NTU Machine Learning Final Project Proposal Notes

Posted on | Post modified | In |

NTU Machine Learning Final Project Proposal Notes

tags: NTU_ML Machine Learning

| Paper | Used Technique / Ingenuity| Suitable / Unsuitable Reason| Replace to | |:—–:|:———————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————- |:—————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————————— |:———-:| | [1] | Models overview
<li>3D maps of gray and/or white matter (deep learning models: six layer CNN, ResNet, and Inception V1)</li><li>vertex wise measurements from the surface-based processing (models BLUP and SVM)</li>
Model 1: Best Linear Unbiased Predictor(BLUP)</br>Model 2: Support Vector Regression</br>Model 3: Six-Layer Convolutional Neural Networks</br>Model 4: Specialized Six-Layer Convolutional Neural Networks for Younger and Older Subjects</br>Model 5: ResNet</br>Model 6: Inception V1</br></br> Additional Experiments<li>Different Types of Model Combination: Linear Regression vs. Random Forest</li><li>Combining Seven (Identical) Convolutional Neural Networks or the Seven Best Epochs</li><li>Influence of the Type of Brain Features on Prediction Accuracy</li> |Suitable:</br>In this field, it’s very clearly on comparing 6 variety models which can help us to know the implementation what we learned in class.</br>Also can aware of the result between high level model and custom level model</br></br>For linear regression and random forest, they trained the ensemble algorithms on a random subset. They repeated this process 500 times to get a bootstrap estimate of the SE of the MAE. | N/A | | [2] |2D and 3D-CNN on age estimation<li>For 2D-CNN, we consider the features as an image of size 168×60 (DH×M) ignoring the days as temporal information.</li><li>However, for 3D-CNN, we consider the features as a 3D volume with temporal information across the days, where each day has 24 hours and an hour is 60 minutes. So to break it down, we represent the features as a three dimensional information of 7×24×60 (D×H×M) minutes.</li>| Unsuitable:</br> Though the topic is interesting, the technique content is less then expectation and the .| No Idea Yet | | [3] | Model for classification:</br>Random Forest, GLMNet, SVM(including e1071, which is a package of LibSVM in R language, LiblinearR, kernlab, Rgtsvm), and xgboost</br></br>Calibration Algorithm(i.e. post-processing):logistic regression(GLM function), BRGLM, GLMNet</br></br>Performance evaluation: HandTill2001| Suitable:</br>The reason is as the same as [1] which also used various methods and compare it to other papers detailed.| N/A |

Appendix

  • The custom model in [1] proposed six-layer CNN

  • Self-defined ResNet in [1]

  • Self-defined Inception V1 in [1]

  • The whole result of experience in [1]

  • Architecture of the proposed DL methods in [2] Architecture of the proposed DL methods

  • The reuslt in [3]

Reference

[1]Machine learning workflows to estimate class probabilities for precision cancer diagnostics on DNA methylation microarray data [2]Estimating Biological Age from Physical Activity using Deep Learning with 3D CNN [3]Ensemble Learning of Convolutional Neural Network, Support Vector Machine, and Best Linear Unbiased Predictor for Brain Age Prediction