International Business Machines Corporation

Globalfoundries Inc.

Bishwaranjan Bhattacharjee

Mohammad Sadoghi Hamedani

Mustafa Canim

Timothy Ray Malkemus

Kenneth Andrew Ross

Adam J Storm

Sriram Kolinjivadi Padmanabhan

Vijayshankar Raman

Guy Maring Lohman

Ronald Jason Barber

Richard Sefton Sidle

Chandrasekaran Mohan

George Andrei Mihaila

Xun Xue

Christian Alexander Lang

Parijat Dube

Matthew Albert Huras

Lipyeow Lim

Michael Jeffrey Winer

John Ronald Kender

Leslie Anne Cranston

Patrick D Watson

Sam Sampson Lightstone

Yuan-Chi Chang

Kavitha Srinivas

Octavian Udrea

Vinod Muthusamy

Aleksander A Slominski

Michael Robert Glass

Ippokratis Pandis

Matthew Leon Hill

Noel C Codella

Daniele Costante Zilio

Michele Merler

Vatche Isahagian

Catherine Suzanne McArthur

Bassem Makni

Leslie Anne Buback

Tony Wen Hsun Lai

Siyu Huo

Sherman Lau

Kwai Wong

Brian Michael Belgodere

John Richard Smith

Philip Shilung Yu

Charles Philippe Tresser

Julian Timothy Dolby

Alfio Massimiliano Gliozzo

Bruce Gilbert Lindsay

Tong Chen

Kenneth Alan Goldman

Zehra Noman Sura

Asit Dan

Paul Christesten Castro

John Steven Langford

Patrick Dantressangle

Mihaela Ancuta Bornea

Yulong Li

Radu Florian

Ioana Roxana Stanoi

Avirup Sil

Naoki Abe

Ramesh Chandra Agarwal

Taesun Moon

Phuong Thanh Nguyen

Matthias Gruetzner

Rameswar Panda

Chidanand V Apte

Hui Wan

Masayasu Muraoka

Sean W McKeough

Yuan Chi Chang

Kaarel Truuvert

Anshul Gupta

Yang Zhao

Carl E Abrams

Zoltan G Toth

Ayushmaan Aggarwal

Niharika Bhattacharjee

Jo-anne Mona Louise Kirton

Aadi Gupta Bhattacharya

Raka Bose

Deepta Bhattacharya Gupta

Renuka Muralidhar

Elina Rani

Matthew Greenberg

Ehab Abdelhamid Mohammed Abdelhamid

Aashka Trivedi

Parul Awasthy

Graeme William Blackwood

Tushar Nagarajan

Fabian Nagel

Matthew A Hilbert

Kathleen M Troidle

Cathy McArthur

Reza Sherkat

Systems, computer-implemented methods, and computer program products to facilitate improving a mono-lingual language model using pre-trained models and parallel data are provided. According to an embodiment, a system can comprise a processor that executes components stored in memory. The computer executable components can comprise a training component that can use a first language model to improve a mono-lingual second language model through mask language modeling (MLM) training and contrastive learning. The computer executable components can comprise a data component that can collect parallel data for the first language model and the mono-lingual second language model. The computer executable components can comprise a tokenizing component that can tokenize the parallel data for the first language model and the mono-lingual second language model. The computer executable components can comprise an alignment component that can create an alignment of the tokenized parallel data.

Mono-lingual language models using parallel data

In an embodiment, a method for fine-tuning a pre-trained neural network for transfer learning, the method comprising obtaining a first target feature vector from a first layer of a pre-trained neural network responsive to a first target data element of a target dataset passing therethrough, obtaining a first source feature vector associated with the first layer of the pre-trained neural network, calculating a first divergence value for the first layer of the pre-trained neural network based at least in part on the first target feature vector and the first source feature vector, and setting a learning rate for the first layer of the pre-trained neural network based at least in part on the first divergence value.

Automated fine-tuning of a pre-trained neural network for transfer learning

A system for ranking machine learning base models for transfer learning purposes is described. The system receives image data in the form an image or an image set and extracts image tags from the images. The image tags are expanded into a set of associated terms using a word embedding database and model. The associated terms are used to query a knowledge database for parent or categorical terms used to rank various matching machine learning base models that may be improved or trained by the image data.

Ranking image sources for transfer learning

A neural architecture search method, system, and computer program product that determines, by a computing device, a best fit language model of a plurality of language models that is a best fit for interpretation of a corpus of natural language and interprets, by the computing device, the corpus of natural language using the best fit language model.

Neural architecture search of language models using knowledge distillation

A computer system, product, and method are provided. The computer system includes an artificial intelligence (AI) platform operatively coupled to a processor. The AI platform includes tools in the form of a machine learning model (MLM) manager, a metric manager, and a training manager. The MLM manager accesses a plurality of pre-trained source MLMs, and inputs a plurality of data objects of a test dataset into each of the source MLMs. The test dataset includes the plurality of data objects associated with respective labels. For each source MLM, associated labels are generated from the inputted data objects and a similarity metric is calculated. The MLM manager selects a base MLM to be used for transfer learning from the plurality of source MLMs based upon the calculated similarity metric. The training manager trains the selected base MLM with a target dataset for the target domain.

Predictive data and model selection for transfer learning in natural language processing

A method includes storing an anchor row vector identification for an anchor row to a local memory. It is determined whether the anchor row vector identification is visible based on isolation requirements. The anchor row vector identification is accessed upon a determination that the anchor row vector identification is visible, and the row vector identification is re-read from the local memory. It is determined whether the anchor row vector identification has not changed since a start of the accessing. Upon a determination that the anchor row vector identification has not changed, read anchor row fields are returned. A first check history is performed on an anchor row history tuple sequence number (TSN) for the anchor row.

In-place updates with concurrent reads in a decomposed state

Whether to train a new neural network model can be determined based on similarity estimates between a sample data set and a plurality of source data sets associated with a plurality of prior-trained neural network models. A cluster among the plurality of prior-trained neural network models can be determined. A set of training data based on the cluster can be determined. The new neural network model can be trained based on the set of training data.

Training transfer-focused models for deep learning

Embodiments for implementing mixed precision learning for neural networks by a processor. A neural network may be replicated into a plurality of replicated instances and each of the plurality of replicated instances differ in precision used for representing and determining parameters of the neural network. Data instances may be routed to one or more of the plurality of replicated instances for processing according to a data pre-processing operation.

Data driven mixed precision learning for neural networks

A method, system and computer program product for analyzing multidimensional data are disclosed. In embodiments, the method comprises obtaining an original set of data having a sequential order and multiple original dimensions; selecting a topic-based summarization scheme to summarize the original set of data; and applying the selected topic-based summarization scheme to the original set of data to transform the original set of data into a new set of data having fewer dimensions than the original set of data, while preserving, within a defined measure, the sequential order of the original set of data. In embodiments, the selecting a topic-based summarization scheme includes selecting a plurality of topics, each of the topic representing a set of the original dimensions. In embodiments, the applying the topic-based summarization scheme includes performing dimensionality reduction on the original set of data to transform the original dimensions to the topics.

Analyzing multidimensional process traces under edit-distance constraint

Techniques for distributing the training of machine learning models across a plurality of computing devices are presented. An example method includes receiving, from a computing device in a distributed computing environment, a request for a set of outstanding jobs for training part of a machine learning model. A system transmits, to the computing device, information identifying the set of outstanding jobs. The system receives, from the computing device, a selected job for execution on the computing device from the set of outstanding jobs. A chunk of training data associated with the selected job and one or more parameters associated with the selected job may be transmitted to the computing device, and the system may take one or more actions with respect to the chunk of data associated with the selected job based on a response from the computing device.

Peer assisted distributed architecture for training machine learning models

A computer-implemented method for data labeling is provided. The computer-implemented method assigns pseudo-labels to unlabeled examples of data using a similarity metric on an embedding space to produce pseudo-labeled examples. A curriculum learning model is trained using the pseudo-labeled examples. The curriculum learning model trained with the pseudo-labeled examples is employed in in a fine-tuning task to enhance classification accuracy of the data.

Generating and augmenting transfer learning datasets with pseudo-labeled images

Aspects include a system, computer program production and computer-implemented method for tagging an image. An image classification engine stored in a memory of a computer device generates a plurality of tags for the image and uses the plurality of tags to generate a relevance subgraph for the image. An embedding engine embeds nodes and edges of the relevance subgraph into fixed dimension vectors of a matrix. A neural network stored in the memory determines a feature vector from the image. A processor applies the feature vector to the matrix to generate a context vector for the image. The context vector is used to tag the image.

Generating and ordering tags for an image using subgraph of concepts

Methods and systems for performing database transactions include executing a first transaction request in a preplay mode that locks the requested data with a prefetch-lock and reads one or more requested data items from storage into a main memory buffer; locking the requested data items with a read/write lock after said data items are read into the main memory buffer; and performing the requested transaction on the data items in the main memory buffer using a processor.

Preplaying transactions that mix hot and cold data

A computer-implemented mitigation and warning method, system, and computer program product including detecting an enclosure temperature via an enclosure temperature probe and activating, via a controller, a thermoelectric device when the enclosure temperature exceeds a predetermined threshold temperature to transfer heat between an enclosure and a phase change material (PCM) to cause the PCM to change phase and cool the enclosure.

Thermal cooling of an enclosure

A method includes logically organizing, by an object hierarchy processor, data objects in a first hierarchy. A portion of the data objects in the first hierarchy logically includes groupings of other data objects. The object hierarchy processor physically organizes the data objects across two or more types of memory in a second hierarchy. Another portion of the data objects in the second hierarchy physically includes groupings of other data objects. Groupings of the data objects in the second hierarchy are dynamically moved across the two or more types of memory. Levels of access of the data objects are tracked using a data structure that maps groupings of the data objects in the first hierarchy onto metadata information including combined access frequencies of the data objects, and current number of accessors to the data objects, in each grouping of the data objects.

Efficient performance of insert and point query operations in a column store

Growing community of inventors

Yorktown Heights, NY, United States of America

Bishwaranjan Bhattacharjee