Abstract: Techniques are provided for reinforcement learning software agents enhanced by external data. A reinforcement learning model supporting the software agent may be trained based on information obtained from one or more knowledge stores, such as online forums. The trained reinforcement learning model may be tested in an environment with limited connectivity to an external environment to meet performance criteria. The reinforcement learning software agent may be deployed with the tested and trained reinforcement learning model within an environment to autonomously perform actions to process requests.

Abstract: A system includes a neural network that includes a Mixture of Experts (MoE) subnetwork between a first neural network layer and a second neural network layer. The MoE subnetwork includes multiple expert neural networks. Each expert neural network is configured to process a first layer output generated by the first neural network layer to generate a respective expert output. The MoE subnetwork further includes a gating subsystem that selects, based on the first layer output, one or more of the expert neural networks and determine a respective weight for each selected expert neural network, provides the first layer output as input to each of the selected expert neural networks, combines the expert outputs generated by the selected expert neural networks in accordance with the weights for the selected expert neural networks to generate an MoE output, and provides the MoE output as input to the second neural network layer.

Abstract: The specification describes methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating a larger neural network from a smaller neural network. One of the described methods includes obtaining data specifying an original neural network and generating a larger neural network from the original neural network. The larger neural network has a larger neural network structure than the original neural network structure. The values of the parameters of the original neural network units and the additional neural network units are initialized so that the larger neural network generates the same outputs from the same inputs as the original neural network, and the larger neural network is trained to determine trained values of the parameters of the original neural network units and the additional neural network units from the initialized values.

Abstract: Embodiments of the present disclosure relate to splitting input data into smaller units for loading into a data buffer and neural engines in a neural processor circuit for performing neural network operations. The input data of a large size is split into slices and each slice is again split into tiles. The tile is uploaded from an external source to a data buffer inside the neural processor circuit but outside the neural engines. Each tile is again split into work units sized for storing in an input buffer circuit inside each neural engine. The input data stored in the data buffer and the input buffer circuit is reused by the neural engines to reduce re-fetching of input data. Operations of splitting the input data are performed at various components of the neural processor circuit under the management of rasterizers provided in these components.

Abstract: A method, apparatus, device and readable medium for transfer learning in machine learning are provided. The method includes: constructing a target model according to the number of classes to be achieved by a target task and a duly-trained source model; obtaining a value of a regularized loss function of the corresponding target model and a value of a cross-entropy loss function of the target model, based on sets of training data in a training dataset of the target task; according to the value of the regularized loss function and the value of the cross-entropy loss function corresponding to each set of training data, updating parameters in the target model by a gradient descent method to implement the training of the target model. The above technical solution avoids excessive constraints on parameters in the prior art, thereby refraining from damaging the training effect of the source model on the target task.

Abstract: Data is received that defines a rule mining run including a scope of a search and at least one data source to be searched. In response, the at least one data source is polled to obtain rules responsive to the rule mining run. Each rule can specify one or more actions to take as part of a computer-implemented process when certain conditions are met. A list of rules (i.e., a proposed subset of the obtained rules) can then be generated using at least one machine learning model. The generated list of rule can then be displayed in a graphical user interface. Related apparatus, systems, techniques and articles are also described.

Abstract: Techniques for optimizing a machine learning model. The techniques can include: obtaining one or more embedding vectors based on a prediction of a machine learning model; mapping the embedding vectors from a higher dimensional space to a 2D/3D space to generate one or more high density points in the 2D/3D space; clustering the high-density points by running a clustering algorithm multiple times, each time with a different set of parameters to generate one or more clusters; applying a purity metric to each cluster to generate a normalized purity score of each cluster; identifying one or more clusters with a normalized purity score lower than a threshold; and optimizing the identifying one or more clusters.

Abstract: Disclosed in the embodiments of the present invention are a method and an apparatus for training an item coding model. The method comprises: acquiring an initial item coding model and a training sample set; using sample user information of training samples in the training sample set as the input for the initial item coding model to obtain the probability of sample item coding information corresponding to the inputted sample user information; adjusting the structural parameters of the initial item coding model to train an item coding model, the item coding model being used for characterizing the correspondence between inputted sample user information and sample item coding information and the correspondence between sample item information and sample item coding information. The present embodiment can use the trained item coding model to implement item recommendation and can use the item coding information as an index to increase retrieval efficiency.

Abstract: Methods and systems for determining event probabilities and anomalous events are provided. In one implementation, a method includes: receiving source data, where the source data is configured as a plurality of events with associated timestamps; searching the source data, where the searching provides a search result including N events from the plurality of events, where N is an integer greater than one, where each event of the N events includes a plurality of field values, where at least one event of the N events can include one or more categorical field values and one or more numerical field values; and for an event of the N events, determining a probability of occurrence for each field value of the plurality of field values; and using probabilities determined for the plurality of field values, determining a probability of occurrence for the event.

Abstract: Various examples of methods and systems are provided for improved predictive analytics. In one example, a method of managing operation of an asset or group of assets of interest includes comparing a generated prediction with one or more prediction range associated with a risk profile assigned to an operational outcome of interest, presenting a notification to an operator in response to the comparison, and incorporating operator-generated input as updated source data for the generation of subsequent predictions. The operator-generated input can comprise an operator-defined selection such as, e.g., acceptance of the notification, or rejection of the notification. The operator-generated input can provide real time or near real time information based upon context-specific knowledge that the operator holds that is substantially independent of historical source data.

Abstract: A non-volatile memory device includes a memory cell array to which an arithmetic internal data is written; and an arithmetic circuitry configured to receive an arithmetic input data and the arithmetic internal data for an arithmetic operation of a neural network with the arithmetic internal data and the arithmetic input data in response to an arithmetic command, perform the arithmetic operation using the arithmetic internal data and the arithmetic input data to generate an arithmetic result data, and output the arithmetic result data of the arithmetic operation of the neural network.

Abstract: A virtual assistant negotiation method is provided, which includes the following steps: transmitting an event information by a first electronic device corresponding to an initiator; receiving a plurality of candidate projects generated by second electronic devices corresponding to a plurality of participants according to the event information by the first electronic devices; selecting a portion of the candidate projects to serve as recommended projects by the first electronic device of the host; and making a decision according to the opinions, for the recommended projects, corresponding to the main participant among the participants by the first electronic device.

Abstract: A recognition method includes retrieving an input including data of a first window size. The method further includes classifying the input based on comparison of warping distance of the input with a pruning threshold.

Abstract: Machine learning models, semantic networks, adaptive systems, artificial neural networks, convolutional neural networks, and other forms of knowledge processing systems are disclosed. An ensemble machine learning system is coupled to a graph module storing a graph structure, wherein a collection of entities and the relationships between those entities forms nodes and connection arcs between the various nodes. A hotfile module and hotfile propagation engine coordinate with the graph module or may be subsumed within the graph module, and implement the various hot file functionality generated by the machine learning systems.

Abstract: Systems and methods are provided relating to a complex adaptive command guided swarm system including an operator section comprising a first command and control section and a plurality of networked swarm of semi-autonomously agent controlled system of systems platforms (SAASoSPs). The first command and control section includes a user interface, computer system, network interface, and plurality of command and control systems executed or running on the computer system. The networked SAASoSPs each include a second command and control section, wherein the second command and control section utilizes artificial intelligence (AI) configured with a combination of both symbolic and probabilistic machine learning for various functions including pattern recognition and new pattern identification.

Abstract: Provided are a system and methodology for iteratively measuring data, as between multiple sets thereof, that accounts for underlying data generation sources and bases. Doing so, via normalization of the data, enables uniformity of interpretation and presentation of the data no matter the machine learning model that produced the data.

Abstract: Provided are an information processing device and the like that facilitate designing a neural network capable of extracting higher-order features. An information processing device includes: an extraction unit that extracts a plurality of subgraphs from a graph including a plurality of nodes and a plurality of edges; a calculation unit that calculates a distance between the plurality of subgraphs extracted; and a design unit that designs a neural network, based on the distance calculated, in that at least a part of the graph is set to an input.

Abstract: Techniques for determining embedding drift score in a machine learning model. The techniques can include: obtaining one or more first embedding vectors based on at least one first prediction of a machine learning model; filtering the first embedding vectors based on a slice of the first prediction; determining a first average vector by averaging each dimension of the filtered first embedding vectors; obtaining one or more second embedding vectors on at least one second prediction of the machine learning model; filtering the second embedding vectors based on a slice of the second prediction; generating a second average vector by averaging each dimension of the filtered second embedding vectors; and determining an embedding drift score based on a distance measure of the first average vector and the second average vector.

Abstract: Methods, computer program products, and systems are presented. The methods include, for instance: inputs of emotion time series data of a user and environmental factor data from one or more data collection device for a user assistance service. A baseline emotion time series is generated and an environmental factor that is likely to have affected changes in state of emotion on a subject is identified by regression analysis. An emotion time series model for the identified environmental factor is produced and prediction of a path to attain a target state of emotion at a certain time in the future is made. Recommendation to achieve the target state of emotion is produced based on the predicted path.

Abstract: A method of implementing a task complexity learning system, including: learning a model for predicting the value of a continuous task variable y based upon an input variable x; learning an encoder that encodes a continuous task variable y into an encoded task value; calculating a loss function based upon the predicted value of y output by the model and the encoded task value output by the encoder; calculating a distortion function based upon the input continuous task variable y and the encoded task value, wherein learning the model and learning the encoder includes minimizing an objective function based upon the loss function and the distortion function for a set of input training data including x, y pairs.