or set in latent space into a plurality of lower-dimensional groupings of latent features; performing one or more operations in latent space on each lower-dimensional groupings of latent features; combining each of the low-dimensional groupings; and outputting the combination for generating the prediction.

Abstract: A computer-implemented method of determining an approximated value of a parameter in a first domain is described. The parameter is dependent on one or more variables which vary in a second domain, and the parameter is determined by a function which relates sets of values of the one or more variables in the second domain to corresponding values in the first domain.

Abstract: Software that uses personalized information pertaining to a user to determine how familiar (or “novel” or “surprising”) a new artifact will be to the user, by performing the following steps: (i) receiving a first dataset pertaining to a first user; (ii) building, utilizing the first dataset, an ontology of artifacts known to the first user, where the ontology includes a domain of food and a plurality of artifacts that include food recipes, and where the artifacts have corresponding characteristics that include food ingredients; (iii) calculating a prior probability distribution for each artifact of the ontology using a probabilistic familiarity algorithm; and (iv) calculating a probabilistic familiarity value for the first artifact with respect to the first user by adding the first artifact to the set of artifacts and calculating the first artifact's prior probability distribution using the probabilistic familiarity algorithm.

Abstract: As part of neural network sensitivity analyses, base outputs of hidden layer nodes of a neural network model for non-perturbed variables can be reused when perturbing the variables. Such an arrangement greatly reduces complexity of the calculations required to generate outputs of the model. Related apparatus, systems, techniques and articles are also described.

Abstract: An intermediate representation of a workflow of one or more modules may be generated to decouple language implementations of the one or more modules. In response to receiving a workflow of one or more modules, the workflow may be analyzed to determine an optimal implementation language for each of the one or more modules to thereby reduce effects of data marshalling. An intermediate representation of the workflow that is configured to decouple any implementation languages associated with the one or more modules may be generated. To allow for decoupling, the intermediate representation may be written in a declarative language. The generated intermediate representation may then be compiled to generate an executable program that corresponds to the workflow and is implemented in the determined optimal language for each of the one or more modules.

Abstract: Anomalous sensors are detected using an apparatus including a processor and one or more computer readable mediums collectively including instructions that, when executed by the processor, cause the processor to obtain a plurality of healthy sensor data, wherein each of the healthy sensor data includes a plurality of sensed values of a corresponding sensor among a plurality of sensors in normal operation, generate a healthy data distribution of at least two sensors among the plurality of sensors based on the plurality of healthy sensor data, and generate a function of a parameter probability distribution of the plurality of sensors under a condition of sensor data of the plurality of sensors based on the healthy data distribution, each parameter indicating whether the corresponding sensor is healthy or anomalous.

Abstract: A method for autofilling an electronic form is provided. Elements of the electronic form are identified. A value for each identified elements of the electronic form is determined. The electronic form is automatically filled with the determined values. During the automatically filling of the electronic form, the determined value is provided in a field corresponding to each of the elements. A user input is received on the provided value. The received user input includes a correction to a first value provided in a first field of the electronic form. An autofill application is trained using the received user input.

Abstract: Machine learning models are powerful artificial intelligence tools that can make determinations based on a variety of factors. Unlike a simple linear model, however, determining the contribution of each variable to the outcome of a machine learning model is a challenging task. It may be unclear which factors contributed heavily toward a particular outcome of the machine learning model and which factors did not have a major effect on the outcome. Being able to accurately determine the underlying causative factors for a machine learning-based decision, however, can be important in several contexts. The present disclosure describes techniques that allow for training and use of non-linear machine learning models, while also preserving causal information for outputs of the models. Relative weight calculations for machine learning model variables can be used to accomplish this, in various embodiments.

Abstract: This invention relates to an artificial memory system and a method of continuous learning for predicting and anticipating human operator's action as response to ego-intention as well as environmental influences during machine operation. More specifically the invention relates to an architecture with artificial memory for interacting with dynamic behaviors of a tool and an operator, wherein the architecture is a first neural network having structures and mechanisms for abstraction, generalization and learning, the network implementation comprising an artificial hierarchical memory system.