Spaces:
Sleeping
Sleeping
| # personal information | |
| ## identification | |
| Singapore Permanent Resident|Chinese citizen | |
| ## address | |
| 17 Jalan Masjid, Singapore | |
| ## contact | |
| yingxu.he1998@gmail.com|+65 91752741|+86 15063250971 | |
| # Working Experience | |
| ## Machine Learning Engineer at Huawei Ltd. | |
| • from Dec 2022 to present | |
| • Built a pipeline to automatically visualize data tables using LSTM network trained on ChatGPT-generated | |
| data with pairwise loss method, achieving 80% recall@5 on 100+ internal test cases. | |
| • Designed and implemented a novel SISR method that enhanced WIFI-signal simulations for office buildings | |
| by achieving 10x speedup compared to physics-based simulation with negligible loss in accuracy (1% MAE) | |
| on over 80 large-scale office layouts. | |
| ## Machine Learning Research Engineer at Dyson Ltd. | |
| • from Sept 2021 to Dec 2022 | |
| • Implemented an object localization model in a few -shot context by semi -supervised training. The model | |
| achieved comparable results to professional software with improved adaptability and robustness . | |
| • Designed and implemented an air quality estimation model, using LGBM, Bayesian Regression, etc., with | |
| geographical and meteorological features . Demonstrat ed its advantages over spatial interpolated methods | |
| and deployed the pipeline with Metaflow framework on AWS services. | |
| ## ML Research Assistant at NUS -Singtel Cyber Security Lab | |
| • from Sept 2020 to July 2021 | |
| • Identif ied anomalies from system logs leveraging DBSCAN and hierarchical clustering for model training . | |
| • Developed an information retrieval method for web -attack strategy identification from system and firewall | |
| logs. The recall@3 rate achieved 80% on 100+ hand -labelled samples . | |
| ## Data Analyst Intern at GIC Pte. Ltd. | |
| • from Dec 2018 to July 2019 | |
| • Deployed an R application that forecasts the mid -term returns of portfolio with visualization using R shiny . | |
| • Optimized the coefficients of a mean reversion forecasting model using the Genetic Algorithm. | |
| ## Data Analyst Intern at PropertyGuru | |
| • from May 2018 to Aug 2018 | |
| • Developed dashboard s in Tableau to analyze the user behaviors and listings’ performance to better match | |
| user demand to agents’ recommendations. | |
| • Implemented a POC to calculate and geographically visualize the liveability score for properties . | |
| # Education | |
| ## Master of Computing in Artificial Intelligence at National University of Singapore | |
| • from Aug 2020 to Sept 2021 | |
| • School of Computing : CAP 4.42/5.0 | |
| • Teaching Assistant : Advanced Analytics and Machine Learning (from Jan 2021 to May 2021) | |
| ## Bachelor of Science (Hons) in Business Analytics at National University of Singapore | |
| • from Aug 2016 to June 2020 | |
| • School of Computing : CAP 4.15/5.0 , Dean’s List in Semester 3 AY 2018/2019 | |
| • Distinction : Analytics Techniques Knowledge Area (awarded in Dec 2020) | |
| • Teaching Assistant : Programming Methodology in python (from Aug 2017 to June 2018) | |
| # Relevant Projects | |
| ## Distilling ChatGPT for finetuning image captioning models | |
| • from Jan 2023 to Present | |
| • Employed Chain -of-Thought with verification prompting technique on ChatGPT to create 10k+ accurate | |
| capt ions from the xView annotations. Fine -tuned a GIT image captioning model and significantly improved | |
| the CIDE r score from 11.59 to 85.93 over 2k RSICD samples. | |
| ## Dialogue Response Generation ( Master Thesis ) at NUS NExT++ Lab | |
| • from Nov 2020 to Aug 2021 | |
| • Built an enriched task -oriented response generation by implementing copy -mechanism on GPT -2 using | |
| Pytorch. The proposed model is capable of naturally incorporating external tips/user reviews about venues | |
| into responses. The generated response outperforms m any state -of-the-art models on user satisfaction. | |
| ## Property Resale Price Prediction | |
| • from Jan 2021 to May 2021 | |
| • Fitted CatBoost, LGBM, XGBoost on 43k pieces of property sales data. Selected features by correlation and | |
| information gain. Engineered new features describing properties’ livability. Reduce d data dimensionality | |
| with WOE encoding. The f inal ensemble methods’ accuracy achieved 5th/64 place. | |
| # Skills | |
| • Python (Pytorch, Tensorflow), R : Machine | |
| Learning, Deep Learning , Data processing | |
| • SQL, Spark: Data query and big data | |
| • Tableau, PowerBI : Visualization development | |
| • Java, Git, Scala, JavaScript, HTML, CSS : Software | |
| Development | |