Machine learning platform for performing large scale data analytics
To address problems that video imaging systems and platforms face when analyzing image and video content for detection and feature extraction, a solution is provided in which accumulating significant amounts of data suitable for training and learning analytics is leveraged to improve over time, the classifiers used to perform the detection and feature extraction, by employing a larger search space and generate additional and more complex classifiers through distributed processing. A distributed learning platform is therefore provided, which is configured for operating on large scale data, in a true big data paradigm. The learning platform is operable to empirically estimate a set of optimal feature vectors and a set of discriminant functions using a parallelizable learning algorithm. A method of adding new data into a database utilized by such a learning platform is also provided. The method comprises identifying an unrepresented sample space; determining new data samples associated with the unrepresented sample space; and adding the new data samples to the database.
1. A distributed learning platform comprising at least one processor and memory, the memory comprising computer executable instructions for:
interfacing the learning platform with a video analysis environment, the video analysis environment being configured to collect data from a plurality of video capture devices in a connected system, to perform a feature analysis on the collected data using at least one classifier to detect and extract features in the video data, and to perform a validation of results of the feature analyses, wherein the validation of results comprises human feedback for one or more of the feature analyses;
obtaining feature analysis and validation results populated in a database by the video analysis environment continually over time;
for new data being added in the database:
identifying an unrepresented sample space;
determining new data samples associated with the unrepresented sample space; and
adding the new data samples to the database;
using the feature analysis and validation results to periodically or continually retrain the at least one classifier in the video analysis environment, comprising:
obtaining validated data having been subjected to at least one feature analysis;
applying at least one pre-processing operation on the validated data;
determining a set of positive samples, a set of negative samples, and a set of features to be analyzed;
creating parallel jobs in a distributed computing environment having a plurality of computing resources;
aggregating results from the plurality of computing resources; and
analyzing the aggregated results to retrain the at least one classifier; and
updating the at least one classifier in the video analysis environment to improve subsequent feature analyses.
2. The learning platform of claim 1 , wherein the platform is further configured for real-world feedback interactivity for selecting a reduced set of samples from a complete set of samples.
3. The learning platform of claim 1 , wherein the platform is configured to enable human intervention to enhance supervision.
4. The learning platform of claim 1 , wherein the parallel jobs comprises a boosting based optimization process.
5. The learning platform of claim 1 , wherein the parallel jobs estimate a set of optimal feature vectors and a set of discriminant functions using a parallelizable learning algorithm.
6. The learning platform of claim 5 , wherein the estimated feature vectors and set of discriminant functions are used for scale invariant multi-class object classification.
7. The learning platform of claim 5 , wherein a feature set is derivable using a bank of non-symmetric kernels.
8. The learning platform of claim 7 , wherein a number of kernels is an exhaustive set of possible offsets and scales.
9. The learning platform of claim 5 , further configured to use at least one unsupervised method to identify relevant data for the learning algorithm.
10. The learning platform of claim 1 , further operable to update previously estimated parameters based upon new data.
11. The learning platform of claim 1 , comprising access to a number of computing resources allowing for processing of a large feature space.
12. The learning platform of claim 1 , further operable for producing a classifier that is operable in real-time through feature dimensionality reduction.
13. The learning platform of claim 1 , wherein the unrepresented sample space is determined by identifying edge cases where the classifier used by the learning platform is insufficient to separate classes.
14. A method of performing distributed learning on a distributed learning platform, the method comprising:
interfacing the learning platform with a video analysis environment, the video analysis environment being configured to collect data from a plurality of video capture devices in a connected system, to perform a feature analysis on the collected data using at least one classifier to detect and extract features in the video data, and to perform a validation of results of the feature analyses, wherein the validation of results comprises human feedback for one or more of the feature analyses;
obtaining feature analysis and validation results populated in a database by the video analysis environment continually over time;
for new data being added in the database:
identifying an unrepresented sample space;
determining new data samples associated with the unrepresented sample space; and
adding the new data samples to the database;
using the feature analysis and validation results to periodically or continually retrain the at least one classifier in the video analysis environment, comprising:
obtaining validated data having been subjected to at least one feature analysis;
applying at least one pre-processing operation on the validated data;
determining a set of positive samples, a set of negative samples, and a set of features to be analyzed;
creating parallel jobs in a distributed computing environment having a plurality of computing resources;
aggregating results from the plurality of computing resources; and
analyzing the aggregated results to retrain the at least one classifier; and
updating the at least one classifier in the video analysis environment to improve subsequent feature analyses.
15. The method of claim 14 , further comprising enabling real-world feedback interactivity for selecting a reduced set of samples from a complete set of samples.
16. The method of claim 14 , further comprising enabling human intervention to enhance supervision.
17. The method of claim 14 , wherein the parallel jobs comprise a boosting based optimization process.
18. The method of claim 14 , wherein the parallel jobs estimate a set of optimal feature vectors and a set of discriminant functions using a parallelizable learning algorithm.
19. The method of claim 18 , wherein the estimated feature vectors and set of discriminant functions are used for scale invariant multi-class object classification.
20. The method of claim 18 , wherein a feature set is derivable using a bank of non-symmetric kernels.
21. The method of claim 20 , wherein a number of kernels is an exhaustive set of possible offsets and scales.
22. The method of claim 18 , further comprising using at least one unsupervised method to identify relevant data for the learning algorithm.
23. The method of claim 14 , further comprising updating previously estimated parameters based upon new data.
24. The method of claim 14 , comprising accessing a number of computing resources allowing for processing of a large feature space.
25. The method of claim 14 , further comprising producing a classifier that is operable in real-time through feature dimensionality reduction.
26. The method of claim 14 , wherein the unrepresented sample space is determined by identifying edge cases where the classifier used by the learning platform is insufficient to separate classes.