Projects

• Deep Neural Network Compression - Per-layer Quantization
  This project tries to simulate currently unsupported fix-pointed computation to reduce the computation overhead of inference caused by using full-precision weights to predict the results.
  
  
• Generalization of "Rethinking the Value of Network Pruning"
  "Rethinking the Value of Network Pruning" states that model architecture is more important than it's weight. This project generalized the result of this paper by conducting similar experiments upon imbalanced datasets.
  
  
• Process Scheduler Implementation
  This project implements a multilevel feedback queue scheduler in NachOS, including the following features: preemptive SJF queue, aging, execution-time estimation, round-robin, assignable priority, priority scheduling.
  
  
• K-NN Accelerator
  This project implements a K-NN hardware Accelerator in Verilog with open-sourced PicoRV32 CPU. This project also implements a Software version of KNN written in provided firmware as a baseline to compare with the accelerator. 
  
  
• FTP Proxy for transfer-rate limitation
  This project implements a FTP Proxy written in C Socket. This proxy can limit data transfer rate by blocking the packets sent to meet the demanded overall transfer rate. 
  
  
• Simple Shell in OS-less Environment
  This project implements a mini-uart based simple shell running on OS-less rasberry-pi3 b+ hardware. The shell is first developed in Linux with QEMU/gdb. Languages included are ARM assembly, linker Script and C