This article contains the summary slides for the Motor Third-Party Liability Claims Analysis and Prediction project.
This article studies motor third-part liability policies in France. I first explored the dataset with data visualization, then used the classic generalized Poisson linear model, ridge regression, lasso regression, and gradient boosted model to make the claim number predictions. I compared the error rates among different models at the end.
In this report, we investigated recent works in using Bayesian Convolutional Neural Networks with variational inference to perform image classification tasks. Based on ideas introduced by Shridhar et al. in 2019, we reproduced the Bayesian LeNet model and tested the performance of our network on image classification tasks with MNIST, CIFAR-10, and CIFAR-100 datasets. We as well compared the validation accuracies generated with Bayesian approach to those generated with frequentist inference under the LeNet architecture. At the end of the paper, we discussed possible reasons for the discrepancies and provided promising future works that can potentially improve the model performance in the original paper.
In this work, we summarized the concepts and methods done in the original paper of Wang et al. in 2017 and reviewed related influential literature. We further reproduced the model and evaluated the performance of the model on CIFAR-10 dataset, where we achieved 84.85% of test accuracy. At the end of the paper, we discussed possible reasons for discrepancies in results between our model and those of the original paper. We also provided promising future works that can potentially improve the model performance in the original paper.
In this project I studied the application of deep reinforcement learning (DRL) in solving problems in the MiniGrid gym environment and a classic control problem CartPole. Specifically, I employed the proximal policy optimization (PPO) algorithm which is a modified version of actor-critic policy gradient method. I choose two testing environments from the MiniGrid environment and the CartPole environment from OpenAI Gym to verify my implementations.
In contrast to the Multi-Layer Perceptron (ML) or fully-connected (FC) nets, Convolutional Neural Networks are a more powerful tool often used in the field of computer vision. In this article, I created a CNN using Tensorflow. The notebook is one of the assignments of the Deep Learning Course at Columbia University.
In this article, I compared the performance of DBScan and MeanShift algorithms on specially generated datasets. In addition, I examine the performance of two popular gradient-based optimization algorithms Adam & Adagrad by writing them from scratch. The notebook is based on one of the assignments of the Advanced Machine Learning Course at Columbia University.
In this task, I created the same MLP network that I created in the previous article, but with Tensforflow library this time. Also, I experimented with t-SNE in visualizing high-dimensional data in 2D for CIFAR10. The notebook is one of the assignments of the Deep Learning Course at Columbia University.
In this article, I wrote a two-layer neural network from scratch and applied it to the CIFAR-10 dataset. The notebook is one of the assignments of the Deep Learning Course at Columbia University.
In this article, I wrote two classifiers– Logistic regression classifier and Softmax classifier with stochastic gradient descent, and applied them to the CIFAR-10 dataset. The notebook is one of the assignments of the Deep Learning Course at Columbia University.
This article uses deep learning to predict the numbers in the images from the MNIST Dataset. I used Tensorflow Keras, setting the activation function as sigmoid function, and output function as softmax function.
This is the legendary Titanic ML competition on Kaggle. I use ridge regression, lasso regression, random forest and gradient boosted model to make the survival prediction and compare the error rates among different models.
This article uses logistic regression and the ROC curve to predict the credit card default rate of clients. I wrote the ROC curve function and compared the error rate in using different thresholds to classify default status.
This article investigates the important bias-variance tradeoff in a linear regression setting under squared loss. The framework is from the course, Statistical Machine Learning at Columbia University.
This article uses ML algorithms– forward selection – to determine which mutations of the Human Immunodeficiency Virus Type 1 (HIV-1) are associated with drug resistance.
This article shares a look-back option pricing program with Monte Carlo Simulation method.
This article shares a pricing program for an insurance contract whose premiums are payable annually for 20 years and provides whole-life annuity benefit, Death/ Total Permanent Disability coverage, and an endowment.
This article shares a program that can be run weekly to generate lists to track newly launched/ discontinued products in the insurance market.
This article shares the code script and the thesis I wrote in 2016, discussing the relationship between oil price and consumption expenditure.
Hosted on GitHub
©2020 Yi-Pei Chan. Powered by orderedlist.