Endpoint device peer ecosystem for federated learning
Aspects of the present disclosure relate to an endpoint device peer ecosystem for federated learning. Lower capability endpoint devices within the peer ecosystem (e.g., within a networked household or office) can collect data, such as user data, usage data, environmental data, etc., while being in communication with a local central device (e.g., a laptop). The local central device can receive a trained machine learning model from a remote computing system, and can either A) receive data collected by the lower capability endpoint devices, aggregate it, and retrain the model, or B) provide the model to the lower capability endpoint devices that individually retrain the model, with the local central device aggregating the retrained models. The local central device can then send the retrained and/or aggregated model back to the remote computing system, without exposing the data collected by the peer ecosystem.
1 . A method for providing an endpoint device peer ecosystem for federated learning, the method comprising:
receiving a global machine learning model from a central computing system, the global machine learning model being trained on parameters aggregated from multiple endpoint devices, wherein the parameters are generated based on one or more local datasets collected by respective endpoint devices of the multiple endpoint devices;
aggregating first data collected by two or more locally networked endpoint devices within a first endpoint device peer ecosystem, wherein the first endpoint device peer ecosystem is a first local network of multiple collocated and/or interconnected first endpoint devices including the two or more locally networked endpoint devices;
generating a first retrained instance of the global machine learning model, wherein generating the first retrained instance of the global machine learning model includes updating the parameters based on the aggregated first data; and
transmitting first updated model data, the first updated model data including at least one of A) the first retrained instance of the global machine learning model, B) the updated parameters based on the aggregated first data, or C) both, to the central computing system,
wherein the central computing system:
obtains second updated model data from a second endpoint device peer ecosystem,
wherein the second endpoint device peer ecosystem is a second local network of multiple collocated and/or interconnected second endpoint devices, and
wherein the second updated model data includes at least one of I) a second retrained instance of the global machine learning model, II) adjusted parameters of the global machine learning model based on aggregated second data collected by two or more of the multiple collocated and/or interconnected second endpoint devices, or III) both;
determines that the second endpoint device peer ecosystem is within a threshold geographic distance of the first endpoint device peer ecosystem;
based on the determining that the second endpoint device peer ecosystem is within the threshold geographic distance of the first endpoint device peer ecosystem, aggregates the first updated model data and the second updated model data; and
updates the global machine learning model based on the aggregated first updated model data and second updated model data.
2 . The method of claim 1 , wherein the first updated model data does not include the aggregated first data received from the two or more locally networked endpoint devices.
3 . The method of claim 1 , wherein the two or more locally networked endpoint devices include two or more of an Internet-of-Things device, a vehicle, a computing device, a mobile device, a wearable device, a smart device, a sensor, a gaming system, a television, a speaker, a WiFi router, or any combination thereof.
4 . The method of claim 1 , wherein the two or more locally networked endpoint devices communicate via at least one of Bluetooth, near field communication, a mesh network, WiFi, cellular network, or any combination thereof.
5 . The method of claim 1 , wherein the retrained instance of the global machine learning model is used to personalize services provided by the two or more locally networked endpoint devices.
6 . The method of claim 1 , wherein another locally networked endpoint device within the first endpoint device peer ecosystem aggregates the first data from the two or more locally networked endpoint devices.
7 . The method of claim 6 , wherein the other locally networked endpoint device is a computing system, a mobile phone, an Internet-of-Things hub, a WiFi router, a gaming system, or a smart television.
8 . The method of claim 6 , wherein the other locally networked endpoint device is selected to aggregate the first data from the two or more locally networked endpoint devices based on at least one of local storage, memory, processing speed, power status, latency, or any combination thereof.
9 . The method of claim 1 , wherein the first updated model data is transmitted to the central computing system via a federated learning tower.
10 . The method of claim 9 , wherein the multiple collocated and/or interconnected second endpoint devices are outside of the first endpoint device peer ecosystem.
11 . A computer-readable storage medium storing instructions that, when executed by a computing system, cause the computing system to perform a process for providing an endpoint device peer ecosystem for federated learning, the process comprising:
receiving a global machine learning model from a central computing system, the global machine learning model being trained on parameters aggregated from multiple endpoint devices, wherein the parameters are generated based on one or more local datasets collected by respective endpoint devices of the multiple endpoint devices;
aggregating first data collected by two or more locally networked endpoint devices within a first endpoint device peer ecosystem, wherein the first endpoint device peer ecosystem is a first local network of multiple collocated and/or interconnected first endpoint devices including the two or more locally networked endpoint devices;
generating a first retrained instance of the global machine learning model, wherein generating the first retrained instance of the global machine learning model includes updating the parameters based on the aggregated first data; and
transmitting first updated model data, the first updated model data including at least one of A) the first retrained instance of the global machine learning model, B) the updated parameters based on the aggregated first data, or C) both, to the central computing system,
wherein the central computing system;
obtains second updated model data from a second endpoint device peer ecosystem,
wherein the second endpoint device peer ecosystem is a second local network of multiple collocated and/or interconnected second endpoint devices, and
wherein the second updated model data includes at least one of I) a second retrained instance of the global machine learning model, II) adjusted parameters of the global machine learning model based on aggregated second data collected by two or more of the multiple collocated and/or interconnected second endpoint devices, or III) both;
determines that the second endpoint device peer ecosystem is within a threshold geographic distance of the first endpoint device peer ecosystem;
based on the determining that the second endpoint device peer ecosystem is within the threshold geographic distance of the first endpoint device peer ecosystem, aggregates the first updated model data and the second updated model data; and
updates the global machine learning model based on the aggregated first updated model data and second updated model data.
12 . The computer-readable storage medium of claim 11 , wherein the first updated model data does not include the aggregated first data received from the two or more locally networked endpoint devices.
13 . The computer-readable storage medium of claim 11 , wherein another locally networked endpoint device within the first endpoint device peer ecosystem aggregates the first data from the two or more locally networked endpoint devices.
14 . The computer-readable storage medium of claim 13 , wherein the other locally networked endpoint device is a computing system, a mobile phone, an Internet-of-Things hub, a WiFi router, a gaming system, or a smart television.
15 . The computer-readable storage medium of claim 13 , wherein the other locally networked endpoint device is selected to aggregate the first data from the two or more locally networked endpoint devices based on at least one of local storage, memory, processing speed, power status, latency, or any combination thereof.
16 . A computing system for providing an endpoint device peer ecosystem for federated learning, the computing system comprising:
one or more processors; and
one or more memories storing instructions that, when executed by the one or more processors, cause the computing system to perform a process comprising:
receiving a global machine learning model from a central computing system, the global machine learning model being trained on parameters aggregated from multiple endpoint devices, wherein the parameters are generated based on one or more local datasets collected by respective endpoint devices of the multiple endpoint devices;
aggregating first data collected by two or more locally networked endpoint devices within a first endpoint device peer ecosystem, wherein the first endpoint device peer ecosystem is a first local network of multiple collocated and/or interconnected first endpoint devices including the two or more locally networked endpoint devices;
generating a first retrained instance of the global machine learning model, wherein generating the first retrained instance of the global machine learning model includes updating the parameters based on the aggregated first data; and
transmitting first updated model data, the first updated model data including at least one of A) the first retrained instance of the global machine learning model, B) the updated parameters based on the aggregated first data, or C) both, to the central computing system,
wherein the central computing system;
obtains second updated model data from a second endpoint device peer ecosystem,
wherein the second endpoint device peer ecosystem is a second local network of multiple collocated and/or interconnected second endpoint devices, and
wherein the second updated model data includes at least one of I) a second retrained instance of the global machine learning model, II) adjusted parameters of the global machine learning model based on aggregated second data collected by two or more of the multiple collocated and/or interconnected second endpoint devices, or Ill) both;
determines that the second endpoint device peer ecosystem is within a threshold geographic distance of the first endpoint device peer ecosystem;
based on the determining that the second endpoint device peer ecosystem is within the threshold geographic distance of the first endpoint device peer ecosystem, aggregates the first updated model data and the second updated model data; and
updates the global machine learning model based on the aggregated first updated model data and second updated model data.
17 . The computing system of claim 16 , wherein the two or more locally networked endpoint devices communicate via at least one of Bluetooth, near field communication, a mesh network, WiFi, cellular network, or any combination thereof.
18 . The computing system of claim 16 , wherein the retrained instance of the global machine learning model is used to personalize services provided by the two or more locally networked endpoint devices.
19 . The computing system of claim 16 , wherein the first updated model data is transmitted to the central computing system via a federated learning tower.
20 . The computing system of claim 19 , wherein the multiple collocated and/or interconnected second endpoint devices are outside of the first endpoint device peer ecosystem.