Summary: In recent years, deep learning has achieved breakthrough success in a series of difficult problems in the field of artificial intelligence.
Model-driven deep learning method
In recent years, deep learning has achieved breakthrough success in a series of difficult problems in the field of artificial intelligence. For example, the face recognition is higher than the person's correct recognition rate, the speech recognition and machine translation are close to the level of simultaneous translation and ‵ 完 finished, the level used for the Go game has reached the victory of the human world champion, The diagnosis for some diseases can match the level of middle and senior professional doctors. Deep learning techniques are now ubiquitous in all areas of information science and are becoming standard methods in their respective fields.
Despite significant progress in deep learning, there is still a lack of theoretical understanding of the relationship between artificial neural network topology and performance. Network topology selection is still an engineering technology and has not become science. This directly leads to the fact that the existing deep learning is mostly a heuristic method lacking theoretical basis. Difficulties in design, difficulty in interpretation, and unpredictable results have become recognized defects in deep learning.
The National Science Review recently published a "Model-driven deep learning" article by Professor Xu Zongben and Professor Sun Jian from the School of Mathematics and Statistics, Xi'an Jiaotong University (National Science Review, 2017, https://doi.org/10.1093/ Nsr/nwx099). This article attempts to address the problem of network topology selection for deep learning in order to implement deep learning methods that can be designed, interpreted, and predictable. This paper proposes a deep learning method combining model-driven and data-driven. As we all know, deep learning is a standard data-driven method. It uses a deep network as a black box to rely on a large amount of data to solve real problems. The model-driven method starts from the goal, mechanism, and a priori to form a cost function of learning. Then solve the problem by minimizing the cost function. The biggest advantage of the model-driven approach is that as long as the model is accurate enough, the quality of the solution can be expected to be optimal, and the solution method is deterministic, but the drawback of the model-driven approach is that it is difficult to accurately model in the application, and to model The pursuit of precision can usually only be a luxury.
The model-driven deep learning method effectively combines the advantages of model-driven and data-driven methods. The standard flow of model-driven deep learning is given in the paper: (1) building model family according to the problem; (2) according to the model Family, design family of algorithms and establish the convergence theory of algorithm families; (3) unfold the algorithm family into deep networks and implement deep learning. The paper also introduces a series of in-depth methods such as model-driven and data-driven research and practice, which demonstrates the effectiveness of the method in solving practical problems.
2-in-1 Laptop,2 In 1 Tablet,Rugged Tablet,Detachable Laptop
C&Q Technology (Guangzhou) Co.,Ltd. , https://www.gzcqteq.com