Researches And Applications Of Traffic Flow Forecasting And Quantitative Analysis Of Celiac Disease By Deep Learning Methods

Because of the existence of "semantic gap",it has long been essential to extract features for nature data based understanding,recognition,and classification tasks.However,only low level features can be extracted with these conventional feature extraction methods,and different feature extraction algorithm need to be designed for different tasks,which requires designers to have a deep understanding of the data in the certain domain and have a prior domain knowledge.The feature extraction method often determines the success of the whole system,and the difficulty of designing and tuning of the low level features is obstructing the progress of the community of artificial intelligence.The emergence of deep learning techniques makes it possible to start from the data themselves and learn the relationship inter and inner the data,which helps to finish the final tasks,through a unified scheme.This scheme is through multi-layer network.Each layer learns different level abstraction of data from the bottom to the top,the higher level the more abstract.At present,the performance of deep learning in image classification,Go game and other fields have been close the level of human beings.Therefore,deep learning based data processing technologies has become the importance research area.Aiming at deep learning based data processing technologies this paper proposes a series of key technologies of deep learning based data processing on traffic flow data and capsule endoscopy video data.In this paper,the main work is as follow:1.We propose a-agree Adaboost stacked autoencoder,which solves the statistical problem,computational problem,and presentational problem of the stacked autoencoders based traffic flow forecasting method under some certain conditions.This adaptive framework ensembles stacked autoencoders with different initial state and different deep architectures.We introduce a data replication strategy to force the sequent stacked autoencoder to take more consideration on the samples with large prediction error.Then the importance of each stacked autoencoder is determined by the discriminative error.The final prediction is by exhaustively search over the feasible solutions,which can satisfy as many stacked autoencoders as possible according to their importance.We conducted our case study on the dataset collected from the motorway of Amsterdam,Netherland,which ends on the ring road of Amsterdam.The results demonstrates that the proposal achieves more accurate performance than the conventional traffic forecasting models,and achieves more accurate performance than individual stacked autoencoder and ensemble of the simple models.2.We propose to integrate a set of representative models into an unified ensemble framework and exploit stacked autoencoder networks(SAE)to select a optimal model or a optimal subset of models to perform traffic flow forecasting in a real-time manner according to the current situation.In order to train the SAE,we acquire the training dataset from historical real data and develop a model-driven mechanism to automatically label them.We then train the SAE based on the training dataset and fine-tune the trained model by taking use the probability obtained in the labeling stage.Once the SAE is established,we can input the current traffic flow and the SAE will choose a optimal model or a optimal subset of models based on the probability distribution.To efficiently and flexibly leverage the probability distribution for forecasting,we implement three strategies to integrate the candidate models according to their probabilities,namely conditional expectation,maximum probability and selective integration.In our implementation,we choose six representative models as the candidates while our framework is extensible to include more models.To validate the effectiveness of the proposed framework,we perform extensive experiments on three representative traffic flow datasets: Netherlands Amsterdam motorways dataset,the dataset of Caltrans Performance Measurement System(PeMS)and the dataset from the Traffic Data Acquisition and Distribution(TDAD)system.Experimental results demonstrate the proposed method outperforms state-of-the-art models and achieves much more accurate forecasting results under large variations and uncertainties by dynamically selecting optimal model(s).3.We propose a quantitative GoogLeNet-based computer-aided celiac disease diagnosis method.We first extract the frames from the capsule endoscopy video of celiac disease patients and the normal controls.Then the informative part of the frames are cropped,and intensity corrected.The frames are pre-rotated before training the GoogLeNet.A confident evaluation method is proposed to figure out how confident the deep network indicates a subject as a celiac disease patient.In the experiment,we find that the deep network not only succeeds to discriminate celiac disease patients by the subtle change on the small intestine mucosa,but severity level of pathology also takes positive correlation with the calculated confident evaluation.This method achieve 100% of sensitivity and specificity patient-wise on the test set by cross-validation.