System and method for distributed learning of universal vector representations on edge devices

1 . A method for learning universal vector representation of concept tokens in a distributed environment comprising a plurality of edge devices, the method comprising:

obtaining, by a candidate edge device from the plurality of edge devices, data from at least two different domains including sensor information, profile information, content information;

determining, by the candidate edge device, a plurality of concept tokens used to train for at least one on-device artificial intelligence (AI) model from the obtained data;

training, by the candidate edge device, the at least one on-device AI model locally available at the candidate edge device using the plurality of concept tokens, wherein the at least one trained on-device AI model is specific to learning of universal vector representation of concept tokens available at the candidate edge device, wherein the universal vector representation indicates a presence of at least one relationship among a plurality of concept representations corresponding to the plurality of the concept tokens;

transmitting, by the candidate edge device, encrypted weights associated with the at least one trained on-device AI model to generate the global AI model by aggregating the weights received from each candidate edge device to a server;

receiving, by the candidate edge device, a global AI model generated based on the encrypted weights for deployment from the server, wherein the global AI model is specific to learning of universal vector representation of the concept tokens available at the candidate edge device and learning of the universal vector representation of concept tokens available across remaining edge devices of the plurality of edge devices in the distributed environment; and

deploying, by the candidate edge device, the global AI model for universal vector representation of concept tokens in the candidate edge device,

wherein the at least one relationship comprising at least one of a relative temporal relationship, an absolute temporal relationship, a relative spatial relationship, an absolute spatial relationship, a user behavior sequence relationship, a tertiary conceptual relationship, or a semantic relationship.

2 . The method of claim 1 , wherein the determining of the plurality of concept tokens comprises at least one of:

determining a plurality of device concept tokens based on sensor information received from the one or more sources available at the candidate edge device;

determining a plurality of profile concept tokens based on one or more users behavior information received from the one or more sources available at the candidate edge device; or

determining a plurality of content concept tokens based on other information received from the one or more sources available at the candidate edge device.

3 . The method of claim 1 , wherein the at least one on-device AI model comprises at least one concept representation generator model for generating at least one concept representation for the plurality of concept tokens and at least one relationship model for indicating the presence of the at least one relationship among the plurality of concept tokens.

4 . The method of claim 3 , wherein the training, by the candidate edge device, of the at least one on-device AI model locally available at the candidate edge device using the plurality of concept tokens comprises:

determining, by the candidate edge device, at least one relationship among the plurality of concept tokens; and

training, by the candidate edge device, the at least one concept representation generator model and the at least one relationship model using the plurality of concept tokens and the determined relationship.

5 . The method of claim 4 , wherein the determining, by the candidate edge device, of the at least one relationship among the plurality of concept tokens comprises determining at least one of temporal relationships indicating time-based relation among the plurality of concept tokens, positional relationships indicating position and location based relations among the plurality of concept tokens, sequential relationships indicating a sequence of relation among the plurality of concept tokens, or conceptual relationships indicating correlations among plurality of concept tokens.

6 . The method of claim 1 , wherein the global AI model comprises at least one global concept representation generator model for generating the universal vector representation for the plurality of concept tokens and at least one global relationship model for indicating the presence of the at least one relationship among the plurality of concept tokens.

7 . The method of claim 6 , further comprising:

receiving, by the candidate edge device, an input comprising one or more concept tokens, and a request to generate universal vector representation of the one or more concept tokens; and

generating, by the candidate edge device, the universal vector representation of the one or more concept tokens by inputting the one or more concept tokens to at least one global concept representation generator model of the global AI model.

8 . The method of claim 6 , further comprising:

receiving, by the candidate edge device, an input comprising two or more concept tokens, and a request to generate relevancy between the two or more concept tokens;

determining, by the candidate edge device, universal vector representation of the two or more concept tokens by inputting the two or more concept tokens to the at least one global concept representation generator model; and

generating, by the candidate edge device, the relevancy between the two or more concept tokens by inputting the determined universal vector representation of the two or more concept tokens to at least one global relationship model of the global AI model.

9 . An edge device for learning universal vector representation of concept tokens in a distributed environment comprising a plurality of edge devices, the edge device comprising:

a memory; and

at least one processor, coupled to the memory, and configured to:

obtain data from at least two different domains including sensor information, profile information, content information,

determine a plurality of concept tokens used to train for at least one on-device artificial intelligence (AI) model from the obtained data,

train the at least one on-device AI model locally available at the edge device using the plurality of concept tokens, wherein the at least one trained on-device AI model is specific to learning of universal vector representation of concept tokens available at the edge device, wherein the universal vector representation indicates a presence of at least one relationship among a plurality of concept representations corresponding to the plurality of the concept tokens,

transmit encrypted weights associated with the at least one trained on-device AI model to generate the global AI model by aggregating the weights received from each candidate edge device to a server,

receive a global AI model for deployment from the server, wherein the global AI model generated based on the encrypted weights is specific to learning of universal vector representation of the concept tokens available at the edge device and learning of the universal vector representation of concept tokens available across remaining edge devices of the plurality of edge devices in the distributed environment, and

deploy the global AI model for universal vector representation of concept tokens in the edge device,

wherein the at least one relationship comprising at least one of a relative temporal relationship, an absolute temporal relationship, a relative spatial relationship, an absolute spatial relationship, a user behavior sequence relationship, a tertiary conceptual relationship, or a semantic relationship.

10 . The edge device of claim 9 , wherein, to determine the plurality of concept tokens, the at least one processor is further configured to:

determine a plurality of device concept tokens based on sensor information received from the one or more sources available at the edge device,

determine a plurality of profile concept tokens based on one or more users behavior information received from the one or more sources available at the edge device, and

determine a plurality of content concept tokens based on other information received from the one or more sources available at the edge device.

11 . The edge device of claim 9 , wherein, the at least one on-device AI model comprises at least one concept representation generator model for generating at least one concept representation for the plurality of concept tokens and at least one relationship model for indicating the presence of the at least one relationship among the plurality of concept tokens.

12 . The edge device of claim 11 , wherein, to train at least one on-device AI model locally available at the edge device using the plurality of concept tokens, the at least one processor is further configured to:

determine at least one relationship among the plurality of concept tokens, and

train the at least one concept representation generator model and the at least one relationship model using the plurality of concept tokens and the determined relationship.

13 . The edge device of claim 12 , wherein, to determine the at least one relationship among the plurality of concept tokens, the at least one processor is further configured to determine at least one of temporal relationships indicating time-based relation among the plurality of concept tokens, positional relationships indicating position and location based relations among the plurality of concept tokens, sequential relationships indicating a sequence of relation among the plurality of concept tokens, or conceptual relationships indicating correlations among plurality of concept tokens.

14 . The edge device of claim 9 , wherein the global AI model comprises at least one global concept representation generator model for generating the universal vector representation for the plurality of concept tokens and at least one global relationship model for indicating the presence of the at least one relationship among the plurality of concept tokens.

15 . The edge device of claim 14 , wherein the at least one processor is further configured to:

receive an input comprising one or more concept tokens, and a request to generate universal vector representation of the one or more concept tokens, and

generate the universal vector representation of the one or more concept tokens by inputting the one or more concept tokens to at least one global concept representation generator model of the global AI model.

16 . A non-transitory computer-readable storage medium, having a computer program stored thereon that performs, when executed by a processor, the method of claim 1 .