US 12,088,719 B2
Method and system for incremental training of machine learning models on edge devices
Subash Sundaresan, Fremont, CA (US)
Assigned to Subash Sundaresan, Fremont, CA (US)
Filed by swarmin.ai, Fremont, CA (US)
Filed on Nov. 8, 2020, as Appl. No. 17/092,289.
Claims priority of provisional application 62/965,099, filed on Jan. 23, 2020.
Prior Publication US 2021/0232981 A1, Jul. 29, 2021
Int. Cl. H04L 9/32 (2006.01); G06F 21/57 (2013.01); G06N 20/00 (2019.01); G06N 20/20 (2019.01); H04L 9/40 (2022.01); G06N 3/098 (2023.01)
CPC H04L 9/3213 (2013.01) [G06F 21/57 (2013.01); G06N 20/00 (2019.01); G06N 20/20 (2019.01); H04L 9/32 (2013.01); H04L 63/08 (2013.01); G06N 3/098 (2023.01)] 16 Claims
OG exemplary drawing
 
1. A system comprising a first edge device that performs an incremental training of one or more machine learning (ML) models, wherein the first edge device is communicatively connected to other edge devices in a peer to peer network, the first edge device comprising:
a processor that is configured to
(a) implement a base version of a first machine learning (ML) model in the first edge device, wherein the first ML model comprises one or more first parameters and one or more first weights associated with the one or more first parameters;
(b) incrementally train the base version of the first ML model upon occurrence of a first data event, by updating the one or more first weights of the first ML model during a first predetermined window of time;
(c) incrementally train a second ML model associated with the second edge device by communicating the one or more first updated weights to the second edge device by enabling a second edge device, in the peer to peer network, wherein the second edge device updates the second ML model based on the one or more first updated weights and incrementally trains the updated second ML model, upon occurrence of a second data event at the second edge device, by updating one or more second weights associated with one or more second parameters of the second ML model during the first predetermined window of time;
(d) update the first ML model based on the one or more second updated weights received from the second edge device to obtain an updated version of the first ML model; and
(e) incrementally train the updated version of the first ML model, upon occurrence of a subsequent data event, by updating the one or more first weights of the updated version of the first ML model during the first predetermined window of time; and
a certifying node that is communicatively connected with the first edge device to regulate the incremental training of the first ML model, wherein the first edge device is configured to register with the certifying node while joining the peer to peer network, wherein the certifying node provides an encrypted key to the first edge device that is used to authenticate any subsequent updates from the first edge device, wherein the first edge device performs the incremental training on the updated versions of the first ML model that is certified and resubmits a subsequent version of the first ML model to the certifying node for certification.