International Business Machines Corporation

Visa International Service Association

Western Digital (fremont), Inc.

Oracle International Corporation

Google Inc.

Hewlett-packard Development Company, L.p.

Palo Alto Research Center Inc.

Axt, Inc.

Facebook, Inc.

Hewlett Packard Enterprise Development LP

Nuance Communications, Inc.

Quantumscape Corporation

Visa International Service Application

Harbin Biyan Technology Co., Ltd.

Wei Zhang

Liang Wang

Hao Yang

Yan Zheng

Li Feng Zhang

Xiaodong Cui

Yufei Ren

Michael Yeh

Yandong Wang

Liang Gou

Junpeng Wang

Yingjian Chen

Alper Buyuktosunoglu

Abdullah Kayi

Amit Ganesh

Kothanda Umamageswaran

Sujatha Muthulingam

Mohamad Towfik Krounbi

Tong Zhang

Ming Mao

Ulrich Alfons Finkler

Satoru Araki

Fei Wang

Yuqing Gao

Yada Zhu

Michel H T Hack

Lingfei Wu

Niloy Mukherjee

Jia Shi

James M A Begole

Michelle Xue Zhou

Hongping Yuan

Guanxiong Li

Hugh Craig Hiner

Liang Gu

Benjamin P Chen

Weiguo Liu

Takashi Matsumoto

Ying Zhang

Lei Wang

Daniel R Tretter

Rui Zhang

Salih Burak Gokturk

Minsik Cho

Jinqiu Zhang

Juan Liu

Barton Allen Smith

Krishnan Meiyyappan

David Shing-Ki Kung

Maurice Kyojin Chu

Danning Yang

Yang Yu

Nicolas Scheffer

Murray Scott Campbell

Baris Sumengen

Ying Hai Chen

Yimin Guo

Wei Tan

Mark Alan Lavin

Shaoping Li

Tim Holme

Mahendra Pakala

Feng Q Liu

Chia-Yu Chen

Tarun Kumar

Rainer Johannes Fasching

Harumi Kuno

Sining Mao

Siyuan Lu

Martin J Hirzel

Yun-Fei Li

Hai Sun

Xiaohong Gong

Dheeraj Pandey

Dmitry Mikhailovich Potapov

Kevin Wilson Warren

Wei Gao

Kumar Bhaskaran

Alexander Tsukerman

Lingyun Miao

Hui Wu

Andrew Paul Breen

Chiranjeet Chetia

Marcus Oliver Freitag

Conrad M Albrecht

Anup Roy

Evgeniy Gabrilovich

Minghua Xu

David Paul Grove

Michele Merler

Zhihong Zhang

Yuhang Wu

Zuoyu Tao

Neil J S MacNaughton

Xiaoyu Yang

Kevin Patrick Murphy

Ildar Khabibrakhmanov

Krishna Kunchithapadam

Rameswar Panda

Aditya Pal

An embodiment includes identifying an initial plurality of sets of hyperparameter values at which to evaluate an objective function that relates hyperparameter values to performance values of a neural network. The embodiment also executes training processes on the neural network with the hyperparameters set to the each of the initial sets of hyperparameter values such that the training process provides an initial set of the performance values for the objective function. The embodiment also generates an approximation of the objective function using splines at selected performance values. The embodiment approximates a point at which the approximation of the objective function reaches a maximum value, then determines an updated set of hyperparameter values associated with the maximum value. The embodiment then executes a runtime process using the neural network with the hyperparameters set to the updated set of hyperparameter values.

Configuring a neural network using smoothing splines

A machine receives a first set of global parameters from a global parameter server. Multiple learner processors in the machine execute an algorithm that models an entity type using the first set of global parameters and a mini-batch of data known to describe the entity type. The machine generates a consolidated set of gradients that describes a direction for the first set of global parameters in order to improve an accuracy of the algorithm in modeling the entity type when using the first set of global parameters and the mini-batch of data. The machine transmits the consolidated set of gradients to the global parameter server. The machine then receives a second set of global parameters from the global parameter server, where the second set of global parameters is a modification of the first set of global parameters based on the consolidated set of gradients.

Parameter data sharing for multi-learner training of machine learning applications

A method is disclosed. The method comprises determining a time series, a subsequence length. The length of the time series may then be determined, and an initial matrix profile may then be computed. The method may then form a processed matrix profile for a first subsequence of the subsequence length by applying the first subsequence to the initial matrix profile. A second subsequence may then be determined from the processed matrix profile. The method may then include comparing the second subsequence to other subsequences in a dictionary and adding it to the dictionary. The subsequences in the dictionary may be used to generate a plurality of subsequence matrix profiles. The method may then include forming an approximate matrix profile using the plurality of subsequence matrix profiles and then determining one or more anomalies in the time series or another time series using the approximate matrix profile.

Error-bounded approximate time series join using compact dictionary representation of time series

Hessian matrix-free sample-based techniques for model explanations that are faithful to the model are provided. In one aspect, a method for explaining a machine learning model {circumflex over (θ)} (e.g., for natural language processing) is provided. The method includes: training the machine learning model {circumflex over (θ)} with training data D; obtaining a decision of the machine learning model {circumflex over (θ)}; and explaining the decision of the machine learning model {circumflex over (θ)} using training examples from the training data D.

Faithful and efficient sample-based model explanations

Methods and wafers for vertical gradient freeze 8 inch gallium arsenide (GaAs) substrates. In disclosed examples, vertical gradient freeze systems for forming gallium arsenide (GaAs) substrates having silicon as a dopant, the system includes a crucible to contain a GaAs liquid melt and seed material during a formation process; one or more heating coils arranged in a plurality of heating zones; and a pedestal to move relative to the crucible, the system operable to control heating of the plurality of heating zones and movement of the pedestal to form a single crystal GaAs substrate.

Method and system for vertical gradient freeze 8 inch gallium arsenide substrates

An artificial intelligence (AI) platform to support random action replay for natural language (NL) learning. A NL conversation is subject to exploration to train a neural network. One or more tuples are leveraged for the training, with each tuple representing an input action, a vector, an output action, and a reward value. An action is sampled from the vector, with the sampling configured to assess a corresponding first gradient. The first gradient is applied to selectively adjust the neural network. As NL input is received and applied to the selectively adjusted neural network, an output corresponding to the NL input is identified and a corresponding action is subject to be executed.

Random action replay for reinforcement learning

A machine learning model can be trained to predict one or more financial indicators using earnings call transcripts augmented with counterfactual information. Using faithful gradient-based method, prediction results with respect to a particular counterfactual information can be explained. Based on the explanation, the counterfactual information determined to have most impact on prediction results can be selected for updating the machine learning model.

Counterfactual neural network learning for contextual enhanced earnings call analysis

Provided are systems for generating a machine learning model for classification tasks using unadversarial training that include a processor to perform an unadversarial training procedure to train a machine learning model to provide a trained machine learning model. When performing the unadversarial training procedure, the processor is programmed or configured to receive a training dataset including a plurality of training samples; generate a noise vector for the plurality of training samples based on a uniform distribution; perturb each training sample of the plurality of training samples; obtain a gradient; generate an updated noise vector based on the gradient; perturb each training sample of the plurality of training samples based on the updated noise vector; and update a model weight of the machine learning model based on the second plurality of perturbed training samples to provide the trained machine learning model. Methods and computer program products are also provided.

System, method, and computer program product for reducing dataset biases in natural language inference tasks using unadversarial training

Provided is a method for normalizing embeddings for cross-embedding alignment. The method may include applying mean centering to the at least one embedding set, applying spectral normalization to the at least one embedding set, and/or applying length normalization to the at least one embedding set. Spectral normalization may include decomposing the at least one embedding set, determining an average singular value of the at least one embedding set, determining a respective substitute singular value for each respective singular value of a diagonal matrix, and/or replacing the at least one embedding set with a product of the at least one embedding set, a right singular vector, and an inverse of the substitute diagonal matrix. The mean centering, spectral normalization, and/or length normalization may be iteratively repeated for a configurable number of iterations. A system and computer program product are also disclosed.

Method, system, and computer program product for normalizing embeddings for cross-embedding alignment

An approach to determine the quality of encodings assigned to a word by a word embedding model. The approach may include determining the asymmetry of two embeddings associated with two words from a word embedding model. The asymmetry of the two words from a preexisting evocation dataset may be determined. The asymmetry of the two embeddings may be compared to the asymmetry from the evocation dataset to generate an encoding quality score.

Word embedding quality assessment through asymmetry

Provided is a system for analyzing features associated with entities using an embedding tree, the system including at least one processor programmed or configured to receive a dataset associated with a plurality of entities, wherein the dataset comprises a plurality of data instances for a plurality of entities. The processor may be programmed or configured to generate at least two embeddings based on the dataset and determine split criteria for partitioning an embedding space of at least one embedding tree associated with the dataset based on feature data associated with an entity and embedding data associated with the at least two embeddings. The processor may be programmed or configured to generate at least one embedding tree having a plurality of nodes based on the split criteria. Methods and computer program products are also provided.

System, method, and computer program product for feature analysis using an embedding tree

Provided is a method for detecting group activities in a network. The method may include receiving interaction data associated with a plurality of interactions. For each account identifier associated with at least one interaction, a value may be determined for each of a first set of categories, and a vector may be generated based on the value for each category. The length of each vector may be determined. At least one relational graph may be generated based on the interaction data. Each relational graph may be associated with a respective category of a second set of categories. At least one cluster of nodes may be determined based on the relational graph(s). A score for each cluster may be determined based on the length of the vector associated with the account identifier of each node of the cluster of nodes. A system and computer program product are also disclosed.

Method, system, and computer program product for detecting group activities in a network

Described are a system, method, and computer program product for dynamic node classification in temporal-based machine learning classification models. The method includes receiving graph data of a discrete time dynamic graph including graph snapshots, and node classifications associated with all nodes in the discrete time dynamic graph. The method includes converting the discrete time dynamic graph to a time-augmented spatio-temporal graph and generating an adjacency matrix based on a temporal walk of the time-augmented spatio-temporal graph. The method includes generating an adaptive information transition matrix based on the adjacency matrix and determining feature vectors based on the nodes and the node attribute matrix of each graph snapshot. The method includes generating and propagating initial node representations across information propagation layers using the adaptive information transition matrix and classifying a node of the discrete time dynamic graph subsequent to the first time period based on final node representations.

System, method, and computer program product for dynamic node classification in temporal-based machine learning classification models

A method includes extracting, by an analysis computer, a plurality of first datasets from a plurality of graph snapshots using a structural self-attention module. The analysis computer can then extract at least a second dataset from the plurality of first datasets using a temporal self-attention module across the plurality of graph snapshots. The analysis computer can then perform graph context prediction with at least the second dataset.

Dynamic graph representation learning via attention networks

Embodiments for training a recommendation system to provide merchant recommendations comprise receiving, by a processor, raw merchant embeddings and raw user embeddings generated from payment transaction records, wherein the raw merchant embeddings include a plurality of embedded features. A generative adversarial network (GAN) is trained to generate modified merchant embeddings from the raw merchant embeddings, where the modified embeddings remove a location feature. Subsequent to training and responsive to receiving a request for merchant recommendations in the target location for the target user, the GAN and a trained preference model are used to generate a list of merchant rankings based on a new set of modified merchant embeddings, past preferences of a target user, and the target location to recommend merchants in the target location.

Unsupervised embeddings disentanglement using a gan for merchant recommendations

Growing community of inventors

Fremont, CA, United States of America

Wei Zhang