Machine learning models and over-fitting considerations

Machine learning models may outperform traditional statistical regression algorithms for predicting clinical outcomes.Proper validation of building such models and tuning their underlying algorithms is necessary to avoid over-fitting and poor generalizability,which smaller datasets can be more prone to.In an effort to educate readers interested in artificial intelligence and model-building based on machine-learning algorithms,we outline important details on crossvalidation techniques that can enhance the performance and generalizability of such models.

Personalized HRTF Prediction Based on Light GBM Using Anthropometric Data

This paper proposes a personalized headrelated transfer function(HRTF)prediction method based on Light GBM using anthropometric data.Considering the overfitting problems of the current training-based prediction methods,we use Light GBM and a specific network structure to prevent over-fitting and enhance the prediction performance.By decomposing and combining the data to be predicted,we set up 90 Light GBM models to separately predict the 90instants of HRTF in log domain.At the same time,the method of 10-fold cross-validation is used to score the accuracy of the model.For models with scores below 80 points,Bayesian optimization is used to adjust model hyperparameters to obtain a better model structure.The results obtained by Light GBM are evaluated with spectral distortion(SD)which can show the fitting error between the prediction and the original data.The mean SD values of both ears on the whole test set are 2.32 d B and 2.28 d B respectively.Compared with the non-linear regression method and the latest method,SD value of Light GBM-based method relatively decreases by 83.8%and 48.5%.

A combined algorithm of K-means and MTRL for multi-class classification

The basic idea of multi-class classification is a disassembly method,which is to decompose a multi-class classification task into several binary classification tasks.In order to improve the accuracy of multi-class classification in the case of insufficient samples,this paper proposes a multi-class classification method combining K-means and multi-task relationship learning(MTRL).The method first uses the split method of One vs.Rest to disassemble the multi-class classification task into binary classification tasks.K-means is used to down sample the dataset of each task,which can prevent over-fitting of the model while reducing training costs.Finally,the sampled dataset is applied to the MTRL,and multiple binary classifiers are trained together.With the help of MTRL,this method can utilize the inter-task association to train the model,and achieve the purpose of improving the classification accuracy of each binary classifier.The effectiveness of the proposed approach is demonstrated by experimental results on the Iris dataset,Wine dataset,Multiple Features dataset,Wireless Indoor Localization dataset and Avila dataset.

Flash flood susceptibility mapping using a novel deep learning model based on deep belief network,back propagation and genetic algorithm

Flash floods are responsible for loss of life and considerable property damage in many countries.Flood susceptibility maps contribute to flood risk reduction in areas that are prone to this hazard if appropriately used by landuse planners and emergency managers.The main objective of this study is to prepare an accurate flood susceptibility map for the Haraz watershed in Iran using a novel modeling approach(DBPGA)based on Deep Belief Network(DBN)with Back Propagation(BP)algorithm optimized by the Genetic Algorithm(GA).For this task,a database comprising ten conditioning factors and 194 flood locations was created using the One-R Attribute Evaluation(ORAE)technique.Various well-known machine learning and optimization algorithms were used as benchmarks to compare the prediction accuracy of the proposed model.Statistical metrics include sensitivity,specificity accuracy,root mean square error(RMSE),and area under the receiver operatic characteristic curve(AUC)were used to assess the validity of the proposed model.The result shows that the proposed model has the highest goodness-of-fit(AUC=0.989)and prediction accuracy(AUC=0.985),and based on the validation dataset it outperforms benchmark models including LR(0.885),LMT(0.934),BLR(0.936),ADT(0.976),NBT(0.974),REPTree(0.811),ANFIS-BAT(0.944),ANFIS-CA(0.921),ANFIS-IWO(0.939),ANFIS-ICA(0.947),and ANFIS-FA(0.917).We conclude that the DBPGA model is an excellent alternative tool for predicting flash flood susceptibility for other regions prone to flash floods.

基于三维分类网络的前列腺辅助诊断

现代医学对数据可视化、科学化的分析需求增加,也增加了对医学影像的依赖性。但对于计算机而言,生物图像极为抽象,生物图像识别至今仍处于探索阶段,同时,对大、复杂三维医学图像特征提取和图像识别难度大。目前采用卷积神经网络对三维医学图像进行训练处理,由于训练数据集数量不足,经常出现过拟合现象。针对这些问题,基于TensorFlow深度学习框架,提出了一种新的前列腺辅助诊断模型。模型优化了深度学习网络层次,采用较少的参数加快训练速度,还能降低过拟合的可能性,此外还利用两种数据扩展方式进行数据扩充,并采用了dropout方法以避免过拟合。训练及测试结果表明,模型能够对大部分前列腺三维图像进行分类,判断出图像是否存在异常,正确率超过70%,优于同种条件下训练出的3DAlexNet网络图片分类模型。

Improve Robustness and Accuracy of Deep Neural Network with L_(2,∞) Normalization

In this paper,the L_(2,∞)normalization of the weight matrices is used to enhance the robustness and accuracy of the deep neural network(DNN)with Relu as activation functions.It is shown that the L_(2,∞)normalization leads to large dihedral angles between two adjacent faces of the DNN function graph and hence smoother DNN functions,which reduces over-fitting of the DNN.A global measure is proposed for the robustness of a classification DNN,which is the average radius of the maximal robust spheres with the training samples as centers.A lower bound for the robustness measure in terms of the L_(2,∞)norm is given.Finally,an upper bound for the Rademacher complexity of DNNs with L_(2,∞)normalization is given.An algorithm is given to train DNNs with the L_(2,∞)normalization and numerical experimental results are used to show that the L_(2,∞)normalization is effective in terms of improving the robustness and accuracy.