Hsiang-Jui (Jerry) Lin

A book introduces leJOS framework on LEGO NXT robots [lejos].
ISBN: 9789862768228

The study focuses on total and paired domination problems on distance-hereditary graph. Discovered an alternative optimal solution to total domination based on split decomposition. The handout can be found in the link.

The thesis combines deep learning and formal methods to tackle the occlusion in image tracking-by-detection problem. It developed a method to establish strong relationship among sensor data and learned representations, and achived short-term prediction of vehicle movements based on the material derivative method. The model is trained on completely labeled data is evaluated on partially occluded data. It shows 0.3 meter displacement error in average.

Develope robotic features on WYSIWYG Android programming. Developed Lego EV3 and MediaTek MT7697 support [ai2-contrib].

The project exploited free storage by abusing Tor’s Onion service. It provided a virtual file system that stored data blocks as RSA public keys in disguise, and distributed them to the rest of Onoin network. (github)

It designed a robotic arm capable of grabbing objects using RealSense. The software combined computer vision and position estimation of objects using Intel RealSense depth camera, and instructed the arm to grab objects accordingly. (github)

The project developed the type level programming language for Rust. It achieved compile-time, type-level arithmetic computations. It was developed to pave the way for building type-safe tensor types in Rust. (github)

It developed building blocks for asynchronous parallel data flow for Rust. It was develped to serve image data processing that mix heavy I/O and computation tasks. It provided iterator-style stream combinators to distribute data and scale on workers, and is designed to support Rust’s async/await in nature. (github)

The project reproduced YOLOv4 object detection model using tch-rs and libtorch in Rust. It was designed with parallel, asynchronous and unordered data flow in mind, provided multi-GPU training, and supports modular deep learning model configurations. (github)

The lab project develops a complete data processing system for traffic scenes. The tasks convered data collection, equipment design, data processing pipeline and data analysis. The vision is to provide a data processing pipeline to generate high precision data from LiDAR to aid the performance of camera based models. The project information can be found in the lab page.

Can write CUDA (example), is able to construct concurrent data strucutres from scratch. (example)

Can design and deploy machine learning models, e.g. SVM (link), image generation models (link), contribute to several machine learning and computer vision libraries, e.g. TensorFlow’s Record File for Rust. (link)

Participated in several Capture the Flag (CTF) contests in college. Can exploit weakness of a system. ([torfs])

Studied game theory, combinatorics, abstract algebra, cryptography, computational logic and discrete differential geometry in college. Had an undergraduate thesis on graph theory ([undergraduate-thesis]).