Learning device, signal processing device, and learning method
A learning data processing unit accepts, as input, a plurality of pieces of learning data for a respective plurality of tasks, and calculates, for each of the tasks, a batch size which meets a condition that a value obtained by dividing a data size of corresponding one of the pieces of learning data by the corresponding batch size is the same between the tasks. A batch sampling unit samples, for each of the tasks, samples from corresponding one of the pieces of learning data with the corresponding batch size calculated by the learning data processing unit. A learning unit updates a weight of a discriminator for each of the tasks, using the samples sampled by the batch sampling unit.
1. A learning device for training a single neural network for a plurality of tasks of different types using learning data whose data size varies from task to task, comprising:
a processor to execute a program; and
a memory to store the program which, when executed by the processor, performs multi-task training of the neural network using stochastic gradient descent by performing processes of,
accepting, as input, corresponding learning data, having a corresponding data size, for each of the plurality of tasks, the corresponding data size for each task varying from task to task;
calculating, for each of the plurality of tasks, a corresponding batch size which meets a condition that a value obtained by dividing the data size of the corresponding learning data for a task by the corresponding batch size for the task is the same between the plurality of tasks;
sampling, for each of the plurality of tasks, samples from the corresponding learning data with the calculated corresponding batch size;
training the single neural network by updating a corresponding weight of a discriminator of the neural network for each of the plurality of tasks, using the samples sampled;
and
repeating the sampling and training until all samples in the learning data have been sampled.
2. A learning device for training a single neural network for a plurality of tasks of different types using learning data whose data size varies from task to task, comprising:
a processor to execute a program; and
a memory to store the program which, when executed by the processor, performs multi-task training of the neural network using stochastic gradient descent by performing processes of,
accepting, as input, corresponding learning data for each of the plurality of tasks, the corresponding learning data for each task having a corresponding data size, the corresponding data size for each task varying from task to task;
calculating, for the plurality of tasks, a respective batch size whose ratio to the corresponding data size for the task has a fixed value between the plurality of tasks;
sampling, for each of the plurality of tasks, samples from the corresponding learning data with calculated corresponding batch size;
training the single neural network by updating a corresponding weight of a discriminator of the neural network for each of the plurality of tasks, using the samples sampled;
and
repeating the sampling and training until all samples in the learning data have been sampled.
3. A signal processing device comprising:
an input information processor to accept input of input information; and
a discriminator to perform a discrimination process using the input information accepted by the input information processor, the discriminator being caused to learn by the learning device according to claim 1 .
4. A signal processing device comprising:
an input information processor to accept input of input information; and
a discriminator to perform a discrimination process using the input information accepted by the input information processor, the discriminator being caused to learn by the learning device according to claim 2 .
5. A learning method of performing multi-task training of a single neural network using stochastic gradient descent for a plurality of tasks of different types using learning data whose data size varies from task to task, the method comprising:
accepting, as input, corresponding learning data, having a corresponding data size for each of the plurality of tasks, the corresponding data size for each task varying from task to task;
calculating, for each of the plurality of tasks, a corresponding batch size which meets a condition that a value obtained by dividing the data size of the corresponding learning data for a task by the corresponding batch size for the task is the same between the plurality of tasks;
sampling, for each of the plurality of tasks, samples from the corresponding learning data with the calculated corresponding batch size;
training the single neural network by updating, a corresponding weight of a discriminator of the neural network for each of the plurality of tasks, using the samples sampled;
and
repeating the sampling and training until all samples in the learning data have been sampled.
6. A learning method of performing multi-task training of a single neural network using stochastic gradient descent for a plurality of tasks of different types using learning data whose data size varies from task to task, the method comprising:
accepting, as input, corresponding learning data, having a corresponding data size, for each of the plurality of tasks, the corresponding data size for each task varying from task to task;
calculating, for the plurality of tasks, a respective batch size whose ratio to the corresponding data size for the task has a fixed value between the plurality of tasks;
sampling, for each of the plurality of tasks, samples from the corresponding learning data with the calculated corresponding batch size;
training the single neural network by updating, a corresponding weight of a discriminator of the neural network for each of the plurality of tasks, using the samples sampled; and
repeating the sampling and training until all samples in the learning data have been sampled.