Implementation of Mask R-CNN on Python 3, Keras, and TensorFlow to
detect the area of damage on a car. The model generates bounding
boxes and segmentation masks for each instance of car in the image.
It's based on Feature Pyramid Network (FPN) and a ResNet50/ResNet101
backbone.Photos of damaged car can be input into model to assess
damage.
Developed a Q&A chatbot that answers question in natural language
based on the content of the given passage that is easy to integrate
with any existing website. This TensorFlow Natural Language Question
Answering model is based on a pre-trained BERT model fine-tuned on
Stanford Question Answering Dataset (SQuAD) 2.0 dataset.
Prediction vs actual result of the FProphet model
Multi-steps forecast result of the FBProphet model
Built an automated system that will fit time-series dataset into
different Statistical and Machine Learning Models for multi-step
forecasting and return result from the most accurate model.
LSTM (Long Short-Term Memory networks) Models for Time Series Forecasting
Univariate Multi-Step time forecasting with Encoder-Decoder Model
Multiple Input Multi-Step forecasting with Stacked LSTM Model
Developed different LSTM (Long Short-Term Memory networks) models to
forecast univariate & multivariate time series dataset. Models are
evaluated with 8 metrics.
Application of the Temporal Fusion Transformer (TFT), a novel
attention-based architecture which combines high-performance
multi-horizon forecasting with interpretable insights into temporal
dynamics to forecast time series data.
Neural NILM: Deep Neural Networks Applied to Energy Disaggregation
Undergraduate research by Yuzhe Lim in Spring 2019 on the topic of
Deep Neural Networks application on NILM (Nonintrusive
load monitoring) for Energy Disaggregation.
Two-Class Boosted Decision Tree Model using Azure Machine Learning Studio
Model training pipeline
Model deployment pipeline
Built a Machine Learning pipeline in Microsoft Azure Machine
Learning Studio to train, evaluate, and deploy a binary
classification model for predicting an individual's income in the
US. The estimator used in this project is a Two-Class Boosted
Decision Tree classifier. The pipeline includes
Data Cleaning: Substitute missing values and exclude irrelevant columns
Upsampling training data to account for Class Imbalance
Training the model and hyperparameter tuning
Scoring and Evaluating the Model
Deploying the Trained Model as a Web Service for inference
A pipeline is also created to compare the performance between one model trained on the upsampled data and the other with the original pre-processed data.
Built a two-layer neural network and an L-layer neural network from scratch.
Implemented all the building blocks of a neural network for image classification model:
Developed 4 different types of recommendation systems using data from
The Movie Database (TMDb) to provide relevant movie suggestions
through unique filtering processes.