Method and information processing apparatus that perform transfer learning while suppressing occurrence of catastrophic forgetting

1. A method executed by an information processing apparatus in order to cause a neural network to learn a T th task corresponding to a T th learning set, the neural network having learned, in sequence, a first task to a (T−1) th task that respectively correspond to a first learning set to a (T−1) th learning set by using the first learning set to the (T−1) th learning set, where T is an integer equal to or larger than 2,

the neural network including a plurality of layers that each include a plurality of units,

each of the plurality of units including a weight parameter corresponding to an input value to the unit,

the method comprising:

for each of the plurality of units included in the plurality of layers, determining an importance degree of the unit in the T th task;

for each of the plurality of layers, determining dissimilar tasks from among the first task to the (T−1) th task, the dissimilar tasks being not similar to the T th task in terms of behaviors in the layer; and

in learning that uses the T th learning set, suppressing updating of weight parameters of the plurality of units included in the plurality of layers in accordance with importance degrees in the dissimilar tasks that have been determined for each of the plurality of layers.

2. The method according to claim 1 , wherein

the weight parameter of a first unit included in a first layer included among the plurality of layers, includes a first weight coefficient corresponding to the input value that is input to the first unit from a second unit included in a second layer included among the plurality of layers,

the suppressing includes updating the first weight coefficient based on a first value that is a smaller one of: a largest value of importance degrees of the dissimilar tasks in the first unit that have been determined with respect to the first layer; and a largest value of importance degrees of the dissimilar tasks in the second unit that have been determined with respect to the second layer, and

the larger the first value, the higher a degree of suppression of updating of the first weight coefficient.

3. The method according to claim 2 , wherein

the suppressing includes

obtaining an error amount of the first unit,

obtaining a first gradient of the first weight coefficient based on the error amount of the first unit,

changing the first gradient to a second gradient based on the first value, and

updating the first weight coefficient based on the second gradient, and

an absolute value of the second gradient is smaller than an absolute value of the first gradient.

4. The method according to claim 1 , wherein

each of the plurality of units in the plurality of layers includes a feature parameter, and

the determining of the dissimilar tasks includes repetitive processing that repeats following with respect to each of the plurality of units included in the plurality of layers:

in order to obtain an output of the neural network based on the T th learning set, obtaining an output value of the unit based on the input value to the unit and on the weight parameter and the feature parameter of the unit;

obtaining an error amount of the unit based on the output of the neural network based on the T th learning set;

obtaining a third gradient of the feature parameter based on the error amount of the unit; and

updating the feature parameter based on the third gradient of the feature parameter.

5. The method according to claim 4 , wherein

the determining of the dissimilar tasks includes

after the repetitive processing has been completed using the T th learning set, obtaining, for each of the plurality of layers, feature vectors that use the feature parameters of the plurality of units included in the layer as elements, and

for each of the plurality of layers, clustering the feature vectors of the first task to the T th task, and

the dissimilar tasks related to each of the plurality of layers are determined based on a result of the clustering.

6. The method according to claim 5 , wherein

the dissimilar tasks related to the layer are tasks which are included among the first task to the (T−1) th task, and which correspond to the feature vectors that have been categorized into clusters different from a cluster of the feature vector of the T th task.

7. The method according to claim 4 , wherein

the obtaining of the output value of the unit includes

obtaining an intermediate value based on the input value to the unit and the weight parameter,

using a function, converting a value based on the feature parameter into an importance degree parameter in a range of 0 to 1, and

multiplying the intermediate value by the importance degree parameter in order to obtain the output value of the unit.

8. The method according to claim 7 , wherein

the importance degree of the unit is determined based on a value of the importance degree parameter.

9. The method according to claim 8 , wherein

the importance degree of the unit increases as the value of the importance degree parameter increases.

10. The method according to claim 7 , wherein

the function is a sigmoid function.

11. The method according to claim 10 , wherein

the importance degree parameter is a value obtained by converting a third value using the sigmoid function, the third value being obtained by multiplying the feature parameter by a second value.

12. The method according to claim 7 , wherein

the weight parameter of a first unit included in a first layer included among the plurality of layers, includes a first weight coefficient corresponding to the input value that is input to the first unit from a second unit included in a second layer included among the plurality of layers,

the repetitive processing includes

determining a first largest value of the importance degree parameter of the first unit and the importance degrees of the first unit in the first task to the (T−1) th task,

determining a fourth gradient of the first weight coefficient based on the error amount of the unit,

obtaining, from the second unit, a second largest value of the importance degree parameter of the second unit and the importance degrees of the second unit in the first task to the (T−1) th task,

changing the fourth gradient to a fifth gradient based on a fourth value that is the smaller one of the first largest value and the second largest value, and

updating the first weight coefficient based on the fifth gradient, and

an absolute value of the fifth gradient is smaller than an absolute value of the fourth gradient.

13. The method according to claim 12 , wherein

a difference between the fifth gradient and the fourth gradient increases as the fourth value increases.

14. An information processing apparatus, comprising:

one or more processors; and

a storage device that stores a program that is executable by the one or more processors,

wherein

the storage device further stores weight parameters that respectively correspond to a plurality of units included in each of a plurality of layers of a neural network that have learned, in sequence, a first task to a (T−1) th task that respectively correspond to a first learning set to a (T−1) th learning set by using the first learning set to the (T−1) th learning set, where T is an integer equal to or larger than 2,

by executing the program, the one or more processors cause the information processing apparatus to execute learning processing that causes the neural network to learn a T th task corresponding to a T th learning set using the T th learning set, and

the learning processing includes

for each of the plurality of units included in the plurality of layers, determining an importance degree of the unit in the T th task,

for each of the plurality of layers, determining dissimilar tasks from among the first task to the (T−1) th task, the dissimilar tasks being not similar to the T th task in terms of behaviors in the layer, and

in learning that uses the T th learning set, suppressing updating of the weight parameters of the plurality of units included in the plurality of layers in accordance with importance degrees of the dissimilar tasks that have been determined for each of the plurality of layers.

15. A non-transitory computer-readable storage medium that stores a program,

wherein

when executed by one or more processors of an information processing apparatus, the program causes the information processing apparatus to execute learning processing that causes a neural network to learn a task corresponding to a learning set, the neural network including a plurality of layers that each include a plurality of units including weight parameters, and

in causing the neural network to learn a T th task corresponding to a T th learning set after the neural network has learned, in sequence, a first task to a (T−1) th task that respectively correspond to a first learning set to a (T−1) th learning set, where T is an integer equal to or larger than 2, the learning processing includes

for each of the plurality of units in the plurality of layers, determining an importance degree of the unit in the T th task,

for each of the plurality of layers, determining dissimilar tasks from among the first task to the (T−1) th task, the dissimilar tasks being not similar to the T th task in terms of behaviors in the layer, and

in learning that uses the T th learning set, suppressing updating of the weight parameters of the plurality of units included in the plurality of layers in accordance with importance degrees of the dissimilar tasks that have been determined for each of the plurality of layers.