Flow: Model Family----depth network----depth learning
Model family: The model has a super-parameter, and gives different parameters corresponding to different models, it forms the model family
Algorithm family: Each model corresponds to a complete algorithm, the entire model family corresponds to an algorithm family to expand the algorithm family into a deep network, the network layer represents the number of iterations, the model's hyper-parameters become the parameters of the network (such as weight, etc.). The network can be trained with a small amount of tagged data.
Advantages relative to model-driven algorithms:
1, can learn the model super parameter, improve the adaptability of the model, improve the accuracy
Relative to data-driven benefits:
1, the network design has the model guidance
2, reduce the data demand
3, reduce the training time
For example, the ADMM algorithm for NMR reconstruction:
Model:
\ (x^*={arg\max}_{x}{\{\frac{1}{2}| | ax-y| | ^2+\sum_{l=1}^{l}\lambda_{l}g (d_{l}x) \}}\)
ADMM Algorithm Solution:
The different choices of \ (g,\lambda,l,d_{l}\) Form different models and constitute the model family.
Generalized Lagrangian functions:
ADMM algorithm iterative update process:
(\beta_{l}=\frac{\alpha_{l}}{\rho_{l}},a=pf\) (known), can be
\ (S (\cdot) \) is a nonlinear shrinkage function. \ (S (\cdot) \) is usually a smooth function.
Network structure:
including the reconstruction Layer \ (x^{(n)}\), convolutional layer \ (c^{(n)}=d_{l}x^{(n)}\), nonlinear transformation layer \ (z^{(n)}\), multiply sub-update layer \ (m^{(n)}\), where the nonlinear transformation function \ (s\) can be approximated by piecewise linear functions, Just learn the value of the function of the interpolation point.
Network Learning Parameters: Hyper-parameters in the model family, each layer can be different.
Network training:
The loss function is
Gradient Descent method training.
Reference documents:
Yangyan,sunjian,lihuibin,xuzongben, deep admm-net for compressive sensing MRI (NIPS2017)
https://arxiv.org/abs/1705.06869
Model-driven deep learning (admm-net)