Scalable online hierarchical meta-learning

1. A method of meta-learning, comprising:

receiving a prediction objective;

extracting a plurality of subsets of data from a distributed dataset;

generating a plurality of local predictions, by a plurality of local learners, wherein each local prediction is based on a different subset of the plurality of subsets of data and the prediction objective;

combining the plurality of local predictions, by a plurality of meta-learners;

generating a covariance matrix corresponding to a correlation between two local learners of the plurality of local learners;

setting a covariance matrix threshold value;

training the two local learners collaboratively upon determining that an absolute value of the covariance matrix is greater than or equal to the covariance matrix threshold value;

training the two local learners independently upon determining that the absolute value of the covariance matrix is less than the covariance matrix threshold value; and

generating a final prediction, by an ensemble learner, based on the combined local predictions.

2. The method of claim 1 , further comprising:

partitioning the plurality of local learners into a plurality of correlated groups based on respective learning models of each of the plurality of local learners,

wherein local learners having a high correlation with each other are partitioned into a same correlated group, and local learners having a low correlation with each other are partitioned into different correlated groups.

3. The method of claim 2 , further comprising:

training local learners in a same correlated group together; and

training local learners in different correlated groups independently.

4. The method of claim 3 , wherein the plurality of local learners are trained online using a single pass on training data used to train the plurality of local learners.

5. The method of claim 3 , further comprising:

updating the respective learning models of each of the plurality of local learners based on at least one training residual of other local learners in the same correlated group,

wherein the training residual comprises a subset of a learning model of a corresponding local learner.

6. The method of claim 5 , further comprising:

transmitting a corresponding training residual of each local learner to the ensemble learner;

generating updated training data, by the ensemble learner, based on the transmitted training residuals;

transmitting the updated training data to the plurality of local learners; and

generating the final prediction, by the ensemble learner, based on the transmitted training residuals.

7. The method of claim 1 , further comprising:

generating a directed acyclic graph comprising the ensemble learner, the plurality of local learners, and the plurality of meta-learners,

wherein the directed acyclic graph illustrates a hierarchical relationship between the ensemble learners, the plurality of local learners, and the plurality of meta-learners.

8. A method of meta-learning, comprising:

receiving a prediction objective;

extracting a plurality of subsets of data, by a plurality of local learners, from a distributed dataset;

selecting a learning model, by an ensemble learner, for each local learner from among the plurality of local learners;

updating the learning model of each local learner, by the plurality of local learners, based on at least one learning model of a correlated local learner;

generating a covariance matrix corresponding to a correlation between two local learners from among the plurality of local learners;

setting a covariance matrix threshold value;

training the two local learners collaboratively upon determining that an absolute value of the covariance matrix is greater than or equal to the covariance matrix threshold value;

training the two local learners independently upon determining that the absolute value of the covariance matrix is less than the covariance matrix threshold value; and

generating a final prediction, by the ensemble learner, based on data output by each local learner from among the plurality of local learners and the prediction objective.

9. The method of claim 8 , further comprising:

partitioning the plurality of local learners into a plurality of correlated groups based on respective learning models of each of the plurality of local learners,

wherein local learners having a high correlation with each other are partitioned into a same correlated group, and local learners having a low correlation with each other are partitioned into different correlated groups.

10. The method of claim 9 , further comprising:

training local learners in a same correlated group together; and

training local learners in different correlated groups independently.

11. The method of claim 10 , wherein the plurality of local learners are trained online using a single pass on training data used to train the plurality of local learners.

12. The method of claim 10 , further comprising:

updating the respective learning models of each of the plurality of local learners based on at least one training residual of other local learners in the same correlated group,

wherein the training residual comprises a subset of a learning model of a corresponding local learner.

13. The method of claim 12 , further comprising:

transmitting a corresponding training residual of each local learner to the ensemble learner;

generating updated training data, by the ensemble learner, based on the transmitted training residuals;

transmitting the updated training data to the plurality of local learners; and

generating the final prediction, by the ensemble learner, based on the transmitted training residuals.

14. The method of claim 8 , further comprising:

generating a directed acyclic graph comprising the ensemble learner, the plurality of local learners, and the plurality of meta-learners,

wherein the directed acyclic graph illustrates a hierarchical relationship between the ensemble learners, the plurality of local learners, and the plurality of meta-learners.

15. A computer readable storage medium embodying instructions executed by a processor to perform a method of meta-learning, comprising:

receiving a prediction objective;

extracting a plurality of subsets of data from a distributed dataset;

generating a plurality of local predictions, by a plurality of local learners, wherein each local prediction is based on a different subset of the plurality of subsets of data and the prediction objective;

combining the plurality of local predictions, by a plurality of meta-learners;

generating a covariance matrix corresponding to a correlation between two local learners of the plurality of local learners;

setting a covariance matrix threshold value;

training the two local learners collaboratively upon determining that an absolute value of the covariance matrix is greater than or equal to the covariance matrix threshold value;

training the two local learners independently upon determining that the absolute value of the covariance matrix is less than the covariance matrix threshold value; and

generating a final prediction, by an ensemble learner, based on the combined local predictions.

16. The computer readable storage medium of claim 15 , further comprising instructions for:

partitioning the plurality of local learners into a plurality of correlated groups based on respective learning models of each of the plurality of local learners,

wherein local learners having a high correlation with each other are partitioned into a same correlated group, and local learners having a low correlation with each other are partitioned into different correlated groups.

17. The computer readable storage medium of claim 16 , further comprising instructions for:

training local learners in a same correlated group together; and

training local learners in different correlated groups independently.

18. The computer readable storage medium of claim 17 , further comprising instructions for:

updating the respective learning models of each of the plurality of local learners based on at least one training residual of other local learners in the same correlated group,

wherein the training residual comprises a subset of a learning model of a corresponding local learner.

19. A computer system configured to perform a method of meta-learning, comprising:

a memory storing a computer program; and

a processor configured to execute the computer program, wherein the computer program is configured to:

receive a prediction objective;

extract a plurality of subsets of data from a distributed dataset;

generate a plurality of local predictions, by a plurality of local learners, wherein each local prediction is based on a different subset of the plurality of subsets of data and the prediction objective;

combine the plurality of local predictions, by a plurality of meta-learners;

generate a covariance matrix corresponding to a correlation between two local learners of the plurality of local learners;

set a covariance matrix threshold value;

train the two local learners collaboratively upon determining that an absolute value of the covariance matrix is greater than or equal to the covariance matrix threshold value;

train the two local learners independently upon determining that the absolute value of the covariance matrix is less than the covariance matrix threshold value; and

generate a final prediction, by an ensemble learner, based on the combined local predictions.

20. The meta-learning system of claim 19 , wherein the computer program is further configured to:

combine the two local predictions, by at least one meta-learner from among the plurality of meta-learners, with another local prediction generated by another meta-learner; and

generate the final prediction, by the ensemble learner, based on the combined local predictions and the another local prediction.

21. The meta-learning system of claim 19 , wherein the computer program is further configured to:

partition the plurality of local learners into a plurality of correlated groups based on respective learning models of each of the plurality of local learners,

wherein local learners having a high correlation with each other are partitioned into a same correlated group, and local learners having a low correlation with each other are partitioned into different correlated groups.

22. The meta-learning system of claim 19 , wherein the plurality of local learners are trained online using a single pass on training data used to train the plurality of local learners.