Feature selection for retraining classifiers

1. A method of managing memory usage of an artificial neural network, comprising:

receiving, after training the artificial neural network with a first set of existing training samples of an existing training set, a new training sample at a device, the new training sample having a same class as the first set of existing training samples in the existing training set stored in a memory of the device;

calculating at least one of a first similarity metric or a second similarity metric for the new training sample, wherein the first similarity metric is associated with a first distance between the new training sample and the first set of existing training samples of the same class as the new training sample, and wherein the second similarity metric is associated with a second distance between the new training sample and a second set of existing training samples, in the existing training set, of a different class than the new training sample;

selectively updating the first set of existing training samples to include the new training sample based on the first similarity metric or the second similarity metric, wherein updating the first set of existing training samples comprises:

selecting an existing training sample from the first set of existing training samples for removal from the memory when updating the first set of existing training samples would cause the memory to exceed a capacity threshold, the selected existing training sample having a smallest sum distance to each point of the first set of existing training samples of the different class in comparison to a sum distance for other existing training samples of the first set of existing training samples; and

removing the selected existing training sample from the memory; and

retraining the artificial neural network with the updated first set of existing training samples.

2. The method of claim 1 , in which the existing training sample is selected based on a bias for retaining the first set of existing training samples according to a proximity to the second set of existing training samples.

3. The method of claim 1 , wherein at least one of the first similarity metric or the second similarity metric are computed based at least in part on a centroid of the first set of existing training samples.

4. The method of claim 1 , wherein the first similarity metric comprises a difference between: a minimum distance between any two points of the first set of existing training samples, and a minimum distance between points of the new training sample and all training samples of the first set of existing training samples.

5. The method of claim 1 , wherein the first similarity metric comprises a difference between: a maximum summed distance between any two points of the first set of existing training samples, and a maximum summed distance between points of the new training sample and all training samples of the first set of existing training samples.

6. The method of claim 1 , wherein:

the first similarity metric comprises a maximum summed distance between each point of the first set of existing training samples and all other sets of existing training samples in other classes,

the second similarity metric comprises a maximum summed distance between points of the new training sample and all other sets of existing training samples in the other classes, and

the new training sample is selectively added to the first set of existing training samples based at least in part on a difference between the first similarity metric and the second similarity metric.

7. An apparatus for managing memory usage of an artificial neural network, comprising:

means for receiving, after training the artificial neural network with a first set of existing training samples of an existing training set, a new training sample at the apparatus, the new training sample having a same class as the first set of existing training samples in the existing training set stored in a memory of the apparatus;

means for calculating at least one of a first similarity metric or a second similarity metric for the new training sample, wherein the first similarity metric is associated with a first distance between the new training sample and the first set of existing training samples of the same class as the new training sample, and wherein the second similarity metric is associated with a second distance between the new training sample and a second set of existing training samples, in the training set, of a different class than the new training sample;

means for selectively updating the first set of existing training samples to include the new training sample based on the first similarity metric or the second similarity metric, wherein updating the first set of existing training samples comprises:

selecting an existing training sample from the first set of existing training samples for removal from the memory when updating the first set of existing training samples would cause the memory to exceed a capacity threshold, the selected existing training sample having a smallest sum distance to each point of the first set of existing training samples of the different class in comparison to a sum distance for other existing training samples of the first set of existing training samples; and

removing the selected existing training sample from the memory; and

means for retraining the artificial neural network with the updated first set of existing training samples.

8. The apparatus of claim 7 , in which the existing training sample is selected based on a bias for retaining the first set of existing training samples according to a proximity to the second set of existing training samples.

9. The apparatus of claim 7 , wherein at least one of the first similarity metric or the second similarity metric are computed based at least in part on a centroid of the first set of existing training samples.

10. The apparatus of claim 7 , wherein the first similarity metric comprises a difference between: a minimum distance between any two points of the first set of existing training samples, and a minimum distance between points of the new training sample and all training samples of the first set of existing training samples.

11. The apparatus of claim 7 , wherein the first similarity metric comprises a difference between: a maximum summed distance between any two points of the first set of existing training samples, and a maximum summed distance between points of the new training sample and all training samples of the first set of existing training samples.

12. The apparatus of claim 7 , wherein:

the first similarity metric comprises a maximum summed distance between each point of the first set of existing training samples and all other sets of existing training samples in other classes,

the second similarity metric comprises a maximum summed distance between points of the new training sample and all other sets of existing training samples in the other classes, and

the new training sample is selectively added to the first set of existing training samples based at least in part on a difference between the first similarity metric and the second similarity metric.

13. An apparatus for managing memory usage of an artificial neural network, comprising:

a memory; and

at least one processor coupled to the memory and configured:

to receive, after training with a first set of existing training samples of an existing training set, a new training sample at the apparatus, the new training sample having a same class as the first set of existing training samples in the existing training set stored in a memory of the apparatus;

to calculate at least one of a first similarity metric or a second similarity metric for the new training sample, wherein the first similarity metric is associated with a first distance between the new training sample and the first set of existing training samples of the same class as the new training sample, and wherein the second similarity metric is associated with a second distance between the new training sample and a second set of existing training samples, in the training set, of a different class than the new training sample;

to selectively update the first set of existing training samples to include the new training sample based on the first similarity metric or the second similarity metric, wherein updating the first set of existing training samples comprises:

selecting an existing training sample from the first set of existing training samples for removal from the memory when updating the first set of existing training samples would cause the memory to exceed a capacity threshold, the selected existing training sample having a smallest sum distance to each point of the first set of existing training samples of the different class in comparison to a sum distance for other existing training samples of the first set of existing training samples; and

removing the selected existing training sample from the memory; and

to retrain the artificial neural network with the updated first set of existing training samples.

14. The apparatus of claim 13 , in which the existing training sample is selected based on a bias for retaining the first set of existing training samples according to a proximity to the second set of existing training samples.

15. The apparatus of claim 13 , wherein the at least one processor is further configured to calculate at least one of the first similarity metric or the second similarity metric based at least in part on a centroid of the first set of existing training samples.

16. The apparatus of claim 13 , wherein the first similarity metric comprises a difference between: a minimum distance between any two points of the first set of existing training samples, and a minimum distance between points of the new training sample and all training samples of the first set of existing training samples.

17. The apparatus of claim 13 , wherein the first similarity metric comprises a difference between: a maximum summed distance between any two points of the first set of existing training samples, and a maximum summed distance between points of the new training sample and all training samples of the first set of existing training samples.

18. The apparatus of claim 13 , wherein:

the first similarity metric comprises a maximum summed distance between each point of the first set of existing training samples and all other sets of existing training samples in other classes,

the second similarity metric comprises a maximum summed distance between points of the new training sample and all other sets of existing training samples in the other classes, and

the at least one processor being configured to update the first set of existing training samples to include the new training sample based at least in part on a difference between the first similarity metric and the second similarity metric.

19. A non-transitory computer-readable medium having program code recorded thereon for managing memory usage of an artificial neural network, the program code being executed by a processor and comprising:

program code to receive, after training the artificial neural network with a first set of existing training samples of an existing training set, a new training sample at a device, the new training sample having a same class as the first set of existing training samples in the existing training set stored in a memory of the device;

program code to calculate at least one of a first similarity metric or a second similarity metric for the new training sample, wherein the first similarity metric is associated with a first distance between the new training sample and the first set of existing training samples of the same class as the new training sample, and wherein the second similarity metric is associated with a second distance between the new training sample and a second set of existing training samples, in the training set, of a different class than the new training sample;

program code to selectively update the first set of existing training samples to include the new training sample based on the first similarity metric or the second similarity metric, wherein updating the first set of existing training samples comprises:

selecting an existing training sample from the first set of existing training samples for removal from the memory when updating the first set of existing training samples would cause the memory to exceed a capacity threshold, the selected existing training sample having a smallest sum distance to each point of the first set of existing training samples of the different class in comparison to a sum distance for other existing training samples of the first set of existing training samples; and

removing the selected existing training sample from the memory; and

program code to retrain the artificial neural network with the updated first set of existing training samples.

20. The computer-readable medium of claim 19 , in which the existing training sample is selected based on a bias for retaining the first set of existing training samples according to a proximity to the second set of existing training samples.

21. The computer-readable medium of claim 19 , further comprising program code to calculate at least one of the first similarity metric or the second similarity metric which are computed based at least in part on a centroid of the first set of existing training samples.

22. The computer-readable medium of claim 19 , wherein the first similarity metric comprises a difference between: a minimum distance between any two points of the first set of existing training samples, and a minimum distance between points of the new training sample and all training samples of the first set of existing training samples.

23. The computer-readable medium of claim 19 , wherein the first similarity metric comprises a difference between: a maximum summed distance between any two points of the first set of existing training samples, and a maximum summed distance between points of the new training sample and all training samples of the first set of existing training samples.

24. The computer-readable medium of claim 19 , wherein:

the first similarity metric comprises a maximum summed distance between each point of the first set of existing training samples and all other sets of existing training samples in other classes,

the second similarity metric comprises a maximum summed distance between points of the new training sample and all other sets of existing training samples in the other classes, and

further comprising program code to selectively update the first set of existing training samples to include the new training sample based at least in part on a difference between the first similarity metric and the second similarity metric.