Jiawei Lu
MSEE student @ Columbia Univeristy
Hi, I’m Jiawei Lu from Columbia University in the City of New York.
I’m in the first year of my MSEE Program in School of Engineering and Applied Science.
My research interest is Deep Learning, Computer Vision, Reinforcement Learning and IoT.
I’m more than happy to establish a connection with you!
- Location
- 403 West 115th St, NY 10025, New York, United States
- jl5999@columbia.edu
- Phone
- (+1) 917-679-9297
- Website
- https://jiawei-lu.com
- GitHub
- JiaweiLu1999
- Jiawei Lu
Education
–
Master in Computer Science & Electrical Engineering
from Columbia University with GPA of 3.80/4.00
Courses
- COMS4732 - Computer Vision II: Learning
- EECS6691 - Advanced Deep Learning
- COMS4705 - Natrual Language Processing
- COMS4771 - Machine Learning
- ECBM4040 - Neural Networks & Deep Learning
- ELEN6885 - Reinforcement Learning
- COMS4118 - Operating Systems I
- EECS4764 - Internet of Things
–
Bachelor in Physics & Electrical Engineering
from Nanjing University with GPA of 4.46/5.00
Courses
- Mathematical Analysis
- Electromagnetism
- Analog Circuits
- Digital Signal Processing
- Fundamentals of Acoustics
- Algorithm Design & Analysis
Projects
Improvements of Active Object Localization with Deep Reinforcement Learning :
–
In this project, we proposed several improvement in four aspects, including using advanced CNNs to generate state representation, defining more flexible action spaces, changing reward function to avoid undesired activity in agent and using mask instead cross for multiple objects. You can find our final paper here.
Highlights
- Replaced feature extractor part of Q-Network with several advanced CNN network in state space.
- Proposed a more flexible 25-action model and used extra trigger training to avoid the unbalance of trigger samples.
- Changed reward function to avoid undesired activity in agent.
- Improved Inhibition of Return mechanism by a new prediction algorithm for multiple objects.
A New Backbone for Hyperspectral Image Construction and Improvement based on Mask Mixture Training and Energy Normalization :
–
We implemented mask mixture training and energy normalization. Those two changes have been proved to successfully improve the robustness of the SSI-ResU-Net. Besides, a combination of those two modifications can further improve the generalizing ability of SSI-ResU-Net. You can find our final paper here.
Highlights
- Implemented a modified version of U-Net named SSI-ResU-Net.
- Utilized Mixed Training and Energy normalization to improve the accuracy of prediction.
Smart Vending Machine :
–
We design a Smart Vending Machine based on Raspberry Pi 4B. We can buy things in two ways: Face Login or Android App. Here is a live demo of our product!
Highlights
- Designed an Android App with Register, Login, TopUp and BuyThings functions.
- Created Cloud Data Base based on Firebase to mange user information.
- Utilized face recognition to enbale fast login. Deep Neural Network is applied to detect the face location in one image, and the face is encoded to a Tensor.
Motion Planning in Understructured Road Environments with Stacked Reservation Grids :
–
We propose a new data structure named stacked reservation grid (SRG). To test our proposed approach, we build a dataset named the Berkeley Deep Drive Drone (B3D) dataset and describe how to develop a validation procedure using the B3D dataset.
Highlights
- Used the Annotation Tool to select pixels in each frame of the video in the data set.
- Designed the corresponding algorithm to improve the picture stability.
- Used the Georeferencer plugin in QGIS and wrote Python script to achieve the correspondence between image pixel coordinates and QGIS coordinates, thus realizing image deformation and re-projection.
- Realized video stabilization by applying color detection to according program.
Application of Reinforcement Learning in Single-Channel Speech Enhancement System :
–
This paper establishes an adaptive Q-Learning model based on the characteristics of DeepXi, and designs a reinforcement learning algorithm named Xi-Q. The Xi-Q algorithmuses Perceptual Evaluation of Speech Quality (PESQ) to design the reward function,and utilizes the estimated a prior SNR calculated by Deep Xi system to give a set ofactions. It selects the actions under the ϵ-greedy strategy to increase the maximumexpected reward by balancing ”exploration-exploitation”.
Highlights
- Designed a reward function for the output a priori signal-to-noise ratio of the Deep Xi system based on the Q-Learning algorithm of reinforcement learning.
- Predicted the action-value function Q between the input signal and the output a priori signal-to-noise ratio estimation using a deep neural network based on a deep learning framework.
- Validated the effectiveness of the proposed reinforcement learning-based self-optimization algorithm Xi-Q Algorithm for Deep Xi system based on PESQ.
A Regional Intelligent Vehicle Dispatching Method for Intersections without Traffic Lights :
–
National Undergraduate Innovation and Entrepreneurship Training Program.
Highlights
- Took advantage of the CNN to predict future traffic flow information.
- Planned the path in stages based on the idea of rasterization and optimal path algorithm according to the prediction of future traffic flow and the real-time change of traffic flow, so as to realize the effect of dynamic path planning.
- Showed the whole scheme in the form of video based on U3D through writing C# script to produce animation effect.
Awards
Winner of Honorable Mention & Group Leader
from Mathematical Contest in Modeling
Highlights
- Led the team to make use of time series model to adjust relevant parameters so the accuracy of test sets was more than 80%.
- Took advantage of Python API to obtain emotional tendency of text semantics, and then predict the rating and attitude of comments.
National Scholarship
from Ministry Of Education
National scholarships are the highest honorary national scholarships available to students in China higher education.
Skills
- Deep Learning
- Level: MasterKeywords:
- Reinforcement Learning
- Level: IntermediateKeywords:
- Kernel Development
- Level: IntermediateKeywords:
- Computer Vision
- Level: IntermediateKeywords:
- Web Development
- Level: JuniorKeywords:
Languages
- Python
- Fluency: Native speaker
- Java
- Fluency: Professional working proficiency
- C
- Fluency: Professional working proficiency
- Matlab
- Fluency: Professional working proficiency
- C++
- Fluency: Limited working proficiency
- Chinese
- Fluency: Native speaker
- English
- Fluency: Professional working proficiency
Interests
- Travelling
- Keywords:
- Coding
- Keywords:
- EDM
- Keywords:
- K-Dramas
- Keywords:
