Nec Corporation

Nec C Orporation

Nec Laboratories Europe Gmbh

Mathias Niepert

Carolin Lawrence

Alberto Garcia Duran

Bhushan Kotnis

Felipe Huici

Mohamed Ahmed

Daniel Onoro-Rubio

Daniel Onoro Rubio

Roberto Gonzalez Sanchez

Maja Schwarz

Konstantin Kutzkov

Nicolas Weber

Lili Jiang

Florian Schmidt

Brandon Malone

Cheng Wang

A method of encoding a query graph having m root nodes and n leaf nodes includes receiving m×n path queries representing the query graph, where each path query starts with a root node and ends with a leaf node, and where the query graph includes one or more missing nodes or edges between nodes. Positions of each node and edge within each path query are encoded independently, where the encoded positions include a positional order, and positional counter values that are reset at a start position of each path query. The missing nodes or edges are masked to produce a masked query that is fed to a transformer encoder. The method can be used for, but is not limited to, medical uses, such as for predicting protein-drug interactions or medical conditions, as well as other applications, to support decision making or optimizing processes or predictions.

Answering complex queries in knowledge graphs with bidirectional sequence encoders usable to support decision making and having medical use applications

A memory for storing data for access by an application program includes an encoded sequential query representing m×n path queries of a query graph having elements including m root nodes and n leaf nodes, and a missing element, wherein m and n are integer values. Each of the path queries includes a root node as a start position and a leaf node as an end position, wherein positional encodings of the elements include a positional order within each path query, wherein the positional encodings of the elements include counter values that are reset at a start position of each of the path queries, and wherein the missing element is masked. The memory can be used for, but is not limited to, medical uses, such as for predicting protein-drug interactions or medical conditions, as well as other applications, to support decision making or optimizing processes or predictions.

A method for encoding a query graph into a sequence representation includes receiving m×n path queries representing a query graph having m root nodes and n leaf nodes, each path query beginning with a root node and ending with a leaf node, and encoding positions of each node and each edge between nodes within each path query independently, wherein the encoded positions include a positional order within each path query. The positional encodings may include no positional ordering between different path queries. The query graph may include one or more missing entities and each missing entity may be masked to produce a masked sequential query, which may be fed to a transformer encoder. The invention can be used for, but is not limited to, medical uses, such as for predicting protein-drug interactions or medical conditions, as well as other applications, and significantly outperforms the optimized model.

Answering complex queries in knowledge graphs with bidirectional sequence encoders useable to support decision making and having medical use applications

Method for encoding a query graph into a sequence representation includes receiving a set of m&#xd7;n path queries representing a query graph, the query graph having a plurality of nodes including m root nodes and n leaf nodes, wherein m and n are integer values, each of the m&#xd7;n path queries beginning with a root node and ending with a leaf node, and encoding positions of each node and each edge between nodes within each of the m&#xd7;n path queries independently, wherein the encoded positions include a positional order within each path query. The positional encodings may include no positional ordering between different path queries. The query graph may include one or more missing entities and each token representing one of the one or more missing entities may be masked to produce a masked sequential query, which may be fed to a transformer encoder.

Answering complex queries in knowledge graphs with bidirectional sequence encoders

A computer-implemented method includes instantiating a neural network including a recurrent cell. The recurrent cell includes a probabilistic state component. The method further includes training the neural network with a sequence of data. In an embodiment, the method includes extracting a deterministic finite automaton from the trained recurrent neural network and classifying a sequence with the extracted automaton.

Recurrent neural networks having a probabilistic state component and state machines extracted from the recurrent neural networks

A method for predicting one or more events includes generating, for features of each of at least two feature types, an intermediate representation using a representation learning model for the at least two feature types. The intermediate representations of the at least two feature types are analyzed using a neural network and at least one neural network model so as to provide a joint representation for predicting certain events. One or more actions to be taken can be determined based on the one or more events predicted by the joint representation.

Method for predicting events using a joint representation of different feature types

A method for predicting a patient outcome from a caretaker episode includes receiving a current episode snapshot of the caretaker episode comprising multi-modal data of the patient from an electronic health records (EHR) system, the multi-modal data including one or more available data modalities and one or more missing data modalities. The multi-modal data is applied as input to an embedding model having a submodel for each of the data modalities. A first embedding is generated for each of the available data modalities. A second embedding is generated for each of the missing data modalities using corresponding embeddings of neighbors in an episode snapshot graph. The first and second embeddings are combined to obtain a complete embedding. The patient outcome is predicted based on the complete embedding for the current episode snapshot using a machine learning component which has been trained using patient outcomes of the historical episode snapshots.

Method and system for predicting patient outcomes using multi-modal input with missing data modalities

A method for accelerating a neural network includes identifying neural network layers that meet a locality constraint. Code is generated to implement depth-first processing for different hardware based on the identified neural network layers. The generated code is used to perform the depth-first processing on the neural network based on the generated code.

Acceleration of neural networks using depth-first processing

A method for learning latent representations of individual users in a personalization system uses a graph-based machine learning framework. A graph representation is generated based on input data in which the individual users are each represented by a node. The nodes are associated with labels. Node vector representations are learned by combining label latent representations from a vertex and neighboring nodes so as to reconstruct the label latent representation of the vertex and updating the label latent representations of the neighboring nodes using gradients resulting from application of a reconstruction loss. A classifier/regressor is trained using the node vector representations and the node vector representations are mapped to personalizations. Actions associated with the personalizations are then initiated.

System and method for multi-modal graph-based personalization

Systems and methods for contextualizing automatic image segmentation and/or regression including defining an hourglass neural network model, which includes defining an encoder configured to generate compression layers, including a bottleneck layer, and defining a decoder including a contextual convolution operation and configured to generate one or more reconstruction layers. The contextual convolution operation includes establishing, for each of the one or more reconstruction layers, a skip connection between the reconstruction layer and prior layer(s) of different spatial dimension, e.g., between the bottleneck layer and the reconstruction layer and/or between reconstruction layers.

Method and system for contextualizing automatic image segmentation and regression

A method for generating a product recommendation in a retail system includes collecting a dataset containing a plurality of entities and attributes for the entities. Relationships between the plurality of entities are generated. The plurality of entities, attributes and relationships are stored in a knowledge graph. A representation of the plurality of entities, attributes and relationships stored in the knowledge graph is learned. Zero-shot learning is performed for a new entity and attributes for the new entity. The new entity and attributes for the new entity are stored in the knowledge graph. A recommendation for a user is generated based on the knowledge graph.

Method and system for multimodal recommendations

A method for providing recommendations to users includes obtaining stored data structure triples and actual ratings associated with the data structure triples; training a machine learning model using the stored data structure triples and associated actual ratings, wherein training the machine learning model includes generating user, product, and review representations based on the stored data structure triples and their associated ratings; predicting, by the machine learning model, ratings using the generated user, product, and review representations; and making recommendations based on the predicted ratings.

Method and system for combining user, item and review representations for recommender systems

A method detects an event or anomaly in real-time and triggers an action based thereon. A stream of data is received from data sources. The data includes at least two categorical features and a real-value measurement. Sketching is performed on the features using min-wise hashing to create sketches of the data. A regression tree is learnt on the sketches so as to estimate a mean squared error. It is determined whether an event or anomaly exists based on the mean squared error. An action is triggered based on at least one of a type, location or magnitude of the determined event or anomaly.

Scalable system and method for real-time predictions and anomaly detection

A method for object density monitoring includes receiving, by a processing server, an input image captured by an image sensor. The method further includes providing an annotated dataset with a target object to be identified in the input image, and providing, by the processing server as output, an object density map generated from the input image. The processing server provides the object density map by using a deep neural network having one or more pairs of a compression layer and a decompression layer connected by gated shortcuts.

System and method for single image object density estimation

A method for monitoring resources in a computing system having system information includes transforming, via representation learning, variable-size information into fixed-size information, and creating a machine learning neural network model and training it the machine learning model to predict future resource usage of an application. The method further includes providing the prediction of further resources usage of the application as an input to an action component, wherein the action component is one of an anomaly detector or a reinforcement learner that drives a scheduler. The method additionally includes performing, by the action component, at least one of scheduling resources within the computing system or detecting a resources usage anomaly.

Growing community of inventors

Heidelberg, Germany

Mathias Niepert