Machine unlearning and retraining of a machine learning model based on a modified training dataset
Systems and methods for training a neural network are described. One or more embodiments of the present disclosure include training a neural network based on a first combined gradient of a loss function at a plurality of sampled elements of a dataset; receiving an insertion request that indicates an insertion element to be added to the dataset, or a deletion request that indicates a deletion element to be removed from the dataset, wherein the deletion element is one of the plurality of sampled elements; computing a second combined gradient of the loss function by adding the insertion element to the dataset or by replacing the deletion element with a replacement element from the dataset; determining whether the first combined gradient and the second combined gradient satisfy a stochastic condition; and retraining the neural network to obtain a modified neural network based on the determination.
1 . A method of training a neural network, comprising:
training a neural network based on a first combined gradient of a loss function at a plurality of sampled elements of a dataset;
receiving an insertion request that indicates an insertion element to be added to the dataset, or a deletion request that indicates a deletion element to be removed from the dataset, wherein the deletion element is one of the plurality of sampled elements;
computing a second combined gradient of the loss function by adding the insertion element to the dataset or by replacing the deletion element with a replacement element from the dataset;
determining whether the first combined gradient and the second combined gradient satisfy a stochastic condition; and
retraining the neural network to obtain a modified neural network based on the determination.
2 . The method of claim 1 , further comprising:
computing the first combined gradient by taking an average of partial gradients for the plurality of sampled elements.
3 . The method of claim 1 , further comprising:
adding a noise parameter to the first combined gradient, wherein the neural network is trained based on the first combined gradient with the added noise parameter.
4 . The method of claim 1 , further comprising:
performing an optimal transport process on the first combined gradient to obtain the second combined gradient.
5 . The method of claim 1 , further comprising:
computing a first partial gradient based on the deletion element, wherein the first combined gradient is based on the first partial gradient;
sampling an additional element from the dataset;
computing a second partial gradient based on the additional element; and
computing the second combined gradient based on the first partial gradient, the second partial gradient, and the first combined gradient.
6 . The method of claim 1 , further comprising:
performing rejection sampling on the sampled elements, wherein the neural network is retrained based on the rejection sampling.
7 . The method of claim 1 , further comprising:
computing a first probability density based on the first combined gradient and a noise parameter;
computing a second probability density based on the second combined gradient and the noise parameter; and
computing a ratio based on the first probability density and the second probability density, wherein the stochastic condition is based on the ratio.
8 . The method of claim 7 , further comprising:
selecting the replacement element from a current batch of the dataset;
determining that the ratio satisfies the stochastic condition based on the replacement element;
accepting the replacement element as a sample of the current batch based on satisfying the stochastic condition; and
evaluating a subsequent batch of the dataset, wherein the subsequent batch is evaluated without retraining the neural network based on the current batch.
9 . The method of claim 7 , further comprising:
determining that the ratio fails to satisfy the stochastic condition;
computing a reflected gradient based on the determination; and
retraining the neural network based on the reflected gradient.
10 . The method of claim 1 , wherein training the neural network further comprises:
computing a first iteration of a model parameter during a first previous training phase;
computing a second iteration of the model parameter during a second previous training phase;
computing a weighted sum of the first iteration of the model parameter and the second iteration of the model parameter; and
computing a third iteration of the model parameter based on the weighted sum and the first combined gradient.
11 . The method of claim 1 , further comprising:
receiving the insertion request; and
retraining the neural network based on the insertion request.
12 . The method of claim 1 , further comprising:
dividing the dataset to a plurality of batches, wherein the training is performed in a plurality of training phases corresponding to the plurality of batches.
13 . The method of claim 1 , further comprising:
receiving a plurality of additional deletion requests, wherein each of the additional deletion requests indicates an additional deletion element of the dataset to be removed; and
retraining the modified neural network based on each of the additional deletion requests.
14 . The method of claim 1 , further comprising:
a total variation distance between an output distribution of the neural network before retraining and a modified output distribution of the modified neural network is less than or equal to a predetermined threshold value.
15 . A method comprising:
operating a neural network that is trained on a dataset including a plurality of minibatches;
receiving a deletion request that indicates a deletion element to be removed from the dataset;
retraining the neural network based on a modified training dataset that does not include the deletion element by performing an optimal transport algorithm for at least one of the minibatches; and
operating the neural network subsequent to a down-time corresponding to the retraining, wherein the down-time is shorter than a second down-time corresponding to a second retraining of the neural network using each of the minibatches.
16 . The method of claim 15 , further comprising:
determining whether a combined loss function gradient corresponding to an initial minibatch of the modified training dataset satisfies a stochastic condition; and
determining whether to retrain the neural network using the initial minibatch based on the determination.
17 . An apparatus for training a neural network, comprising:
a processor and a memory storing instructions and in electronic communication with the processor, the processor being configured to execute the instructions to:
train a neural network based on a first combined gradient of a loss function at a plurality of sampled elements of a dataset;
receive an insertion request that indicates an insertion element to be added to the dataset, or a deletion request that indicates a deletion element to be removed from the dataset, wherein the deletion element is one of the plurality of sampled elements;
compute a second combined gradient of the loss function by adding the insertion element to the dataset or by replacing the deletion element with a replacement element from the dataset;
determine whether the first combined gradient and the second combined gradient satisfy a stochastic condition; and
retrain the neural network to obtain a modified neural network based on the determination.
18 . The apparatus of claim 17 , further comprising:
a rejection sampling component configured to perform rejection sampling on the sampled elements, wherein the neural network is retrained based on the rejection sampling.
19 . The apparatus of claim 17 , the instructions further configured to:
perform an optimal transport process on the first combined gradient to obtain the second combined gradient.