when the level of the network deepens, it still takes days or even weeks to learn. And in
order to build a better network, deep learning is always accompanied by various trials
and errors, which creates the need to shorten the learning time as much as possible. For
this reason, distributed computing can be considered on multiple GPUs, but this is a
problem that is difficult to overcome because it involves multiple problems such as
communication between machines and synchronization of data. Considering that
memory capacity and bus bandwidth may also become bottlenecks, the number of data
bits flowing through the network should be reduced as much as possible. In order to
achieve high speed, bit number reduction is also a subject of concern. In the training of
the network, a large data set is usually needed, and the knowledge of the existing public
data set can be transferred to the own network, and a part of the trained model is
directly applied to the similar model, such as a small amount. The data only trains part
of the network and can achieve good results on small data sets. This migration learning
technology is one of the future research directions. Through deep learning, you can also
draw artistic paintings, such as drawing a painting with a Van Gogh style, inputting two
images, one content image, another style image, and a new image will be generated. In
the process of learning, the intermediate data of the network is similar to the intermediate data of the content image, and an input image such as a similar style of the
content image can be output [10]. Different from the above research, there is also a
network model that can generate new images without any input image. For example,
based on deep learning, it is possible to generate images of the living room from zero.
This technology is mainly developed by letting the two neural networks of generator
and discriminator learn from each other. The generator that will eventually grow up
will have the ability to draw images that are enough to confuse the real [11]. Deep
learning has shown excellent performance in many aspects and has become a research
hotspot. However, there are still many works to be done in the theoretical research of
the model. More research work on deep learning requires researchers to carry out.
References
1. Shi C, Pun CM (2019) Adaptive multi-scale deep neural networks with perceptual loss for
panchromatic and multispectral images classification. Inf Sci
2. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep
convolutional neural networks. In: Advances in neural information processing systems,
pp 1097–1105
3. Zeiler MD, Fergus R (2014) Visualizing and understanding convolutional networks. In:
European conference on computer vision. Springer, Cham, pp 818–833
4. Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409–1556
5. Szegedy C, Liu W, Jia Y et al (2015) Going deeper with convolutions. In: Proceedings of the
IEEE conference on computer vision and pattern recognition, pp 1–9
6. He K, Zhang X, Ren S et al (2016) Deep residual learning for image recognition. In:
Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770–778
7. Jaderberg M, Simonyan K, Zisserman A (2015) Spatial transformer networks. In: Advances
in neural information processing systems, pp 2017–2025
Research on System Architecture Based on Deep Learning …
521
order to build a better network, deep learning is always accompanied by various trials
and errors, which creates the need to shorten the learning time as much as possible. For
this reason, distributed computing can be considered on multiple GPUs, but this is a
problem that is difficult to overcome because it involves multiple problems such as
communication between machines and synchronization of data. Considering that
memory capacity and bus bandwidth may also become bottlenecks, the number of data
bits flowing through the network should be reduced as much as possible. In order to
achieve high speed, bit number reduction is also a subject of concern. In the training of
the network, a large data set is usually needed, and the knowledge of the existing public
data set can be transferred to the own network, and a part of the trained model is
directly applied to the similar model, such as a small amount. The data only trains part
of the network and can achieve good results on small data sets. This migration learning
technology is one of the future research directions. Through deep learning, you can also
draw artistic paintings, such as drawing a painting with a Van Gogh style, inputting two
images, one content image, another style image, and a new image will be generated. In
the process of learning, the intermediate data of the network is similar to the intermediate data of the content image, and an input image such as a similar style of the
content image can be output [10]. Different from the above research, there is also a
network model that can generate new images without any input image. For example,
based on deep learning, it is possible to generate images of the living room from zero.
This technology is mainly developed by letting the two neural networks of generator
and discriminator learn from each other. The generator that will eventually grow up
will have the ability to draw images that are enough to confuse the real [11]. Deep
learning has shown excellent performance in many aspects and has become a research
hotspot. However, there are still many works to be done in the theoretical research of
the model. More research work on deep learning requires researchers to carry out.
References
1. Shi C, Pun CM (2019) Adaptive multi-scale deep neural networks with perceptual loss for
panchromatic and multispectral images classification. Inf Sci
2. Krizhevsky A, Sutskever I, Hinton GE (2012) Imagenet classification with deep
convolutional neural networks. In: Advances in neural information processing systems,
pp 1097–1105
3. Zeiler MD, Fergus R (2014) Visualizing and understanding convolutional networks. In:
European conference on computer vision. Springer, Cham, pp 818–833
4. Simonyan K, Zisserman A (2014) Very deep convolutional networks for large-scale image
recognition. arXiv preprint arXiv:1409–1556
5. Szegedy C, Liu W, Jia Y et al (2015) Going deeper with convolutions. In: Proceedings of the
IEEE conference on computer vision and pattern recognition, pp 1–9
6. He K, Zhang X, Ren S et al (2016) Deep residual learning for image recognition. In:
Proceedings of the IEEE conference on computer vision and pattern recognition, pp 770–778
7. Jaderberg M, Simonyan K, Zisserman A (2015) Spatial transformer networks. In: Advances
in neural information processing systems, pp 2017–2025
Research on System Architecture Based on Deep Learning …
521