Amazon Technologies, Inc.

University of Minnesota

George Karypis

Karthik Gurumoorthy Subramanya Bharathy

Yida Wang

Mu Li

Bradley R Bebee

Trishul Amit Madhukar Chilimbi

Shuai Zheng

Zhen Zhang

Divyakala Vel

Justin Chiu

Dominique Welle LaSalle

Pascal Asselot

Daniel Lawrence Bork

Brian Joshua Berman

Xiang Song

Erich Lawson Spangenberg

Ankit Gupta

Adesoji Adeshina

A resource set which includes multiple servers with a respective plurality of training computing devices is identified for training a machine learning model. The resource set is subdivided into partition groups, such that each partition group can store a respective replica of state information of the model. The model is trained using the partition groups. The training comprises a multi-stage gathering of a portion of the state information at training computing devices of a particular partition group. Different types of communication channels between training computing devices are used in respective stages of the gathering, including inter-server communication channels in one stage and an intra-server communication channel during another stage. A trained version of the model is stored.

Distributed training of machine learning models

A computer-implemented method performed at least in part by a graph database configured to store at least one graph dataset. The method includes receiving a graph query configured to be performed against a machine learning model, and communicating the graph query with a machine learning system that is configured to use the machine learning model to obtain model inference results and communicate those model inference results to the graph database application. The graph database provides query results based at least in part on the model inference results to an entity.

Using graph queries to obtain results from machine learning models

A system and method for a patent search and analytics software tool that finds prior art for each claim limitation/element by breaking up claims into individual claim limitations. Once the claims are separated, the system finds the best prior art for each of the individual separate, different claim limitations/elements. Additionally, the software finds the best prior art for entire claims, including non-patent literature (NPL) searching. The system takes into account the limitations of the claim under consideration (query claim of query patent), the text of the art, the link structure of the citation network, and the patent classification and then constructs a network that consists of two types of nodes: (i) the art (patents and non-patent literature) and the (ii) classes of the patent classification. Each art node is linked to all the art nodes that it cites and is linked to all the classification nodes that it belongs to.

Automatically separating claim into elements/limitations and automatically finding art for each element/limitation

Multiple parallel slave processes and a master process are assigned to a node executing an operating system such that the operating system maintains a ready queue comprising a list of one or more processes that are ready to be executed by at least one processing core. A parallel slave process takes an action that causes the operating system to keep the parallel slave process out of the ready queue. Based on receiving an indication that the parallel slave process is to be kept out of the ready queue, the master process sets the parallel slave process to a blocking state, selects a second parallel slave process that is in a runnable state but is currently kept from being in the ready queue, and takes an action that causes the operating system to add the parallel slave process that is in the runnable state to the ready queue.

Parallel processing with cooperative multitasking

Growing community of inventors

Bloomington, MN, United States of America

George Karypis