Prediction guided sequential data learning method

1. A computerized prediction guided learning method for classification of sequential time-lapse image data, comprising the steps of:

a) inputting a time-lapse image data sequence into a computer memory of a computer hosting a computerized machine learning, wherein the time-lapse image data sequence is delayed by delay buffering the time-lapse image data sequence for a pre-determined period of time to generate a delayed time-lapse image data sequence;

b) performing by a computer program of the computerized machine learning a prediction learning using the time-lapse image data sequence as truth data directly for prediction by the delayed time-lapse image data sequence without explicit labeling data to generate an initial machine learning classifier, wherein the prediction learning is performed by a self-supervised prediction learning to train the initial machine learning classifier using the delayed time-lapse image data sequence as input to predict the time-lapse image data sequence which is effectively the pre-determined period of time ahead of the delayed time-lapse image data sequence;

c) inputting semantic label data into the computer memory; and

d) performing by a computer program of the computerized machine learning a prediction guided semantic learning using the time-lapse image data sequence, the initial machine learning classifier containing learned feature representation and the semantic label data to generate an output machine learning classifier and applying the output machine learning classifier to the time-lapse image data sequence to generate a semantic classification, wherein the semantic classification labels pixels of the time-lapse image data sequence, and wherein the feature representation is fixed in the initial classifier.

2. The computerized prediction guided learning method of claim 1 , wherein the prediction guided semantic learning is selected from a group consisting of deep network, recurrent network, and traditional machine learning methods.

3. The computerized prediction guided learning method of claim 1 , wherein the self-supervised prediction learning comprises a deep network.

4. The computerized prediction guided learning method of claim 1 , wherein the self-supervised prediction learning comprises a recurrent network.

5. The computerized prediction guided learning method of claim 1 , wherein the self-supervised prediction learning includes a traditional machine learning selected from a group consisting of decision tree classifier, random forest classifier, support vector machine, kernel estimator, mixture of Gaussian classifier, and nearest neighbor classifier.

6. The computerized prediction guided learning method of claim 3 , wherein the deep network of the self-supervised prediction learning comprises a fully convolutional network.

7. A computerized prediction guided learning method for classification of sequential time-lapse image data, comprising the steps of:

a) inputting a time-lapse image data sequence into a computer memory of a computer hosting a computerized machine learning, wherein the time-lapse image data sequence is delayed by delay buffering the time-lapse image data sequence for a pre-determined period of time to generate a delayed time-lapse image data sequence;

b) performing by a computer program of the computerized machine learning a prediction learning using the time-lapse image data sequence as truth data directly without explicit labeling data to generate an initial machine learning classifier, wherein the prediction learning is performed by a self-supervised prediction learning to train the initial machine learning classifier using the delayed time-lapse image data sequence as input to predict the time-lapse image data sequence which is effectively the pre-determined period of time ahead of the delayed time-lapse image data sequence;

c) inputting label data into the computer memory; and

d) performing by a computer program of the computerized machine learning a prediction guided update learning using the time-lapse image data sequence, the initial machine learning classifier containing learned feature representation and the label data to generate an output machine learning classifier and applying the output machine learning classifier to the time-lapse image data sequence to generate an image data classification for a targeted classification task, wherein the feature representation is fixed in the initial classifier.

8. The computerized prediction guided learning method of claim 7 , wherein the prediction guided update learning is selected from a group consisting of deep network, recurrent network, and traditional machine learning methods.

9. The computerized prediction guided learning method of claim 7 , wherein the self-supervised prediction learning comprises a deep network.

10. The computerized prediction guided learning method of claim 7 , wherein the self-supervised prediction learning comprises a recurrent network.

11. The computerized prediction guided learning method of claim 7 , wherein the self-supervised prediction learning includes a traditional machine learning selected from a group consisting of decision tree classifier, random forest classifier, support vector machine, kernel estimator, mixture of Gaussian classifier, and nearest neighbor classifier.

12. The computerized prediction guided learning method of claim 9 , wherein the deep network of the self-supervised prediction learning comprises a fully convolutional network.

13. A computerized prediction guided learning method for classification of sequential time-lapse image data, comprising the steps of:

a) inputting a time-lapse image data sequence into a computer memory of a computer hosting a computerized machine learning, wherein the time-lapse image data sequence is delayed by delay buffering the time-lapse image data sequence for a pre-determined period of time to generate a delayed time-lapse image data sequence;

b) performing by a computer program of the computerized machine learning tool a prediction learning using the time-lapse image data sequence as truth data directly without explicit labeling data to generate an initial machine learning classifier, wherein the prediction learning is performed by a self-supervised prediction learning to train the initial machine learning classifier using the delayed time-lapse image data sequence as input to predict the time-lapse image data sequence which is effectively the pre-determined period of time ahead of the delayed time-lapse image data sequence;

c) inputting semantic and label data into the computer memory; and

d) performing by a computer program of the computerized machine learning a prediction guided update and semantic learning using the time-lapse image data sequence, the initial machine learning classifier containing learned feature representation and the semantic and label data to generate an output machine learning classifier, and applying the output machine learning classifier to the time-lapse image data sequence to generate a semantic classification and a data classification for a targeted classification task wherein the semantic classification labels pixels of the time-lapse image data sequence, wherein the feature representation is fixed in the initial classifier.

14. The computerized prediction guided learning method of claim 13 , wherein the prediction guided update and semantic learning is selected from a group consisting of deep network, recurrent network, and traditional machine learning methods.

15. The computerized prediction guided learning method of claim 13 , wherein the self-supervised prediction learning comprises a deep network.

16. The computerized prediction guided learning method of claim 13 , wherein the self-supervised prediction learning comprises a recurrent network.

17. The computerized prediction guided learning method of claim 13 , wherein the self-supervised prediction learning includes a traditional machine learning selected from a group consisting of decision tree classifier, random forest classifier, support vector machine, kernel estimator, mixture of Gaussian classifier, and nearest neighbor classifier.

18. An apparatus for computerized prediction guided learning for classification of sequential data, comprising:

a memory for storing a data sequence and semantic and/or label data, wherein the data sequence is delayed for a pre-determined period of time to generate a delayed data sequence;

a computerized machine learning for performing a prediction learning using the data sequence and the delayed data sequence without explicit labeling data to generate an initial machine learning classifier, wherein the prediction learning is performed by a self-supervised prediction learning to train the initial machine learning classifier using the delayed data sequence as input to predict the data sequence which is effectively the pre-determined period of time ahead of the delayed data sequence, and performing a prediction guided semantic learning using the data sequence, the initial machine learning classifier containing learned feature representation and the semantic and/or label data to generate an output machine learning classifier and semantic and/or classification, wherein the feature representation is fixed in the initial classifier,

wherein the self-supervised prediction learning is implemented by a deep network or a recurrent network including an input layer, a plurality of hidden layers, and an output layer,

wherein in a training phase, the data sequence is processed in a feedback way by the output layer, and then by the plurality of hidden layers, and then by the input layer to update parameters, and in a classification phase, the delayed data sequence is processed in a feed- forward way by the input layer, and then by the plurality of hidden layers, and then by the output layer to generate the initial machine learning classifier.

19. The computerized prediction guided learning method of claim 18 , wherein the deep network of the self-supervised prediction learning comprises a fully convolutional network.