From
2018/5/12 v1
2019/7/11 v2
-
Named Entity Recognition
• Compared with different NN structures, incl. Bi-LSTM, CNN+LSTM – F1 score 83.2%
• Utilised different embedding function, incl. Glove and Tensorflow Elmo -
Sentiment Classification NLP
• Compared with different structures, incl. Bi-LSTM + Global Soft Attention and SVM – Acc 0.771
-
Speech Recognition Industrial Project
• Utilised Bi-LSTM to create language speech recognition system transforming audio data to text
• Audio → MFCC → NN → Target: CTC Loss (Beam Search, Edit Distance) -
Kaggle - House Prices: Advanced Regression Techniques
• Achieved top 6% and accuracy of 0.11 for the RMSE
• DNN, Linear Regression & Stacking Model (Lasso, Elastic Net, SVR, Kernel Ridge, Bayesian Ridge, Ridge) -
Kaggle - Histopathologic Cancer Detection
• Two CNN structures, incl. Fastai DenseNet 201 & NASNet + global max/average pooling – Acc 95.9% -
Machine Learning Mobile Application – Second Hand Car Selling
• Model 1: Utilised NN to create a price prediction system
• Model 2: Utilised Latent Dirichlet Allocation to create a topic system
• Model 3: Utilised Google Cloud Vision API to create an image recognition system
• Workflow: Created ML models (Python) → Imported 3 models to API → Created a mobile application -
Machine Learning Startup Project – Crime Warning System
• Market Demand: Theft offences occurred at a rate of 45.8 per 1,000 population in 2017 in London
• Workflow: Dimension Reduction → ARIMA & LSTM → Validation → Output Prediction
• Function: Once users walks on a road at a specific time, they will be reminded if the time is dangerous -
Kaggle – Titanic Machine Learning from Disaster
• Achieved top 2% and accuracy of 83.7% for the prediction of the data
• Utilised 2 methods to predict labels respectively, incl. DNN & Random Forest