Research the use of Large Multimodal Models in Medicine

Research novel augmentation methods for increasing robustness of neural network models. e.g using GANs for image translation.

Tutor for lecture "Introduction to Deep Learning" at Technical University of Munich, winter semester 21/22. I am responsible for checking exercises and tutorial slides before publishing it, answering students' questions about the lecture on online forum (Piazza) and hold an office hour every week.

I help to set up the website for ScanNet Data Efficient benchmark. This benchmark is based on ScanNet benchmark, a benchmark for evaluating several scene understanding tasks such as 2D/3D semantic label, 2D/3D semantic instance prediction using ScanNet dataset. The dataset is RGB videos containing 2.5 million views in more than 1500 scans, annotated with 3D camera poses, surface reconstructions, and instance-level semantic segmentations. ScanNet Data Efficient benchmark aims to evaluate algorithms trained on less data than original ScanNet dataset, hence the name Data Efficient. An algorithm having high performance on ScanNet Data Effcient is expected to learn a good representation of data. My contribution: - Set up the web interface based on existing ScanNet benchmark interface using Javascript - Use MySQL to create database to store user's information as well as performance result of their algorithm - Use PHP in the backend to link Python scripts which evaluates several scene understanding tasks

During my internship, I help to develop a prototype for new forklift truck using Arduino Due. The toy truck runs on a track, which simulates the narrow path in storing center. Along the track are many RFID tags, which send signal to the truck when to stop to load or unload goods. The truck is equipped with an RFID reader to receive and analyze the signal. An encoder is attached to the truck to measure distance travel and the speed. The vehicle is powered by current supplied to the it directly through the track. As it travels to the end of the track, it automatically reverses the direction and travels backwards, thanks to an H-bridge circuit. The information of the truck, such as speed, distance travel, information from the tags, is displayed on an LCD attached to it, as well as transmitted to the server through an antenna. Thanks to that, operator can supervise all operations of the truck and interrupt when things happen. All functions are set using Arduino and MATLAB Simulink, with the help of C. Simulink is a software used to simulate real-time operation. It uses graphical programming rather than text programming, meaning Simulink blocks are connected to create a complete logical program. It offers many predefined blocks for Arduino. However, many components do not have their own block, which requires me to create new Simulink blocks for them using C and turn it into MATLAB code.

-Develop an autonomous 4-wheeled vehicle that can be controlled byvoice and automatically plans a path to a goal defined by user’s voice-Use Robotic Operating System(ROS) and SLAM algorithm for pathplanning and CMUSphinx for voice recognition-The robot can avoid collision that suddenly appears and plan another route towards the goal-Thanks to Kinect sensor, a 3D map is also constructed as the robot moves