Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a neural network using normalized target outputs. One of the methods includes updating current values of the normalization parameters to account for the target output for the training item; determining a normalized target output for the training item by normalizing the target output for the training item in accordance with the updated normalization parameter values; processing the training item using the neural network to generate a normalized output for the training item in accordance with current values of main parameters of the neural network; determining an error for the training item using the normalized target output and the normalized output; and using the error to adjust the current values of the main parameters of the neural network.

Abstract: Techniques are described herein that are capable of performing intent-based scheduling via a digital personal assistant. For instance, an intent of user(s) to perform an action (a.k.a. activity) may be used to schedule time (e.g., on a calendar of at least one of the user(s)) in which the action is to be performed. Examples of performing an action include but are not limited to having a meeting, working on a project, participating in a social event, exercising, and reading.

Abstract: A hardware-based programmable deep learning processor (DLP) is proposed, wherein the DLP comprises with a plurality of accelerators dedicated for deep learning processing. Specifically, the DLP includes a plurality of tensor engines configured to perform operations for pattern recognition and classification based on a neural network. Each tensor engine includes one or more matrix multiplier (MatrixMul) engines each configured to perform a plurality of dense and/or sparse vector-matrix and matrix-matrix multiplication operations, one or more convolutional network (ConvNet) engines each configured to perform a plurality of efficient convolution operations on sparse or dense matrices, one or more vector floating point units (VectorFPUs) each configured to perform floating point vector operations, and a data engine configured to retrieve and store multi-dimensional data to both on-chip and external memories.

Abstract: A machine learning device, which performs a task using a plurality of industrial machines and learns task sharing for the plurality of industrial machines, includes a state variable observation unit which observes state variables of the plurality of industrial machines; and a learning unit which learns task sharing for the plurality of industrial machines, on the basis of the state variables observed by the state variable observation unit.

Abstract: A method, system and computer readable medium for generating a cognitive insight comprising: receiving training data, the training data being based upon interactions between a user and a cognitive learning and inference system; performing a plurality of machine learning operations on the training data; generating a cognitive profile based upon the information generated by performing the plurality of machine learning operations; and, generating a cognitive insight based upon the profile generated using the plurality of machine learning operations.

Abstract: A method, system and computer readable medium for generating a cognitive insight comprising: receiving training data, the training data being based upon interactions between a user and a cognitive learning and inference system; performing a hierarchical topic machine learning operation on the training data; generating a cognitive profile based upon the information generated by performing the hierarchical topic machine learning operation; and, generating a cognitive insight based upon the cognitive profile generated using the hierarchical topic machine learning operation.

Abstract: A method of determining the effect that changes in input variables have on changes in the output of a time series model, between two instances of time, produces variable attributions that satisfy the Shapley fairness properties of efficiency, symmetry, linearity, and null player.

Abstract: Systems and methods are described herein. In one embodiment, the method includes receiving a goal associated with a predicate-object pair; receiving utilization data including a plurality of predicate-object pairs including the predicate-object pair associated with the goal; determining a prediction model comprising a plurality of nodes that form a hierarchical structure including a root node and two or more leaf nodes and organized based on one or more of an information gain and a business gain, the two or more leaf nodes including a leaf node associated with the predicate-object pair of the goal; identifying nodes in the hierarchical structure that trace a path from the root node to the node associated with the goal; and causing a recommendation for at least partial completion of the goal to be presented to a user, the recommendation based on the one or more nodes that trace the path.

Abstract: The present invention relates to artificial neural networks, for example, deep neural networks. In particular, the present invention relates to a compression method considering load balance for deep neural networks and the device thereof. More specifically, the present invention relates to how to compress dense neural networks into sparse neural networks in an efficient way so as to improve utilization of resources of the hardware platform.

Abstract: A method and a system are provided for predicting a non-linear relationship between a plurality of parameters in a deep neural network framework. The method comprises receiving, by an application server, a plurality of parameter values associated with the plurality of parameters. The method further comprises selecting, by the application server, an activation function based on a desired output. In an embodiment, the desired output is based on an industry type and an application area of the plurality of parameters. The method further comprises predicting, by the application server, the non-linear relationship between the plurality of parameters by modelling the deep neural network framework based on the selected activation function.

Abstract: The Automata Processor Workbench (AP Workbench) is an application for creating and editing designs of AP networks (e.g., one or more portions of the state machine engine, one or more portions of the FSM lattice, or the like) based on, for example, an Automata Network Markup Language (ANML). For instance, the application may include a tangible, non-transitory computer-readable medium configured to store instructions executable by a processor of an electronic device, wherein the instructions include instructions to represent an automata network as a graph.

Abstract: In many environments, rules are trained on historical data to predict an outcome likely to be associated with new data. Described is a ruleset which predicts the probability of a particular outcome. Roughly described, an individual identifies a ruleset, where each of the rules has a plurality of conditions and also indicates a rule-level probability of a predetermined classification. The conditions indicate a relationship (e.g. ‘<’ or ‘!<’) between an input feature and a corresponding value. The rules are evaluated against input data to derive a certainty for each condition, and aggregated to a rule-level certainty. The rule probabilities are combined using the rule-level certainty values to derive a probability output for the ruleset, which can be used to provide a basis for decisions. In an embodiment, the per-condition certainty values are fuzzy values aggregated by fuzzy logic. A novel genetic algorithm can be used to derive the ruleset.

Abstract: Systems, methods, and non-transitory computer-readable media can collect past user information and churn data for a plurality of users. A churn prediction model is trained using the past user information and churn data. A churn propensity score is calculated for a present user based on the churn prediction model, the churn propensity score indicative of the likelihood of the present user to churn.

Abstract: A system may include a decentralized communication network and multiple processing devices on the network. Each processing device may have an artificial intelligence (AI) chip, the device may be configured to generate an AI model, determine the performance value of the AI model on the AI chip, receive a chain from the network where the chain contains a performance measure. If the performance value of the AI model is better than the performance measure, then the processing device may broadcast the AI model to the network for verification. If the AI model is verified by the network, the device may update the chain with the performance value so that the chain can be shared by the multiple processing devices on the network. Any processing device on the network may also verify an AI model broadcasted by any other device. Methods for generating the AI model are also provided.

Abstract: An autonomous agent maintains and updates an underlying model in a dynamic system. The autonomous agent receives minimum acceptable criteria from secondary users of the underlying model. The agent then compares current output samples of the model with the minimum acceptable criteria. If output samples do not meet the minimum acceptable criteria, then the agent formulates alternative model improvement actions and evaluates each alternative action by modeling rewards associated with it. The agent executes the alternative model improvement action having the highest reward.

Abstract: Training a target neural network comprises providing a first batch of samples of a given class to respective instances of a generative neural network, each instance providing a variant of the sample in accordance with the parameters of the generative network. Each variant produced by the generative network is compared with another sample of the class to provide a first loss function for the generative network. A second batch of samples is provided to the target neural network, at least some of the samples comprising variants produced by the generative network. A second loss function is determined for the target neural network by comparing outputs of instances of the target neural network to one or more targets for the neural network. The parameters for the target neural network are updated using the second loss function and the parameters for the generative network are updated using the first and second loss functions.

Abstract: An example method described herein involves receiving a data input; identifying a plurality of topics in the data input; determining an underrepresented set of data for a first set of topics of the plurality of topics based on a plurality of knowledge graphs associated with the first set of topics; calculating a score for each topic of the first set of topics based on a representative learning technique; determining that the score for a first topic of the first set of topics satisfies a threshold score; selecting a topic specific knowledge graph based on the first topic; identifying representative objects that are similar to objects of the data input based on the topic specific knowledge graph; generating representation data that is similar to the data input based on the representative objects to balance the underrepresented set of data with a set of data associated with a second set of topics of the plurality of topics; and performing an action associated with the representation data.

Abstract: A system may include a decentralized communication network and multiple processing devices on the network. Each processing device may have an artificial intelligence (AI) chip, the device may be configured to generate an AI model, determine the performance value of the AI model on the AI chip, receive a chain from the network where the chain contains a performance measure. If the performance value of the AI model is better than the performance measure, then the processing device may broadcast the AI model to the network for verification. If the AI model is verified by the network, the device may update the chain with the performance value so that the chain can be shared by the multiple processing devices on the network. Any processing device on the network may also verify an AI model broadcasted by any other device. Methods for generating the AI model are also provided.

Abstract: Application tool recommendations are described. Initially, application usage data is captured indicating tools used and actions performed by existing users of an application. This application usage data is converted into human-readable words describing the tools used and actions performed. This allows natural language processing techniques to be applied to the converted data. Through natural language processing, importance scores for the tools and actions can be computed and tasks performed with the application determined. The natural language processing techniques are also used to build task prediction models based on the importance scores and determined tasks. These task prediction models indicate probabilities of the determined tasks to be next performed by a current application user. A task having the highest probability of being next performed is predicted as the next task. Tool recommendations associated with the predicted next task are then presented to aid the user with the predicted next task.

Abstract: Methods, systems, and apparatus, including computer programs encoded on computer storage media, for reinforcement learning. One of the methods includes selecting an action to be performed by the agent using both a slow updating recurrent neural network and a fast updating recurrent neural network that receives a fast updating input that includes the hidden state of the slow updating recurrent neural network.