Methods and apparatus for incremental learning using stored features

1 . A computerized method for incrementally training a classifier using stored sets of representative features, the method comprising:

accessing a first set of training images for a first class;

processing the first set of training images, using a feature extraction technique, to determine a set of features for the first class;

determining, using a selection technique, a set of representative features from the set of features for the first class, wherein said determination comprises:

determining, based on the set of features, a mean of the features for the first class, and

determining, based on the mean, the set of representative features, wherein the set of representative features is a subset of the set of features;

training the classifier using the set of representative features for the first class;

adding the set of representative features for the first class to an exemplar set of representative features for a plurality of classes;

accessing a second set of training images for a second class; and

re-training the classifier using the second set of training images and at least part of the exemplar set of representative features.

2 . The method of claim 1 , wherein re-training the classifier using the second set of training images and at least part of the exemplar set of representative features comprises:

generating a first modified representative feature for the first class, comprising:

selecting a first representative feature in the set of representative features for the first class;

determining a noise component for the first representative feature; and

generating, based on the first representative feature and the noise component, the first modified representative feature.

3 . The method of claim 2 , further comprising generating a second modified representative feature, comprising:

selecting a second representative feature from the set of representative features for the first class; and

determining the second modified representative feature based on a difference between values of the first representative feature and the second representative feature.

4 . The method of claim 3 ,

wherein the method further comprises:

determining, using the second set of training images, a second set of features for the second class; and

re-training the classifier comprises training the classifier using:

the second set of features; and

at least one of the first representative feature, the first modified representative feature, and the second modified representative feature.

5 . The method of claim 4 , further comprising:

determining a second set of representative features from the second set of features for the second class; and

adding the second set of representative features for the second class to the exemplar set.

6 . The method of claim 1 , further comprising executing a testing phase, comprising:

receiving a new image;

determining a new set of features for the new image;

executing, using the new set of features, the classifier to generate a first set of predictions for the new image;

executing, using the exemplar set of representative features, a machine learning model to generate a second set of predictions for the new image; and

determining a predicted class for the new image based on the first set of predictions and the second set of predictions.

7 . The method of claim 6 , wherein determining the predicted class for the new image comprises:

generating, based on the first set of predictions and the second set of predictions, a final set of predictions; and

determining the predicted class based on the final set of predictions.

8 . The method of claim 6 , wherein executing the machine learning model using the exemplar set of representative features comprises:

accessing a plurality of sets of stored representative features in the exemplar set, wherein each set of stored representative features is associated with an associated class and an associated training step in which the classifier was re-trained; and

executing the machine learning model using the plurality of sets of stored representative features to generate the second set of predictions for the new image.

9 . The method of claim 8 , wherein:

generating the first set of predictions comprises generating, for each class of the plurality of sets of stored representative features, a first value that is indicative of a prediction of whether the new set of features belongs to the class; and

generating the second set of predictions comprises generating, for each step of the plurality of sets of stored representative features, a second value that is indicative of a prediction of whether the new set of features belongs to the step.

10 . The method of claim 7 , wherein generating the final set of predictions comprises:

determining a weighting factor based on (a) a maximum prediction of the second set of predictions and a minimum prediction of the second set of predictions and (b) a normalization constant;

adjusting the second set of predictions based on the weighting factor; and

adding the first set of predictions to the adjusted second set of predictions.

11 . A non-transitory computer-readable media comprising instructions for incrementally training a classifier using stored sets of representative features, wherein the instructions, when executed by one or more processors on a computing device, are operable to cause the one or more processors to:

access a first set of training images for a first class;

process the first set of training images, using a feature extraction technique, to determine a set of features for the first class;

determine, using a selection technique, a set of representative features from the set of features for the first class, wherein said determination comprises:

determining, based on the set of features, a mean of the features for the first class, and

determining, based on the mean, the set of representative features, wherein the set of representative features is a subset of the set of features;

train the classifier using the set of representative features for the first class;

add the set of representative features for the first class to an exemplar set of representative features for a plurality of classes;

access a second set of training images for a second class; and

re-train the classifier using the second set of training images and at least part of the exemplar set of representative features.

12 . The non-transitory computer-readable media of claim 11 , wherein the instructions are further operate to cause the one or more processors to:

generate a first modified representative feature for the first class by:

selecting a first representative feature in the set of representative features for the first class;

determining a noise component for the first representative feature; and

generating, based on the first representative feature and the noise component, the first modified representative feature.

13 . The non-transitory computer-readable media of claim 12 , wherein the instructions are further operable to cause the one or more processors to:

generate a second modified representative feature by:

selecting a second representative feature from the set of representative features for the first class; and

determining the second modified representative feature based on a difference between values of the first representative feature and the second representative feature.

14 . The non-transitory computer-readable media of claim 13 , wherein the instructions are further operable to cause the one or more processors to:

determine, using the second set of training images, a second set of features for the second class; and

wherein re-training the classifier comprises training the classifier using:

the second set of features; and

at least one of the first representative feature, the first modified representative feature, and the second modified representative feature.

15 . The non-transitory computer-readable media of claim 14 , wherein the instructions are further operable to cause the one or more processors to:

determine a second set of representative features from the second set of features for the second class; and

add the second set of representative features for the second class to the exemplar set.

16 . The non-transitory computer-readable media of claim 11 , wherein the instructions are further operable to cause the one or more processors to execute a testing phase, comprising:

receiving a new image;

determining a new set of features for the new image;

executing, using the new set of features, the classifier to generate a first set of predictions for the new image;

executing, using the exemplar set of representative features, a machine learning model to generate a second set of predictions for the new image; and

determining a predicted class for the new image based on the first set of predictions and the second set of predictions.

17 . The non-transitory computer-readable media of claim 16 , wherein determining the predicted class for the new image comprises:

generating, based on the first set of predictions and the second set of predictions, a final set of predictions; and

determining the predicted class based on the final set of predictions.

18 . The non-transitory computer-readable media of claim 16 , wherein executing the machine learning model using the exemplar set of representative features comprises:

accessing a plurality of sets of stored representative features in the exemplar set, wherein each set of stored representative features is associated with an associated class and an associated training step in which the classifier was re-trained; and

executing the machine learning model using the plurality of sets of stored representative features to generate the second set of predictions for the new image.

19 . The non-transitory computer-readable media of claim 18 , wherein:

generating the first set of predictions comprises generating, for each class of the plurality of sets of stored representative features, a first value that is indicative of a prediction of whether the new set of features belongs to the class; and

generating the second set of predictions comprises generating, for each step of the plurality of sets of stored representative features, a second value that is indicative of a prediction of whether the new set of features belongs to the step.

20 . The non-transitory computer-readable media of claim 17 , wherein generating the final set of predictions comprises:

determining a weighting factor based on (a) a maximum prediction of the second set of predictions and a minimum prediction of the second set of predictions and (b) a normalization constant;

adjusting the second set of predictions based on the weighting factor; and

adding the first set of predictions to the adjusted second set of predictions.