US 12,462,075 B2
Resource prediction system for executing machine learning models
Yao Yang, Sunnyvale, CA (US); Andrew Hoonsik Nam, San Francisco, CA (US); Mohamad Mehdi Nasr-Azadani, San Francisco, CA (US); Teresa Sheausan Tung, Tustin, CA (US); Ophelia Min Zhu, San Lorenzo, CA (US); Thien Quang Nguyen, San Jose, CA (US); and Zaid Tashman, San Francisco, CA (US)
Assigned to Accenture Global Solutions Limited, Dublin (IE)
Filed by Accenture Global Solutions Limited, Dublin (IE)
Filed on Feb. 23, 2021, as Appl. No. 17/182,538.
Prior Publication US 2022/0269835 A1, Aug. 25, 2022
Int. Cl. G06F 30/20 (2020.01); G06F 11/34 (2006.01); G06N 20/00 (2019.01)
CPC G06F 30/20 (2020.01) [G06F 11/3461 (2013.01); G06N 20/00 (2019.01)] 16 Claims
OG exemplary drawing
 
1. A system comprising:
non-transitory memory storing instructions executable to select a suggested hardware platform executing a machine learning model under constraints; and
a processor configured to execute the instructions to:
obtain input data including a targeted objective and the constraints;
determine configurations of candidate machine learning models based on the input data;
virtually execute, in a first virtualization process, the candidate machine learning models defined by a hyper-parameter set corresponding to the configurations based on the input data for a selection of machine learning model;
obtain a virtual performance metrics set of each respective candidate machine learning model from virtual execution of the candidate machine learning models;
generate an evaluation score of each respective candidate machine learning model based on a similarity of the virtual performance metrics set of each respective candidate machine learning model to the targeted objective;
select a deployable machine learning model having the evaluation score that meets a predetermined criterion from among the candidate machine learning models;
virtually execute, in a second virtualization process, the deployable machine learning model selected from the candidate machine learning models on each of candidate hardware platforms according to the constraints, the second virtualization process being performed by a hardware performance model to simulate the deployable machine learning model, the hardware performance model being trained based on probabilistic programming and comprising at least one generative model;
generate an assessment report of the virtual performance metrics set of the deployable machine learning model executed on each of the candidate hardware platforms;
select the suggested hardware platform meeting the predetermined criterion from among the candidate hardware platforms for running the selected machine learning model, the suggested hardware platform probabilistically satisfying the targeted objective under the constraints when combined with the deployable machine learning model for execution;
output a model of an implementation of resource prediction based on the suggested hardware platform;
automatically perform, using a resource prediction twin, the resource prediction and allocation of tasks associated with the resource prediction according to the outputted model, wherein the resource prediction twin includes a generative model that is trained, using probabilistic programming, on parameters of an inference model that is initialized with virtual parameters to generate real parameters as input to the outputted model; and
based on the real parameter as input to the model, dynamically assigning a predetermined portion of required tasks to a first hardware device from the candidate hardware platforms and other portion of required tasks to a second hardware device depending on the performance of each respective hardware device in order to optimally accomplish the targeted objective.