During undergraduates, I had one year as a visiting student at NTU, Singapore where I worked with
Professor Siew Kei Lam and
Dr. Thinh Pham on a hardware acceleration
project for ORB object detection in self-driving cars system.
During the time at NUS, I studied the effects of different motion prediction models on the performance of motion planning algorithms.
I am interested in autonomous driving, especially in motion prediction and planning.
My current research focuses on planning under uncertainty by formulating the problem as a POMDP.
I am fascinated to explore how we can have an interpretable, safe, and robust planning for self-driving car.
Towards that ends, I posit the consideration of three guiding design concepts:
(i) robust vision to detect objects (e.g., occupancy network, scene flow estimation),
(ii) jointly trained all modules in stack to reduce error propagation and
(iii) language models to intelligently plan and explain the decision.
In this work, I extend our developed engine to cater for three main requirements from Y3 Technoligies: compatibility constraints (relationship between vehicles and locations), dynamic time/distance matrix (due to road construction) and optimized costs for each delivery routes.
I developed two neural network models: i) WCGAN combined with linear regression on tabular inputs and ii) Spatial-Temporal network in addition with VAE on images data to predict the crimes' locations and occurences. I also implemented local search to plan ahead of the patrol shift work plans with minimum time shift between locations.
Collaborative Urban Delivery Optimization
Fujitsu-SMU Corp Lab, March 2019 - March 2021
In this work, I used machine learning to predict the estimated time of arrival (ETA) between two locations. I proposed a time-slot division strategy to find the lowest number of vehicles before the neighborhood search and post-optimized the schedule plans with redistributions of orders while violating minimum of soft constraints. I also engineered the frontend and backend of the system.
Sashankh Chengavalli Kumar, Tran Phong, Wong Wei Hao
In this project, we aim to study deep learning approaches to a text classification task, by attempting to predict the genre of a film based on its plot in a single-label multi-class classification problem.
We designed cache-efficient (cache-oblivious and cache-adaptive) algorithms for matrix multiplication and experimented on a simulated cache and a real computer.
We developed Deep Q-Learning to automate the Duckiebot to follow the lane and reduce speed near the traffic light. We also employed Autoencoder network to extract good features.
Vietnamese Character Recognition
HCMUT, Vietnam, Sep. 2017
Le Minh Quang, Tran Phong
We collected comprehensive data of all Vietnamese characters to support the thesis proposal and applied a standard CNN model to recognize Vietnamese characters with 95% top-1 accuracy.
LC3b CPU
HCMUT, Vietnam, Feb. 2017
Tran Phong
We designed a core CPU (LC3b) architecture of Illinois University with Verilog language, verified by VCS tool in Linux environment, and evaluated on FPGA (DEII Kit of Altera company). We proposed a pipeline architecture that functioned properly sixteen machine instructions and achieved 125 Mhz.
