On-device machine learning calibration
AssigneeSamsara Inc.
InventorsNarendran Rajan, Sharan Srinivasan, Jing Wang, Yan Wang
Life of this patent
Estimating the term…
Abstract
A method and apparatus of utilizing an edge device comprising storing in a memory a pre-installed trained machine learning model, the machine learning model comprising a fixed base portion and a flexible portion. The method further comprises, when the edge device is installed, collecting data using one or more sensors on the edge device, and using the fixed base portion of the machine learning model on the data for a first aspect classification, while using the flexible portion of the machine learning model for a second aspect classification. The method further comprises calibrating the flexible portion of the machine learning model during the use.
Claims20
A method of utilizing an edge device comprising:
- storing in a memory a pre-installed trained machine learning model, the machine learning model comprising a fixed base portion and a flexible portion;
- initializing the flexible portion as a general probabilistic model, with a default configuration;
- calibrating the flexible portion by refining parameters based on an output of the fixed base portion of the machine learning model;
- collecting data using one or more sensors on the edge device;
- using the fixed base portion of the machine learning model on the data from the one or more sensors for a first aspect classification, the first aspect classification comprising the output of the fixed base portion;
- and
- using the flexible portion of the machine learning model on the data from the one or more sensors for a second aspect classification.
A method of utilizing a vehicle safety device comprising:
- storing in a memory a pre-installed trained machine learning model, the machine learning model comprising a fixed base portion and a flexible portion;
- initializing the flexible portion as a general probabilistic model, with a default configuration;
- collecting image data using one or more sensors on the vehicle safety device;
- using the fixed base portion of the machine learning model to evaluate the image data for a first aspect classification;
- using the flexible portion of the machine learning model to evaluate the image data for a second aspect classification;
- adjusting one or more parameters of the general probabilistic model of the flexible portion on the vehicle safety device, based on an output of the machine learning model and user feedback.
An edge device comprising:
- a memory to store a pre-installed trained machine learning model, the machine learning model comprising a fixed base portion and a flexible portion;
- one or more sensors to collect data;
- a processor to run the fixed base portion of the machine learning model on the data from the one or more sensors for a first aspect classification;
- the processor to run the flexible portion of the machine learning model on the data from the one or more sensors for a second aspect classification;
- wherein the flexible portion of the pre-installed trained machine learning model is initialized with a default configuration, and the processor calibrating the flexible portion of the machine learning model refines one or more parameters, based on the first aspect classification from the fixed base portion of the machine learning model.
Dependencies are inferred from the claim text.