| CPC G06N 20/20 (2019.01) [G06F 18/2411 (2023.01); G06N 3/08 (2013.01); G06N 3/126 (2013.01); G06N 5/043 (2013.01); G06N 7/01 (2023.01); G06N 20/00 (2019.01)] | 4 Claims |
|
1. A system comprising:
a plurality of electronic devices that each include a first processor and a sensor, the first processor being configured to perform an automated decision making function with a local model; and
a second processor,
the second processor being configured to
receive features data and a local model result through a communication device from the first processor of an electronic device of the plurality of electronic devices only in response to the confidence score of the local model result being lower than a first threshold, the features data and the local model result being for use in updating a global model stored at the second processor, the global model being used in common with the plurality of electronic devices including the electronic device, the local model result including a first predicted value obtained by applying a local model to features data derived from raw data obtained at the electronic device, wherein the local model is obtained by use of a first machine learning process and the confidence score is a probability estimate associated with the first predicted value,
update the global model based on the received features data and the received local model result by use of a second machine learning process,
send the updated global model to all of the electronic devices including the electronic device, wherein the first processor updates the local model with the updated global model, the updated local model being the updated global model, when an evaluated confidence score of the updated global model is not lower than a second threshold, the evaluated confidence score being associated with a second predicted value using features data by applying the updated global model,
send, to the first processor, a request for pre-processed data for use in updating the global model when the evaluated confidence score of the updated global model is lower than the second threshold, and
receive the pre-processed data through the communication device from the first processor based on the request, the global model being updated with features extracted from the pre-processed data in the electronic device,
the first processor of an electronic device of the plurality of electronic devices being configured to:
obtain the raw data at the electronic device;
apply the local model on the electronic device to features data to obtain the local model result, the features data being derived from the raw data, the local model result including the first predicted value using the features data by applying the local model, wherein the local model is obtained by use of the first machine learning process;
generate the confidence score of the local model result, the confidence score being a probability estimate associated with the first predicted value;
send, to the second processor, the features data and the local model result through the communication device for use in updating a global model only in response to the confidence score of the local model result being lower than the first threshold, the global model being used in common with the plurality of electronic devices, and not send, to the second processor, the features data and the local model result through the communication device for use in updating a global model in response to the confidence score of the local model result being equal to or not being lower than the first threshold;
receive the updated global model data from the second processor when the confidence score of the updated global model is not lower than the second threshold;
receive, from the second processor, a request for pre-processed data for use in updating the global model when the evaluated confidence score of the updated global model is lower than the second threshold; and
send the pre-processed data through the communication device to the second processor based on the request,
wherein the pre-processed data includes fusion of the raw data collected from multiple sensors of the plurality of electronic devices, by performing at least one of resampling, interpolation, and filtering the raw data.
|