Data set cleaning for artificial neural network training
A technology for cleaning a training data set for a neural network using dirty training data starts by accessing a labeled training data set that includes relatively dirty labeled data elements. The labeled training data set is divided into a first subset A and a second subset B. The procedure includes cycling between the subsets A and B, including producing refined model-filtered subsets of subsets A and B to provide a cleaned data set. Each refined model-filtered subset can have improved cleanliness and increased numbers of elements.
1. A computer-implemented method for cleaning training data for a neural network, comprising:
accessing a labeled training data set;
using a first subset of the labeled training data set to train a first model of the neural network;
filtering a second subset of the labeled training data set using the first model to provide a first model-filtered subset of the second subset;
using the first model-filtered subset of the second subset to train a first refined model of the neural network;
filtering the first subset using the first refined model to provide a first refined model-filtered subset of the first subset;
using the first refined model-filtered subset of the first subset to train a second refined model of the neural network; and
filtering the second subset of the labeled training data set using the second refined model to provide a second refined model-filtered subset of the second subset.
2. The method of claim 1 , including:
combining the first refined model-filtered subset of the first subset and the second refined model-filtered subset of the second subset to provide a filtered training set, training an output model of a target neural network using the filtered training set, and saving the output model in memory.
3. The method of claim 1 , wherein the second refined model-filtered subset has a greater number of elements than the first model-filtered subset.
4. The method of claim 1 , wherein the first subset and the second subset do not overlap.
5. The method of claim 1 , wherein said filtering the first subset using the first refined model includes:
executing the neural network using the first refined model over the first subset to produce classification data classifying data elements of the first subset; and
selecting data elements of the first subset having labels matching the classification data to provide the first refined model-filtered subset of the first subset.
6. The method of claim 2 , further including loading the output model in an instance of the target neural network in an inference engine.
7. The method of claim 1 , including iteratively:
(i) using a previously provided refined model-filtered subset of one of the first subset and the second subset to train an instant refined model of the neural network;
(ii) filtering another of the first subset and the second subset using the instant refined model to provide an instant refined model-filtered subset of the other of the first subset and the second subset; and
(iii) determining whether an iteration criterion is met, and if not, then executing (i) to (iii), and if so, then using a combination of a selected one of the refined model-filtered subsets of the first subset and a selected one of the refined model-filtered subsets of the second subset to produce a trained model for the neural network.
8. The method of claim 7 , further including loading the trained model in an instance of the neural network in an inference engine.
9. A computer system configured to clean training data for a neural network, comprising:
one or more processors and memory storing computer program instructions configured to execute a process comprising:
accessing a labeled training data set;
using a first subset of the labeled training data set to train a first model of the neural network;
filtering a second subset of the labeled training data set using the first model to provide a first model-filtered subset of the second subset;
using the first model-filtered subset of the second subset to train a first refined model of the neural network;
filtering the first subset using the first refined model to provide a first refined model-filtered subset of the first subset;
using the first refined model-filtered subset of the first subset to train a second refined model of the neural network; and
filtering the second subset of the labeled training data set using the second refined model to provide a second refined model-filtered subset of the second subset.
10. The system of claim 9 , the process including:
combining the first refined model-filtered subset of the first subset and the second refined model-filtered subset of the second subset to provide a filtered training set, training an output model of a target neural network using the filtered training set, and saving the output model in memory.
11. The system of claim 9 , wherein the second refined model-filtered subset has a greater number of elements than the first model-filtered subset.
12. The system of claim 9 , wherein said filtering the first subset using the first refined model includes:
executing the neural network using the first refined model over the first subset to produce classification data classifying data elements of the first subset; and
selecting data elements of the first subset having labels matching the classification data to provide the first refined model-filtered subset of the first subset.
13. The system of claim 9 , the process including iteratively:
(i) using a previously provided refined model-filtered subset of one of the first subset and the second subset to train an instant refined model of the neural network;
(ii) filtering another of the first subset and the second subset using the instant refined model to provide an instant refined model-filtered subset of the other of the first subset and the second subset; and
(iii) determining whether an iteration criterion is met, and if not, then executing (i) to (iii), and if so, then using a combination of a selected one of the refined model-filtered subsets of the first subset and a selected one of the model-filtered subsets of the second subset to produce a trained model for a target neural network.
14. The system of claim 13 , the process including loading the trained model in an instance of the target neural network in an inference engine.
15. A computer program product configured to support cleaning training data for a neural network, comprising a non-transitory computer readable memory storing computer program instructions configured to execute a process comprising:
accessing a labeled training data set;
using a first subset of the labeled training data set to train a first model of the neural network;
filtering a second subset of the labeled training data set using the first model to provide a first model-filtered subset of the second subset;
using the first model-filtered subset of the second subset to train a first refined model of the neural network;
filtering the first subset using the first refined model to provide a first refined model-filtered subset of the first subset;
using the first refined model-filtered subset of the first subset to train a second refined model of the neural network; and
filtering second subset of the labeled training data set using the second refined model to provide a second refined model-filtered subset of the second subset.
16. The computer program product of claim 15 , wherein the second refined model-filtered subset has greater number of elements than the first model-filtered subset.
17. The computer program product of claim 15 , wherein the first subset and the second subset do not overlap.
18. The computer program product of claim 15 , the process including iteratively:
(i) using a previously provided refined model-filtered subset of one of the first subset and the second subset to train an instant refined model of the neural network;
(ii) filtering another of the first subset and the second subset using the instant refined model to provide an instant refined model-filtered subset of the other of the first subset and the second subset; and
(iii) determining whether an iteration criterion is met, and if not, then executing (i) to (iii), and if so, then using a combination of a selected one of the refined model-filtered subsets of the first subset and a selected one of the model-filtered subsets of the second subset to produce a trained model for the neural network.
19. The computer program product of claim 18 , the process including loading the trained model in an instance of the neural network in an inference engine.
20. The computer program product of claim 15 , wherein said filtering the first subset using the first refined model includes:
executing the neural network using the first refined model over the first subset to produce classification data classifying data elements of the first subset; and
selecting data elements of the first subset having labels matching the classification data to provide the first refined model-filtered subset of the first subset.