Table 1 Summary and induction of typical DNNs

SAE

Stacked AE layers + a classifier

1) Unsupervised learning

2) Layer-by-layer unsupervised learning, followed by reverse supervised fine-tuning

3) Compression and distributed representation of dataset

1-dimensional (1D): Data de-noising [67]

2-dimensional (2D): Target recognition [68]

3-dimensional (3D): High spectral image classification [69]

DBN

Stacked RBM layers + a classifier

1) Unsupervised learning

2) Layer-by-layer unsupervised learning, followed by reverse supervised fine-tuning

3) Learning probability distribution on generative structure

1D: Speech recognition [70]

2D: Radar image recognition [71]

3D: High spectral image classification [72]

CNN

Alternately appeared convolution and sampling layers + a classifier

1) Supervised learning

2) Automatic learning of convolution kernels

3) Obtains the essence of the samples through a local convolution operation

1D: Speech recognition [73]

2D: Image classification [74]

Face recognition [75]

3D: Video classification [76]

RNN

Input layer + closed-loop hidden layers + output layer

1) Unsupervised learning

2) Feedback loop taking the output of the previous moment as input

3) Mainly used for modeling time-series signals

1D: Text classification [77]

Machine translation [78]

2D: Image annotation [79]

Emotional test [80]

3D: Video analysis [81]

GAN

Generator + discriminator

1) Semi-supervised learning

2) Learn the probability distribution of the training samples and establish correlation between the input and output

1D: Image generation [82]

2D: Image retrieval [83]

3D: Video prediction [84]