US 12,380,370 B1
Hardware-aware automated machine learning (AutoML) model creation and optimization
Jeremi Wójcicki, Milan (IT); Evan Petridis, Palo Alto, CA (US); Justin Ormont, Seattle, WA (US); and Adisa Bolić, Sarajevo (BA)
Assigned to Eta Compute, Inc., Sunnyvale, CA (US)
Filed by Eta Compute, Inc., Sunnyvale, CA (US)
Filed on Dec. 30, 2024, as Appl. No. 19/005,923.
Claims priority of provisional application 63/619,192, filed on Jan. 9, 2024.
Int. Cl. G06N 20/00 (2019.01)
CPC G06N 20/00 (2019.01) 20 Claims
OG exemplary drawing
 
1. A system for automated machine learning (Auto ML) creating and optimizing ML models, the system comprising:
a model store unit for:
storing machine learning (ML) model architectures;
storing trained ML models and trained ML hardware models derived from ML model architectures;
storing runtime test metrics data corresponding to each of the stored trained ML models and trained ML hardware models;
storing ML advised-models created based on the runtime test metrics data corresponding to each of the stored trained ML models and trained ML hardware models;
a model meta-services (MMS) unit for:
accessing the machine learning (ML) model architectures;
accessing the trained ML models and trained ML hardware models;
accessing the runtime test metrics data corresponding to the trained ML models and trained ML hardware models;
one of creating, updating or replacing ML meta-models based on the runtime test metrics data corresponding to ML model architectures, the stored trained ML models, trained ML hardware models and ML advised-models and runtime test metrics;
answering to models producer and consumer (MPC) queries by outputting advice on creation, training and optimization of ML advised-models;
a models producer and consumer (MPC) unit for:
selecting a selected ML advised-model from the ML advised-models;
selecting ML test inputs and ML test outputs for testing the selected ML advised-model;
selecting types of test metrics for testing the selected ML advised-model;
testing the selected ML advised-model using the ML test inputs and ML test outputs to provide runtime test metrics data for the selected types of test metrics, wherein the runtime test metrics data are for ML output predictions made by the selected ML advised-model input with the ML test inputs, as compared to the ML test outputs;
training and optimizing the selected ML advised-model to become an optimized trained ML advised-model using the runtime test metrics data;
sending the optimized trained ML advised-model to the model store unit for storing as one of the stored ML advised-models;
sending the runtime test metrics data to the model store unit for storing as part of the runtime test metrics data;
querying model meta-services (MMS) with the MPC query for advice on creation, training and optimization of ML advised-models;
the MPC unit using the meta-models from the MMS unit to optimize the optimized trained ML advised-model using queries that the store unit matches to the meta-models to more efficiently and automatically improve and/or optimize a predictive quality of the optimized trained ML advised-model and improve the performance of resource use and accuracy of further MPC jobs by optimizing the optimized trained ML advised-model of a trained ML hardware models of a selected hardware processor; and
a validator for creating a validated version of the stored ML advised-model by writing the optimized trained ML advised-model to a selected hardware processor to be a validated optimized trained ML advised-model.