Network training device, network training system, network training method, and computer program product

1. A network training device adapted to train a neural network, the neural network including a first network that converts an input signal to a first signal, a second network that converts the first signal to a second signal, and a third network that converts the second signal to an output signal, the device comprising:

a first training unit configured to train the first network;

a second training unit configured to train the second network; and

a third training unit configured to train the third network, wherein

the first training unit trains the first network as an encoder of a first autoencoder that encodes an input signal for training into a first signal for training having lower dimensionality than the input signal for training and decodes the first signal for training into the input signal for training,

the second training unit trains the second network by backpropagation by using a second signal for training corresponding to the first signal for training as supervised data, and

the second signal for training is generated by an encoder of a second autoencoder that encodes a third signal for training into the second signal for training having lower dimensionality than the third signal for training and decodes the second signal for training into the third signal for training.

2. The device according to claim 1 , wherein

the third signal for training is an output signal for training, and

the third training unit trains the third network as a decoder of the second autoencoder.

3. The device according to claim 1 , further comprising a fourth training unit configured to fine-tune the entire neural network in which the first network trained by the first training unit, the second network trained by the second training unit, and the third network trained by the third training unit are connected.

4. The device according to claim 1 , wherein at least one of the first network and the third network includes a convolutional neural network.

5. The device according to claim 1 , further comprising a parameter change unit configured to change, regarding at least one of the first network, the second network, and the third network, at least one of a depth of a layer of a network, number of nodes in the layer, number of channels or a filter size in a case in which the network includes the convolutional neural network, in accordance with a user operation.

6. The device according to claim 1 , further comprising a display control unit configured to cause at least one of reproducibility of the first autoencoder and reproducibility of the second autoencoder to be displayed.

7. A network training system adapted to train a neural network by using a plurality of information processing devices, the neural network including a first network that converts an input signal to a first signal, a second network that converts the first signal to a second signal, and a third network that converts the second signal to an output signal, the system comprising:

a first training unit configured to train the first network;

a second training unit configured to train the second network; and

a third training unit configured to train the third network, wherein

the first training unit trains the first network as an encoder of a first autoencoder that encodes an input signal for training into a first signal for training having lower dimensionality than the input signal for training and decodes the first signal for training into the input signal for training,

the second training unit trains the second network by backpropagation by using a second signal for training corresponding to the first signal for training as supervised data, and

the second signal for training is generated by an encoder of a second autoencoder that encodes a third signal for training into the second signal for training having lower dimensionality than the third signal for training and decodes the second signal for training into the third signal for training.

8. A network training method adapted to train a neural network, the neural network including a first network that converts an input signal to a first signal, a second network that converts the first signal to a second signal, and a third network that converts the second signal to an output signal, the method comprising:

training the first network;

training the second network; and

training the third network, wherein

at the training the first network, the first network is trained as an encoder of a first autoencoder that encodes an input signal for training into a first signal for training having lower dimensionality than the input signal for training and decodes the first signal for training into the input signal for training,

at the training the second network, the second network is trained by backpropagation by using a second signal for training corresponding to the first signal for training as supervised data, and

the second signal for training is generated by an encoder of a second autoencoder that encodes a third signal for training into the second signal for training having lower dimensionality than the third signal for training and decodes the second signal for training into the third signal for training.

9. A computer program product comprising a computer readable medium including programmed instructions, the instructions causing a computer to perform a function of training a neural network, the neural network including a first network that converts an input signal to a first signal, a second network that converts the first signal to a second signal, and a third network that converts the second signal to an output signal, the instructions causing the computer to perform:

training the first network;

training the second network; and

training the third network, wherein

at the training the first network, the first network is trained as an encoder of a first autoencoder that encodes an input signal for training into a first signal for training having lower dimensionality than the input signal for training and decodes the first signal for training into the input signal for training,

at the training the second network, the second network is trained by backpropagation by using a second signal for training corresponding to the first signal for training as supervised data, and

the second signal for training is generated by an encoder of a second autoencoder that encodes a third signal for training into the second signal for training having lower dimensionality than the third signal for training and decodes the second signal for training into the third signal for training.