Implicit curriculum learning
Systems and techniques for facilitating implicit curriculum learning are provided. These allow for improved machine learning systems that can automatically execute curriculum learning without drawbacks such as pre-sorting data into different epochs which may have varying degrees of difficulty (e.g. easiest first, then harder epochs). Applicant's techniques can be executed more efficiently by automatically iterating over a data set, which need not be manually separated into different epochs. Thus, a system can access a neural network and a set of labeled data candidates. In various aspects, the system can perform a plurality of training epochs on the neural network based on the set of labeled data candidates. In various instances, the system can iteratively update the set of labeled data candidates as the plurality of training epochs are performed, by removing, after each training epoch, a dropout percentage of those labeled data candidates which the neural network correctly classified during the training epoch.
1 . A system, comprising:
a processor; and
a non-transitory computer-readable medium having stored thereon computer-executable instructions that are executable by the system to cause the system to perform operations comprising:
accessing, by the processor, a neural network and a set of labeled data candidates based on a first electronic activity of a plurality of users, at least some of the set of labeled data candidates representative of anomalous activity;
performing, by the processor, a plurality of training epochs on the neural network based on the set of labeled data candidates and an implicit metric to provide a trained neural network, wherein the implicit metric is based on a classification accuracy of the set of labeled data candidates at each training epoch, wherein performing the plurality of training epochs comprises:
iteratively updating the set of labeled data candidates as each training epoch of the plurality of training epochs is performed, by:
initializing a correctly-classified subset and an incorrectly-classified subset as empty sets;
identifying, after each given training epoch in the plurality of training epochs, the correctly-classified subset and the incorrectly-classified subset within the set of labeled data candidates based on the classification accuracy of a given data candidate, wherein, for the given data candidate, the identification comprises:
in response to the classification by the neural network matching a ground-truth classification, allocate the given data candidate to the correctly-classified subset, and
in response to the classification by the neural network not matching the ground-truth classification, allocate the given data candidate to the incorrectly-classified subset;
removing, from the set of labeled data candidates after each given training epoch, a dropout percentage of the correctly-classified subset to generate an updated set of labeled data candidates, wherein the updated set of labeled data candidates includes an entirety of the incorrectly-classified subset, wherein the dropout percentage is configured to modify the updated set of labeled data candidates such that more data candidates of the incorrectly-classified subset are represented in a next training epoch relative to data candidates in the correctly-classified subset, and wherein the updated set of labeled data candidates is treated as the set of labeled data candidates during the next training epoch that immediately follows the given training epoch; and
wherein, at each given training epoch, the neural network training is performed using a backpropagation algorithm based on a set of errors/losses determined for each given labeled data candidate of the set of labeled data candidates according to the classification by the neural network and the ground-truth classification to generate a given trained neural network, and wherein the given trained neural network is utilized as the neural network during the next training epoch; and
determining, by the trained neural network, anomalous activity in the system based on data representative of a second electronic activity of the plurality of users,
wherein the trained neural network is output based on the plurality of training epochs and a training termination criterion, the training termination criterion being based on a size of the incorrectly-classified subset at each given training epoch.
2 . The system of claim 1 , wherein the iterative updating of the set of labeled data candidates causes, without manual intervention, the set of labeled data candidates to become progressively smaller after each training epoch, and
wherein each successive set of labeled data candidates comprises a higher ratio of incorrectly-classified candidates compared to correctly-classified candidates as the plurality of training epochs are performed.
3 . The system of claim 1 , wherein the operations further comprise:
computing the dropout percentage based on a size of the correctly-classified subset or the size of the incorrectly-classified subset.
4 . The system of claim 1 , wherein the operations further comprise:
computing the dropout percentage based on a total number of training epochs in the plurality of training epochs.
5 . The system of claim 1 , wherein the operations further comprise:
ceasing to perform the plurality of training epochs and ceasing to iteratively update the set of labeled data candidates, once the training termination criterion has been reached.
6 . The system of claim 5 , wherein the training termination criterion is based on a size of the correctly-classified subset or the size of the incorrectly-classified subset.
7 . The system of claim 1 , wherein each data candidate in the set of labeled data candidates is an electronic transaction that is labeled as either fraudulent or non-fraudulent.
8 . A computer-implemented method, comprising:
performing, by a computer system comprising a processor, a plurality of training epochs on a neural network to output a trained neural network based on a set of labeled data candidates representative of electronic activity of a plurality of users and an implicit metric, the implicit metric being based on a classification accuracy of the electronic activity as anomalous activity at each epoch based on the set of labeled data candidates;
iteratively updating, by the computer system, the set of labeled data candidates as the plurality of training epochs are performed, wherein the iteratively updating includes:
initializing a correctly-classified subset and an incorrectly-classified subset as empty sets;
identifying, by the computer system and after each given training epoch in the plurality of training epochs, the correctly-classified subset and the incorrectly-classified subset within the set of labeled data candidates, wherein the correctly-classified subset includes labeled data candidates allocated thereto that the neural network correctly classified as corresponding to anomalous activity during the given training epoch;
removing, by the computer system and from the set of labeled data candidates, a dropout percentage of the correctly-classified subset to generate an updated set of labeled data candidates, wherein the updated set of labeled data candidates is treated as the set of labeled data candidates during a next training epoch that immediately follows the given training epoch;
training the neural network using a backpropagation algorithm based on a set of errors/losses determined for each given labeled data candidate of the set of labeled data candidates according to the classification by the neural network and a ground-truth classification to generate a trained neural network, and wherein the trained neural network is utilized as the neural network during the next training epoch; and
determining, by the trained neural network, anomalous activity based on the electronic activity of the plurality of users,
wherein the trained neural network is output based on the plurality of training epochs and a training termination criterion, the training termination criterion being based on a size of the incorrectly-classified subset at each given training epoch or relative to a minimum error level.
9 . The computer-implemented method of claim 8 ,
wherein the incorrectly-classified subset includes labeled data candidates that the neural network incorrectly classified as anomalous activity during the given training epoch, and wherein the updated set of labeled data candidates includes an entirety of the incorrectly-classified subset.
10 . The computer-implemented method of claim 9 , wherein the dropout percentage is based on a ratio of the size of the incorrectly-classified subset to a size of the correctly-classified subset.
11 . The computer-implemented method of claim 10 , wherein the dropout percentage varies inversely with the ratio of the size of the incorrectly-classified subset to the size of the correctly-classified subset.
12 . The computer-implemented method of claim 9 , further comprising:
ceasing, by the computer system, to perform the plurality of training epochs and ceasing, by the computer system, to iteratively update the set of labeled data candidates, based on determining that the size of the incorrectly-classified subset is below a minimum threshold.
13 . The computer-implemented method of claim 9 , further comprising:
ceasing, by the computer system, to perform the plurality of training epochs and ceasing, by the computer system, to iteratively update the set of labeled data candidates, based on determining that a ratio of the size of the incorrectly-classified subset to a size of the correctly classified subset is below a minimum threshold.
14 . The computer-implemented method of claim 8 , wherein the dropout percentage varies inversely with a total number of training epochs in the plurality of training epochs.
15 . A computer program product for facilitating implicit curriculum learning, the computer program product comprising a computer-readable medium having program instructions embodied therewith, the program instructions executable by a computer system to cause the computer system to perform operations comprising:
performing a plurality of training epochs on a neural network to generate a trained neural network based on a set of labeled data candidates representative of electronic activity of a plurality of users and an implicit metric, the implicit metric being based on a classification accuracy of the set of labeled data candidates at each training epoch; and
iteratively updating the set of labeled data candidates as the plurality of training epochs are performed, by:
classifying each labeled data candidate by the neural network according to a data type;
comparing the classification assigned by the neural network to a ground-truth classification associated with each labeled data candidate;
identifying, after each given training epoch in the plurality of training epochs, a correctly-classified subset and an incorrectly-classified subset within the set of labeled data candidates, wherein the correctly-classified subset includes labeled data candidates that the neural network correctly classified as anomalous activity during the given training epoch, and wherein the incorrectly-classified subset includes labeled data candidates that the neural network incorrectly classified during the given training epoch;
computing a dropout percentage based on a size of the correctly-classified subset and a size of the incorrectly-classified subset;
removing, from the set of labeled data candidates, the dropout percentage of the correctly-classified subset to generate an updated set of labeled data candidates, wherein the updated set of labeled data candidates includes an entirety of the incorrectly-classified subset, wherein the updated set of labeled data candidates is treated as the set of labeled data candidates during a next training epoch that immediately follows the given training epoch;
generating, for the updated set of labeled data candidates, a set of errors/losses, wherein an error/loss is determined for the given labeled data candidate according to the classification by the neural network and the ground-truth classification; and
training the neural network based on the set of errors/losses and a backpropagation algorithm to generate a trained neural network model, and wherein the trained neural network is utilized as the neural network during the next training epoch; and
determining, by the trained neural network, anomalous activity based on the electronic activity of the plurality of users,
wherein the trained neural network is output based on the plurality of training epochs and a training termination criterion, the training termination criterion being based on a size of the incorrectly-classified subset at each given training epoch.
16 . The computer program product of claim 15 , wherein the iterative updating causes the set of labeled data candidates to become progressively smaller, and
wherein each successive set of labeled data candidates comprises a higher ratio of incorrectly-classified candidates compared to correctly-classified candidates as the plurality of training epochs are performed, thereby causing the neural network to learn how to label based on the incorrectly-classified candidates without forgetting how to label the correctly-classified candidates.
17 . The computer program product of claim 15 , wherein the neural network is a fraud detection model that is configured to label an inputted online transaction as either fraudulent or non-fraudulent.
18 . The computer program product of claim 17 , wherein a data candidate from the set of labeled data candidates represents an online transaction that corresponds to a ground-truth label of fraudulent or non-fraudulent.
19 . The computer program product of claim 15 , wherein the given training epoch involves backpropagation to iteratively update parameters of the neural network.
20 . The computer program product of claim 15 , wherein the operations further comprise:
ceasing the performing the plurality of training epochs and ceasing the iteratively updating, based on the size of the incorrectly-classified subset being less than a minimum threshold.