Query Understanding (summer intern / 2014)
In this project, we improved the knowledge graph (KG) recommendation system for Yahoo Taiwan search engine. This project involved two stage: KG intent detection and KG recommendation. In the KG intent detection stage, we used logistic regression classifier to detect whether to provide KG to users. In the recommendation stage, we first used semantics analysis to find out some synonyms of predicate terms, and then crawled on the knowledge graph to find out the best KG.
- Teamwork with Michael Liao and Kevin Lee. - Awards: 2014 Yahoo Intern Technical Awards. (1 out of 11 groups) |
Spoken Term Detection and Spoken Content Retrieval (2014)
In this project, we used spoken term detection (STD) and spoken content retrieval (SCR) techniques on NTCIR-11 SpokenDoc&Query task. In SCR, we used pseudo relevance feedback, query expansion, and graph-based re-ranking techniques.
- Teamwork with Po-wei Chou. - Publication: NTCIR 2014 [pdf] [slide] [poster] |
Travel Note Retrieval System (2013)
In this project, we built a retrieval system for travel notes using either image or text query. Users can use text-text, text-image, image-image modes in this retrieval system. In text retrieval part, we implemented OKAPI/BM25 as vector space model and rocchio relevance feedback; in image retrieval, we used features, such as SIFT, Gabor and color histograms, and we also applied sparse representation. The similarity score was calculated as cosine similarity.
- Teamwork with Wei-Chiu Ma and Wei-che (Tony) Tsai. - Result: [report] |
Right Whale Detection on ICML 2013 Data (2013)
In this project, we attended the ICML 2013 right whale detection competition on Kaggle. we first transformed the audio signal into spectrogram, then used template matching approached to find out similar patterns to the sounds of right whale in the specific range of frequency. In addition, we extracted features from template matching, then applied gradient boosting classifier to detect the right whale sounds.
- Awards: fourth place in this competition. - Honors: first place in Music Signal Analysis and Retrieval class. (1 out of 40+ teams) |
Classification on KDD 2012 Click Data (2013)
In this project, we aimed to solve the bipartite ranking problem over the products recommendation data of KDD 2012 competition. Given the features extracted with 10% missing data, we first implemented expectation-maximization (EM) algorithms to make up the missing data, applied dimension reduction and feature selection techniques, and then used classifiers such as naive bayes, linear regression, logistic regression, KNN, SVM, random forest, matrix factorization and gradient boosting.
- Honors: First place in Machine Learning class (1 out of 60+ teams). - Teamwork with Wen-Ding Lee and Chia-Han Kuo. - Report: [report] |
Smart Phone Games (2012)
In this project, we built a smart phone game on Android platform using TCP/IP protocol. Two users can connect to each other and start the battle mode.
- Teamwork with Ting-chun Wang and Jong-chi Su. - Demo: [video] |
In this project, we built a smart phone game on Android platform using TCP/IP protocol, augmented reality (AR) and image processing. First two teams in this game wear YELLOW and BLUE clothes, connect each other, and aim to the other team's member through screen and then shoot. We used image processing techniques to detect the color patch, and then applied AR technology into this game.
- Teamwork with Ting-chun Wang and Jong-chi Su. - Demo: [ppt] [video_trailor] [video_short] [video] |
Object Recognition (2012)
Face Recognition (2011)
In this project, we implemented Eigenfaces and Fisherfaces on face images as features, and then classified face images using support vector machine (SVM). Furthermore, if the confidence score of classification was too low, we took it as non-seen face. In this system, we provided the function that user can add new faces and re-train the SVM model.
- Teamwork with Chi-Yun (Diana) Tsai. - Demo: [video] - Result: [report] |
Lego Robot (2011)
In this project, we constructed a LEGO robot as a ping-pong shooter with motor and camera. We used OpenCV for image processing to detect the red target in tasks.
There were three tasks for this project: 1. fixed point and target. 2. fixed point and moving target. 3. moving point and fixed/moving target. - Teamwork with Ru-shane (Carolyn) Hua, Zi-Yi Yang and Den-Han Yu - Demo: fixed point, fixed target [video] fixed point, moving target [video] moving point, fixed target [video] |