b05702057 Goto Github PK

anomaly-detection

I used KNN, PCA, and different decoders to detect anomaly in the testing data, and compare their accuracy. The goal of my project is to make sure that the correct predictions of a model is actually correct. For instance, people usually focus on the wrong predictions to modify the model. However, if we overlook the anomalies, the model can actually perform worse than our perception.

audi-used-car-market-analysis

This report includes the python web crawler code to collect used cars data, the R code of regression analysis, and a comprehensive report that can help you understand how people value the used cars of Audi.

binary-income-prediction-model

In this project, I implemented two models, the generative model and the logistic regression model to infer whether a person earns more than $50000 per year. To improve the accuracy, I tried three main operations. First of all, I did tiral and error to find the best feature combination. Also, I normalized all the data in the numpy array. Lastly, I tried regularizations to avoid overfitting.

bomberman-game

I wrote a bomberman game that includes two players and several game items, like shoes and magic powder which can speed up the players or strengthen the power of their bombs. Also, I made the game items appear on the map or under the wooden box randomly.

chocopy-compiler-subset

In this project, I started from the previous Chocopy compiler to design and implement the heap-manipulating parts to define and use classes.

chocopy-compiler-without-heap-memory

In this project, I started from the previous mini compiler to design and implement a bigger compiler for all but the heap-manipulating parts of ChocoPy.

data-structures-and-algorithms-for-criteo-big-data

In this project, I implemented C++ stl vectors to store 6.4GB data containing 1.6M clicking information of the company named Criteo. There are 4 main functions. The first one is to get the information whether a user makes a purchase of the product he clicks at a specific time. The second one is to find the common purchase of two users. The third one is two know the conversion rate(purchases/clicks) of specific users. The last one is to know all the products a user buys. These operations can cost a lot of time because there are 1.6M pieces of data. My main work was to store these data in a specific way so that I could get information by implementing binary search.

dynamic-load-balancer

english-mandarin-translation-model

I Built a sequence to sequence model with advanced machine learning techniques like attention, schedule sampling, and beam search to translate English into Mandarin. The attention mechanism can consider more complicated relationships between words of a sentence. The schedule sampling mechanism can adjust the way the model translate a word. The beam search mechanism can help consider much more situations.

fault-tolerant-surfstore-system-with-raft

In this project, I implemented the log replication part in Raft to handle server crashes or network partition, provider clients with the majority answer.

fault-tolerant-tribbler

In this project, I extended the previous Tribbler system and made it fault-tolerant.

file-system-call-implementation

Nachos is an incomplete operating system. Initially, Nachos have the creat, open, read, write, close, and unlink file system calls that cannot be accesses by user processes. My implementation converts the existing system calls to an interface that can be used by user processes. With the interface, user processes can execute the calls easily and safely, without ruining the underlying operating system.

food-image-classification

In this project, I implemented a CNN model to classified food pictures. To improve the accuracy, I conducted data normalization, dropout, weight decay, and so on. Also, I implemented the techniques of explainable AI, trying to interpret how my model learned so that I could enhance it.

google-search-result-web-crawler

This code can help automatically download all the search results of the topic you are interested in as PDF files in your computer.

group-chocopy-wasm-compiler

Python REPL (similar to Jupyter/Colab Notebooks) compiled to WebAssembly in your browser. Written primarily in TypeScript, this was a multi-team class project for CSE 231: Advanced Compiler Design

hotel-booking_company-revenue-prediction

In this project, We did trial and error with models like the neural network, random forest, xg-boost, and so on to know which model is the best to predict the revenue and the cancellation rate. We also used advanced techniques of early stopping, soft labels, and ensembling to improve the prediction accuracy.

landscape-image-classification

I Built a classification model with unsupervised learning algorithms like Kmeans and dimension reduction techniques like PCA, t-SNE, and autoencoder. The model can also be used to reconstruct the original image.

mail-query-system

In this project, we implemented different data structures and search algorithms to query the mails, including the unordered map in stl.

mini-compiler

In this project, I started from a baseline compiler provide in class, adding binary arithmetic operators and a few built-in Python functions.

moba-game-matching-system

Nachos is an incomplete operating system. In this project, I implemented the alarm function to make a thread sleep temporarily, a join function to block a thread until its child finishes, a condition variable class to wait for or signal other threads, a sleepFor function to wake a thread when it has waited too long, and a cancel function to cancel the alarm when a thread is waken up because of the condition variable. Then, I designed a game matching system that can start games after a given number of players with the same ability join.

model-attacking-defending

In this project, I implemented FGSM and the basic iterative method to attack a pre-trained model. Then I tried to protect my model by doing randomization to the images before I feed them into my model.

network-compression

In this project, I hoped to compress the model size to less than 300000 bytes so that the model can be used on a wider variety of devices. Thus, I conducted network compression techniques like architecture design, knowledge distillation, network pruning, and weight quantization.

operation-management-algorithm-design

In this project, we designed an algorithm to optimize a factory operation considering the floating material price each day, the storage cost, the startup cost, the initial storage, the productivity capacity constraint, the number of machines, the conversion rates between stations, and so on.

operation-research-computing-center-scheduling

In this project, we designed an algorithm to optimize the operation to perform works from student submissions, utilizing linear programming to consider the number of computers each task needs and the time each subtask takes.

pm2.5-prediction-model

In this project, I used data which includes the amount of PM10, PM2.5, NO, NO2, the rainfall, temperatures, and so on to predict the PM2.5 in the future. First of all, I conduct EDA to do the feature selection. Then I tried the different gradient descent algorithms and the closed-form solution to find the global minima. Lastly, I adjust the learning rate and training epoch to get a better result.

quant-trading-webpage

I Designed a professional website for a quant trading firm with EJS and RESTful Routes, creating servers with Express, and utilizing MongoDB to manage client data

scalable-tribbler-infrastructure

In this project, I utilized the previous storage RPC to build a Twitter-like system for clients to post and follow others.

storeage-rpc

In this project, I implemented a key-value storage service that is called via RPCs.

stream-cipher-decryption

In this project, I iterated all the possible timestamps until I successfully decrypted some texts that make sense, utilizing the same timestamp as the random seed to get the original file.

street-voice-data-analysis-report

In this project, we crawled data from the leaderboard of Streetvoice and analyze the data to infer how artists should promote themselves when they have songs to release.