De-centralised learning for re-indentification
A method for generating an optimised domain-generalisable model for re-identification of a target in a set of candidate images. The method optimises a local feature embedding model for domain-specific feature representation at each client of a plurality of clients, then receives, at a central server, information on changes to the local feature embedding model at each respective client resulting from the optimising step, and then updates a global feature embedding model based on the changes to the local feature embedding model. The method further receives, at each client from the central server, information representative of the updates to the global feature embedding model, then maps, at each client, on to the respective local feature embedding model at least a portion of the received updates, and subsequently updates, at each client, the respective local feature embedding model based on the mapped updates. The steps are repeated until convergence criteria are met, wherein the global feature embedding model is the optimised domain-generalisable model for re-identification of a target in a set of candidate images.
1. A method for generating an optimised domain-generalizable model for re-identification of a target in a set of candidate images, comprising:
optimising a local feature embedding model for domain-specific feature representation at each client of a plurality of clients, the local feature embedding model at each client of a plurality of clients optimised for a respective data set associated with a domain of each client of the plurality of clients;
receiving, at a central server from each client of the plurality of clients, information on changes to the local feature embedding model at each respective client resulting from the optimising step;
updating, at the central server, a global feature embedding model for domain-generalizable feature representation based on the changes to the local feature embedding model at each respective client of at least a subset of the plurality of clients;
receiving, at each client of the plurality of clients from the central server, information representative of the updates to the global feature embedding model;
mapping, at each client of the plurality of clients, on to the respective local feature embedding model, at least a portion of the received updates to the global feature embedding model;
updating, at each client of the plurality of clients, the respective local feature embedding model based on the mapped updates; and
repeating each of the previous steps until convergence criteria are met for the optimisation of the local feature embedding model at each of the plurality of local clients, wherein the global feature embedding model is the optimised domain-generalizable model for re-identification of a target image in a set of candidate images.
2. The method of claim 1 , wherein the respective data set associated with a domain of each client of the plurality of clients is an independent data set.
3. The method of claim 2 , wherein each independent data set is non-overlapping.
4. The method of claim 1 , wherein updating the global feature embedding model based on the changes to at least a subset of the local feature embedding model at each respective client, comprises:
selecting a subset of the plurality of clients;
aggregating, at the central server, the information on changes to the local feature embedding model received from each client of the selected subset of the plurality of client to determine the aggregate changes; and
updating the global feature embedding model based on the aggregate changes.
5. The method of claim 4 , wherein aggregating the information on changes to the local feature embedding model received from each client of the selected subset of the plurality of client, comprises:
averaging the information on changes to the local feature embedding model received from each client of the selected subset of the plurality of clients.
6. The method of claim 4 , wherein, prior to aggregating the information, the further comprises:
applying white noise to the information on changes to the local feature embedding model at each respective client of the selected subset of the plurality of clients.
7. The method of claim 1 , wherein receiving, at each client of the plurality of clients from the central server, information representative of the updates to the global feature embedding model comprises receiving, at each client of the plurality of clients from the central server, the global feature embedding model.
8. The method of claim 7 , wherein mapping, at each client of the plurality of clients, on to the respective local feature embedding model, at least a portion of the received updates to the global feature embedding model, further comprises:
determining, at each local client of the plurality of clients, a probability distribution for the respective local feature embedding model applied to the data set associated with a domain of the respective client;
determining, at each local client of the plurality of clients, a probability distribution for the global feature embedding model applied to the data set associated with a domain of the respective client;
determining, at each local client of the plurality of clients, a divergence between the probability distribution for the respective local feature embedding model and the probability distribution for the global feature embedding model;
based on the determined divergence, identifying, at each local client of the plurality of clients, the updates to the global feature embedding model that are relevant to the respective local feature embedding model; and
updating, at each local client of the plurality of clients, the respective local feature embedding model based on the identified relevant updates to the global feature embedding model.
9. The method of claim 1 , wherein after convergence criteria are met for the optimisation of the local feature embedding model at each of the plurality of local clients, the method further comprises:
applying the global feature embedding model to characterize a target;
applying the global feature embedding model to characterize each image of a set of candidate images; and
using a distance metric to identify the candidate images having sufficient similarity to the target image.
10. The method of claim 1 , further comprising:
introducing a further client to the plurality of clients;
setting the local feature embedding model at the further client as the global feature embedding model; and
repeating each of the steps of the method until the convergence criteria are met for the optimisation of the local feature embedding model at each of the plurality of local clients, including the further client.
11. A system for generating an optimised domain-generalizable model for re-identification of a target image in a set of candidate images, comprising:
a central server, hosting a global feature embedding model for domain-generalizable feature representation; and
a plurality of clients, each client hosting a local feature embedding model for domain-specific feature representation, each client of the plurality of clients having access to a respective data set associated with a domain of each client of the plurality of clients;
the central server being configured to:
receive, from each client of the plurality of clients, information on changes to the local feature embedding model at each respective client;
update the global feature embedding model for feature representation, based on the changes to the local feature embedding model at each respective client of at least a subset of the plurality of clients;
send, to each of the clients of the plurality of clients, information representative of the updates to the global feature embedding model; and
each client of the plurality of clients being configured to:
optimise the local feature embedding model for the respective data set associated with the domain of each client of the plurality of clients;
send, to the central server, information on changes to the local feature embedding model;
receive, from the central server, information representative of the updates to the global feature embedding model;
map on to the local feature embedding model, at least a portion of the received updates to the global feature embedding model; and
update the local feature embedding model based on the mapped updates.
12. The system of claim 11 , wherein the respective data set associated with a domain of each client of the plurality of clients is an independent data set.
13. The system of claim 11 , wherein the central server being configured to update the global feature embedding model based on the changes to at least a subset of the local feature embedding model at each respective client, comprises the central server being configured to:
select a subset of the plurality of clients;
aggregate the information on changes to the local feature embedding model received from each client of the selected subset of the plurality of client to determine the aggregate changes; and
update the global feature embedding model based on the aggregate changes.
14. The system of claim 13 , wherein the central server being configured to aggregate the information comprises the central server being configured to:
average the information on changes to the local feature embedding model received from each client of the selected subset of the plurality of clients.
15. The system of claim 13 , wherein each of the local clients is further configured to:
apply white noise to the information on changes to the local feature embedding model, prior to sending to the central server.
16. The system of claim 11 , wherein each client of the plurality of clients being configured to receive, from the central server, information representative of the updates to the global feature embedding model comprises:
each client of the plurality of clients being configured to receive the global feature embedding model.
17. The system of claim 11 , wherein each client of the plurality of clients being configured to map on to the local feature embedding model, at least a portion of the received updates to the global feature embedding model, comprises each client of the plurality of clients being configured to:
determine a probability distribution for the respective local feature embedding model applied to the data set associated with a domain of the respective client;
determine a probability distribution for the global feature embedding model applied to the data set associated with a domain of the respective client;
determine a divergence between the probability distribution for the respective local feature embedding model and the probability distribution for the global feature embedding model;
based on the determined divergence, identify the updates to the global feature embedding model that are relevant to the local feature embedding model; and
update the local feature embedding model based on the identified relevant updates to the global feature embedding model.
18. The system of claim 11 , the central server is further configured to:
deploy the global feature embedding model to a customer client;
wherein the customer client is configured to:
apply the global feature embedding model to characterize a target;
apply the global feature embedding model to characterize each image of a set of candidate images associated with the domain of the customer client; and
use a distance metric to identify the candidate images having sufficient similarity to the target image.