Energy-efficient retraining method of generative neural network for domain-specific optimization

1 . An energy-efficient retraining method of a generative neural network for domain-specific optimization, the energy-efficient retraining method comprising:

retraining, by a mobile device, a pretrained generative neural network model with respect to some data of a new user dataset;

comparing, by the mobile device, the pretrained generative neural network model and a generative neural network model retrained for each layer with each other in terms of a relative change rate of weights;

selecting, by the mobile device, specific layers having high relative change rate of weights, among layers of the pretrained generative neural network model, as layers to be retrained; and

performing, by the mobile device, weight update for only the layers selected in step (c),

wherein the mobile device divides a difference between a weight of the pretrained generative neural network model and a weight of the retrained generative neural network model by a size of the weight of the pretrained generative neural network model in order to calculate the relative change rate of weights,

wherein, in step (c), the mobile device selects k continuous or discontinuous layers from a random m-th layer, instead of retraining all n layers of the model,

wherein, in step (d), the mobile device selects the k discontinuous layers by selecting layers having relative change rates of weights before and after retraining exceeding a specific threshold value, maintains original weights without error propagation and weight update for continuous layers that are located before a first one of the discontinuously selected layers and that are not selected, performs forward propagation in only a first epoch of a retraining step, and reuses a result of the forward propagation of the first epoch in repeated retraining epochs thereafter,

wherein the mobile device performs forward propagation, error propagation, and weight update for only each of the discretely selected layers, and performs forward propagation and error propagation without weight update for layers that are located after a first one of the discretely selected layers.

2 . An energy-efficient retraining method of a generative neural network for domain-specific optimization, the energy-efficient retraining method comprising:

retraining, by a mobile device, a pretrained generative neural network model with respect to some data of a new user dataset;

comparing, by the mobile device, the pretrained generative neural network model and a generative neural network model retrained for each layer with each other in terms of a relative change rate of weights;

selecting, by the mobile device, specific layers having high relative change rate of weights, among layers of the pretrained generative neural network model, as layers to be retrained; and

performing, by the mobile device, weight update for only the layers selected in step (c),

wherein the mobile device divides a difference between a weight of the pretrained generative neural network model and a weight of the retrained generative neural network model by a size of the weight of the pretrained generative neural network model in order to calculate the relative change rate of weights,

wherein, in step (c), the mobile device selects k continuous or discontinuous layers from a random m-th layer, instead of retraining all n layers of the model,

wherein, in step (d), the mobile device maintains original weights without weight update for unselected layers, after selecting the k continuous layers, does not perform even back propagation for unselected layers before a first one of the selected layers, performs forward propagation in only a first epoch of retraining, and reuses a result of the forward propagation of the first epoch in repeated retraining epochs thereafter,

wherein the mobile device receives a result obtained from the generative neural network, converts quality of the result of the generative neural network into a score, and performs the error propagation using the score as a loss function,

wherein the mobile device sets the loss function in order to generate a result in which distribution of generative data of a generative neural network to be trained approximates distribution of data at a time of initial training, and performs training.

3 . The energy-efficient retraining method according to claim 1 , wherein the mobile device receives a result obtained from the generative neural network, converts quality of the result of the generative neural network into a score, and performs the error propagation using the score as a loss function.

4 . An energy-efficient retraining method of a generative neural network for domain-specific optimization, the energy-efficient retraining method comprising:

retraining, by a mobile device, a pretrained generative neural network model with respect to some data of a new user dataset;

comparing, by the mobile device, the pretrained generative neural network model and a generative neural network model retrained for each layer with each other in terms of a relative change rate of weights;

selecting, by the mobile device, specific layers having high relative change rate of weights, among layers of the pretrained generative neural network model, as layers to be retrained; and

performing, by the mobile device, weight update for only the layers selected in step (c),

wherein the mobile device divides a difference between a weight of the pretrained generative neural network model and a weight of the retrained generative neural network model by a size of the weight of the pretrained generative neural network model in order to calculate the relative change rate of weights,

wherein, in step (c), the mobile device selects k continuous or discontinuous layers from a random m-th layer, instead of retraining all n layers of the model,

wherein, in step (d), the mobile device selects the k discontinuous layers by selecting layers having relative change rates of weights before and after retraining exceeding a specific threshold value, maintains original weights without error propagation and weight update for continuous layers that are located before a first one of the discontinuously selected layers and that are not selected, performs forward propagation in only a first epoch of a retraining step, and reuses a result of the forward propagation of the first epoch in repeated retraining epochs thereafter,

wherein the mobile device performs forward propagation, error propagation, and weight update for only each of the discretely selected layers, and performs forward propagation and error propagation without weight update for layers that are located after a first one of the discretely selected layers,

wherein the mobile device receives a result obtained from the generative neural network, converts quality of the result of the generative neural network into a score, and performs the error propagation using the score as a loss function.