Application Research Of Deep Learning In Prognosis Assessment

Deep Learning(DL) is composed of multiple hidden layers. It is a kind of network model simulating human brain mechanism, and is a learning process to extract feature through the mapping of the Low-level data to the High-level. Deep Learning is capable of being effectively and efficiently utilized to represent complex function and to classify complex data. So far, it has already successfully applied in speech recognition, image recognition and the like. Deep Learning has two learning models, i.e. unsupervised and supervised model. The unsupervised model includes Convolutional Neural Network (CNN). Deep Belief Network (DBN) is a classical supervised DL model. In this paper, Deep Belief Network is researched.The traditional Deep Belief Networks utilizes reconstruction error as the evaluation criterion of Restricted Boltzmann Machine (RBM) networks in training process, which could reflect the likelihood between RBM network and training samples to some extent. However, it is not reliable. Maximum Information Coefficient (MIC), which is based on the estimations of Shannon entropy and conditional entropy, identifies interesting relationships between pairs of variables in large data sets and can be utilized to capture a subset of highly related features. The MIC could be used as a criterion for evaluating a learned model since it is robust to outliers. In order to construct the model that fits data well and to reduce classification error, this paper proposed a Deep Belief Networks Based on MIC method. MIC is utilized not only to reduce dimensions but also to improve the reliability of the network learning. Classification experiments are performed on handwriting data sets of MNIST and USPS by several traditional methods and Deep Belief Networks Based on MIC method. The results show that the latter can effectively improve the recognition rate.Besides, it’s difficult for the traditional Deep Belief Networks to determine network parameter setting, e.g. the learning-rate of network. When the value of learning-rate is too large, the network convergence speed is high but the learned network is easily instable. On the contrary, when the value of learning-rate is too small, the instability in the network might be avoided but the network convergence speed is low. Hence, it is very important to determine an appropriate learning-rate. The traditional method to obtain the parameters of network is through the try-error experience or the learning. The global optimal of parameters could be acquired by utilizing Genetic Algorithm (GA). In order to effectively and efficiently determine a global optimal of parameters that fit data well and reduce classification error, this paper advanced a Deep Belief Network Based on Genetic Algorithm method. Classification experiments are performed on handwriting data sets of MNIST and USPS by several traditional methods and Deep Belief Networks Based on Genetic Algorithm method,. The results exhibit that the latter can improve the recognition rate and the learning speed effectively.The evaluation of the prognosis of ICU has important significance to alleviate the shortage of medical resources and the rational allocation of resources. When the Deep Belief Network is applied to ICU data, an ICU evaluation system Based on Deep Belief Networks is brought forward. Classification experiments are performed on ICU data sets by traditional Deep Belief Networks methods, Deep Belief Networks Based on MIC method and Deep Belief Networks Based on Genetic Algorithm method, the results indicate that the Deep Belief Network is able to play an important role in the prognosis of ICU.