Real time feedback from a machine learning system
A technique for providing real time feedback from a machine learning system is provided that includes a method and system for interactively training machine learning models. In particular, by separating processing and analysis using static and dynamic models that are trained differently, the disclosed technique enables interactive training and prediction of machine learning models to increase the speed of generating new predictions based on real time feedback. In some cases, a dynamic model is applied to the output of a static model to generate an analysis, a correction of the analysis is received, and the correction is used to retrain the dynamic model. An updated analysis is generated based on reapplying the dynamic model to the output of the static model without having to retrain the static model.
1. A computer system for providing real time feedback from a machine learning system, comprising:
a processor; and
a memory coupled with the processor, wherein the memory is configured to provide the processor with instructions which when executed cause the processor to:
define a target field based on a first interaction of a user with the computer system;
access a corpus of files to be analyzed, the corpus of files being identified by the user;
apply a static machine learning model to at least one of the files to generate an output;
automatically apply a dynamic learning model to the output of the static machine learning model to generate an analysis of the output;
determine a correction of the analysis, the correction being based on a second interaction of the user with the computer system;
retrain the dynamic machine learning model based on the correction; and
generate an updated analysis of other files in the corpus of files based at least in part on reapplying the dynamic machine learning model to respective outputs of the static machine learning model for each of the other files of the corpus.
2. The system of claim 1 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed.
3. The system of claim 1 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files and storing the normalized files.
4. The system of claim 1 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files, storing the normalized files, and running an algorithm on the normalized files.
5. The system of claim 1 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files, storing the normalized files, and running an algorithm on the normalized files, wherein running the algorithm depends on the file type.
6. The system of claim 1 wherein the analysis of the file is based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file.
7. The system of claim 1 wherein the analysis of the file comprises a predicted specification based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file.
8. The system of claim 1 wherein the analysis of the file comprises a predicted specification so that the analyzed file becomes a predicted file, wherein the predicted specification is based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file, and wherein:
determining the correction comprises receiving a ground truth specification for the predicted file so that the predicted file becomes a new ground truth file;
retraining the dynamic machine learning model based on the correction comprises retraining the dynamic machine learning model in response to receiving the ground truth specification for the predicted file; and
generating an updated analysis of other files in the corpus of files based at least in part on reapplying the dynamic machine learning model to the respective outputs of the static machine learning model for each of the other files of the corpus comprises receiving, from the retrained dynamic machine learning model, a predicted specification of the target field for at least one of the other files in the corpus of files so that the file becomes a predicted file based at least in part on the retrained dynamic machine learning model.
9. A computer program product for providing real time feedback from a machine learning system, the computer program product being embodied in a tangible non-transitory computer readable storage medium and comprising computer instructions for execution by a computer system to perform operations comprising:
defining a target field based on a first interaction of a user with the computer system;
accessing a corpus of files to be analyzed, the corpus of files being identified by the user;
applying a static machine learning model at least one of the files to generate an output;
automatically applying a dynamic learning model to the output file of the static machine learning model to generate an analysis of the output;
determining a correction of the analysis, the correction being based on a second interaction of the user;
retraining the dynamic machine learning model based on the correction; and
generating an updated analysis of other files in the corpus of files based at least in part on reapplying the dynamic machine learning model to respective outputs of the static machine learning model for each of the other files of the corpus.
10. A method for operating a computer system, the method comprising:
defining a target field based on a first interaction of a user with the computer system;
accessing a corpus of files to be analyzed, the corpus of files being identified by the user;
applying a static machine learning model at least one of the files to generate an output;
automatically apply a dynamic learning model to the output of the static machine learning model to generate an analysis of the output;
determining a correction of the analysis, the correction being based on a second interaction of the user;
retraining the dynamic machine learning model based on the correction; and
generating an updated analysis of other files in the corpus of files based at least in part on reapplying the dynamic machine learning model to respective outputs of the static machine learning model for each of the other files of the corpus.
11. The method of claim 10 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed.
12. The method of claim 10 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files and storing the normalized files.
13. The method of claim 10 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files, storing the normalized files, and running an algorithm on the normalized files.
14. The method of claim 10 wherein applying a static machine learning model comprises pre-processing the corpus of files to be analyzed including by normalizing the files, storing the normalized files, and running an algorithm on the normalized files, wherein running the algorithm depends on the file type.
15. The method of claim 10 wherein the analysis of the file is based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file.
16. The method of claim 10 wherein the analysis of the file comprises a predicted specification based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file.
17. The method of claim 10 wherein the analysis of the file comprises a predicted specification so that the analyzed file becomes a predicted file, wherein the predicted specification is based at least in part on a ground truth file generated in response to receiving a ground truth specification of the target field for the ground truth file, and wherein:
determining the correction comprises receiving a ground truth specification for the predicted file so that the predicted file becomes a new ground truth file;
retraining the dynamic machine learning model based on to the correction comprises retraining the dynamic machine learning model in response to receiving the ground truth specification for the predicted file; and
generating an updated analysis of other files in the corpus of files based at least in part on reapplying the dynamic machine learning model to the respective outputs of the static machine learning model for each of the other files of the corpus comprises receiving, from the retrained dynamic machine learning model, a predicted specification of the target field for at least one of the other files in the corpus of files so that the file becomes a predicted file based at least in part on the retrained dynamic machine learning model.