Research Projects in Semiconductor Optics
- Spectroscopy and Simulation Revealed Strong Interaction of Insulating States in Moiré Superlattice
- Fabricated a WSe2/WS2 moiré Superlattice with piezo stage. Performed SHG spectroscopy to determine the crystal orientation.
- Implemented Microwave Impedance Microscopy to study the low conductivity at correlated insulating states.
- Performed PL spectroscopy to study the strong interaction between interlayer excitons and correlated electrons in superlattice.
- Identified integer filling of Mott insulator and fractional fillings of Wigner crystals, with opposite valley polarization.
- Metasurface Integrated 2D Material WSe2–SiN system for Photonic Crystal Cavities
- Fabricated SiN metasurface by LPCVD and inductive-coupled plasma etching. Coupled monolayer WSe2 to SiN metasurface.
- Simulated the guided mode resonances (GMR) of the WSe2-SiN photonic crystal cavities by RCWA, FDTD and COMSOL.
- Performed Energy-Momentum Spectroscopy in self-built back-focal-plane imaging setup. Observed Rabi splitting of 18meV.
- Demonstrated the existence of exciton–polariton with high Q factor~143, inspiring future applications in lasers and displays.
- Unit Valley Polarization Quantum Bits Control in WSe2/MoSe2 Heterostructure
- Fabricated twisted heterostructure under microscope. Wrote electrodes by E-beam lithography. Characterized devices by AFM.
- Identified spin-triplet and singlet interlayer excitons with PL spectroscopy by tuning electric, magnetic field and temperature.
- Conducted helicity resolved PLE spectroscopy, discovered 100% valley polarization for future quantum computing application
Research Projects in Machine Learning
- Accelerating Decentralized Momentum SGD in Large-batch Deep Learning for Object Detection
- Conducted baseline model based on PmSGD and DmSGD for object detection on Cifar-10 and ImageNet dataset using Pytorch.
- Derived convergence analysis of SGD models in both non-convex and strongly convex scenarios for algorithm optimization.
- Evaluated performance of DecentLAM with state-of-art models (Faster-RCNN, YOLO) on COCO dataset, achieved linear speedup as PmSGD, reduced inconsistency bias exponentially by (1-β^2), saved 60% communication costs per iteration.
- Stochastic Optimization and Data Compression in Distributed Learning for Image Classification
- Developed deep learning model based on LeNet5 and FCNNs for image classification on Fashion-MNIST and Cifar10 datasets.
- Applied Top-k Sparsification and Low-bit Quantization for data compression, saved memory usage by 77%.
- Simulated the distributed/federated learning based on Error Compensated SGD, got 93.6% accuracy with 20% time reduction.
- Real-time Parking Space Detection
- Captured 600+ parking lots images from video with OpenCV and pre-processed images based on Grayscale Transform.
- Applied multiple image processing methods for image augmentation, expanded train dataset size by 10 times.
- Conducted image segmentation based on Hough Transform, Harris Corner Detector and Canny Edge Detector.
- Performed feature detection and matching, homography estimation with MOPS descriptor, SIFT descriptor and RANSAC.
- Built and trained CNN model based on VGG16 and MobileNet for identification using Keras, got 92.4% validation accuracy.
- Created a real-time parking space detector to predict the available or occupied parking spots from 600+ parking lots
- Used Car Market Forecast
- Mined and engineered 400k historical trades over 12 months for used car markets to predict the car price for future trades.
- Performed EDA and visualized the time series and the correlation matrix of features with price using python.
- Engineered 40 key features including Make, Model and Mileage, performed PCA to extract the usable data after encoding.
- Predicted car price with multiple tree-based prediction models (XGBoost, LightGBM, Catboost), achieved MAE as 180.
- Designed a fully connected Neural Networks with activation function of softplus, tuned the regressor with 5-fold cross validation and trained with iterative decreasing learning rate, reduced MAE by 26.7%.
- News Recommendation System
- Developed data pipeline to clean and label 3M clicks data over 0.36M articles from MySQL database with python.
- Calculated cosine similarity and SVD Matrix Factorization to quantify the similarity between articles.
- Lessened recommendation list from item-based collaborative filtering (ItemCF), top-k clicking news and embedding retrieval.
- Identified 12 key metrics for recommendation model, including news category, words count, user logs and click timestamp.
- Trained GBDT+LR (Logistic Regression), LightGBM and DIN (Deep Interest Network) model, performed model stacking for predictions with the evaluation metric of mean reciprocal rank (MRR), achieved MRR as 0.3110.
- Built a recommendation system to predict 5 most recommended articles based on user’s click history, targeting for 0.3M users
