Method and information processing apparatus for performing transfer learning while suppressing occurrence of catastrophic forgetting

1 . A method to be executed by an information processing apparatus, the method comprising:

training a neural network having learned first to (T−1)-th tasks (where T is an integer of two or more) sequentially using first to (T−1)-th training sets to learn a T-th task using a T-th training set, wherein the neural network includes a plurality of layers, and the plurality of layers each include a plurality of units, the plurality of units each have a weight coefficient associated with each input to a unit, and an importance parameter indicating importance of the weight coefficient,

wherein the training the neural network comprises:

adjusting a first weight coefficient based on the importance parameter of the first weight coefficient and a first gradient of the first weight coefficient that is determined using the T-th training set;

after the learning with the T-th training set has been completed, determining, based on the T-th training set, respective second gradients of a plurality of second weight coefficients included in a first layer among the plurality of layers, wherein the plurality of second weight coefficients include the first weight coefficient, and calculating the importance parameter of the first weight coefficient to be used in learning of a (T+1)-th task based on the respective second gradients of the plurality of second weight coefficients, wherein the plurality of layers include first to T-th heads that are for providing outputs of the neural network respectively corresponding to the first to T-th tasks,

wherein the calculating the importance parameter of the first weight coefficient to be used in the learning of the (T+1)-th task comprises:

normalizing a second gradient of the first weight coefficient based on the respective second gradients of the plurality of second weight coefficients; and

determining respective fourth gradients of the plurality of second weight coefficients in a t-th task (where t is an integer from 1 to T) based on an output from a t-th head based on the T-th training set;

normalizing a fourth gradient of the first weight coefficient in the t-th task based on the respective fourth gradients of the plurality of second weight coefficients in the t-th task; and

determining a maximum value among the normalized fourth gradients of the first to T-th tasks, and

wherein the maximum value is the normalized second gradient of the first weight coefficient.

2 . The method according to claim 1 , wherein

the adjusting of the first weight coefficient comprises:

calculating a third gradient by adjusting the first gradient of the first weight coefficient based on the importance parameter of the first weight coefficient; and

adjusting the first weight coefficient based on the third gradient,

the third gradient has a smaller absolute value than the first gradient, and

the higher the importance of the first weight coefficient indicated by the importance parameter of the first weight coefficient is, the greater a difference between the third gradient and the first gradient is.

3 . The method according to claim 1 , wherein an absolute value of the second gradient of the first weight coefficient is normalized to a value that is not smaller than zero and is smaller than one.

4 . The method according to claim 1 , wherein the normalizing the second gradient of the first weight coefficient comprises calculating an average and variance of the respective second gradients of the plurality of second weight coefficients.

5 . The method according to claim 1 , wherein the importance parameter of the first weight coefficient to be used in the learning of the (T+1)-th task is a greater one of the normalized second gradient of the first weight coefficient and the importance parameter of the first weight coefficient used in the learning of the T-th task.

6 . An information processing apparatus comprising:

one or more processors; and

a storage device, operatively connected to the one or more processors, that stores a program that can be executed by the one or more processors, wherein

the storage device further stores, for each of a plurality of units included in each of a plurality of layers of a neural network that has sequentially learned first to (T−1)-th tasks (where T is an integer of two or more) sequentially using first to (T−1)-th training sets, a weight coefficient associated with each input to a unit and an importance parameter indicating importance of the weight coefficient,

by executing the program, the one or more processors cause the information processing apparatus to execute training processing for causing the neural network to learn a T-th task using a T-th training set, and

the training processing comprises:

adjusting a first weight coefficient based on the importance parameter of the first weight coefficient and a first gradient of the first weight coefficient that is determined using the T-th training set; and

after learning with the T-th training set has been completed, determining, based on the T-th training set, respective second gradients of a plurality of second weight coefficients included in a first layer among the plurality of layers, wherein the plurality of second weight coefficients include the first weight coefficient, and calculating the importance parameter of the first weight coefficient to be used in learning of a (T+1)-th task based on the respective second gradients of the plurality of second weight coefficients, wherein the plurality of layers include first to T-th heads that are for providing outputs of the neural network respectively corresponding to the first to T-th tasks, and

wherein the calculating the importance parameter of the first weight coefficient to be used in the learning of the (T+1)-th task comprises:

normalizing a second gradient of the first weight coefficient based on the respective second gradients of the plurality of second weight coefficients;

determining respective fourth gradients of the plurality of second weight coefficients in a t-th task (where t is an integer from 1 to T) based on an output from a t-th head based on the T-th training set;

normalizing a fourth gradient of the first weight coefficient in the t-th task based on the respective fourth gradients of the plurality of second weight coefficients in the t-th task; and

determining a maximum value among the normalized fourth gradients of the first to T-th tasks, and

wherein the maximum value is the normalized second gradient of the first weight coefficient.

7 . A non-transitory computer-readable recording medium storing a program, wherein

upon being executed by one or more processors of an information processing apparatus, the program causes the information processing apparatus to execute training processing for causing a neural network having sequentially learned first to (T−1)-th tasks (where T is an integer of two or more) using first to (T−1)-th training sets to learn a T-th task using a T-th training set,

for each of a plurality of units included in each of a plurality of layers of the neural network, the information processing apparatus has stored therein a weight coefficient associated with each input to a unit and an importance parameter indicating importance of the weight coefficient, and

the training processing comprises:

adjusting a first weight coefficient based on the importance parameter of the first weight coefficient and a first gradient of the first weight coefficient that is determined using the T-th training set; and

after learning with the T-th training set has been completed, determining, based on the T-th training set, respective second gradients of a plurality of second weight coefficients included in a first layer among the plurality of layers, wherein the plurality of second weight coefficients include the first weight coefficient, and calculating the importance parameter of the first weight coefficient to be used in learning of a (T+1)-th task based on the respective second gradients of the plurality of second weight coefficients, wherein the plurality of layers include first to T-th heads that are for providing outputs of the neural network respectively corresponding to the first to T-th tasks, and

wherein the calculating the importance parameter of the first weight coefficient to be used in the learning of the (T+1)-th task comprises:

normalizing a second gradient of the first weight coefficient based on the respective second gradients of the plurality of second weight coefficients;

determining respective fourth gradients of the plurality of second weight coefficients in a t-th task (where t is an integer from 1 to T) based on an output from a t-th head based on the T-th training set;

normalizing a fourth gradient of the first weight coefficient in the t-th task based on the respective fourth gradients of the plurality of second weight coefficients in the t-th task; and

determining a maximum value among the normalized fourth gradients of the first to T-th tasks, and

wherein the maximum value is the normalized second gradient of the first weight coefficient.

8 . A system configured to perform the method of claim 1 .