Project · AI & machine learning
Applied Neural Networks
Prediction across tabular, image, and time-series data.
How should a model change when its inputs shift from rows and columns to images and chronological sequences? I explored that question by building three predictive systems in Python, handling the full workflow from data cleaning and feature engineering to model training, validation, and performance analysis.
- Python
- TensorFlow
- Keras
- pandas
- scikit-learn
Three models, three kinds of data
NYC Taxi Fare Prediction
Tabular · Regression
A dense regression network that estimates taxi fares from passenger count, trip distance, pickup time, day of the week, and geographic coordinates. The workflow covers data cleaning, temporal and geographic feature engineering, standardization, training and validation curves, and evaluation with R‑squared and mean squared error.
Chinese Numeral Recognition
Image · Classification
A 15-class neural-network classifier for 64×64 handwritten Chinese numeral images: normalized pixel values, dropout regularization, separate training, validation, and test sets, and error analysis through a confusion matrix and misclassified examples.
Melbourne Temperature Forecasting
Time series · Forecasting
A comparison of Simple RNN and LSTM architectures on rolling 30-day temperature sequences, each evaluated with mean absolute error and used to generate a recursive seven-day minimum-temperature forecast.
The process
Prepare and clean the data
Engineer or normalize features
Select an architecture for the data type
Train and validate the model
Evaluate predictions and investigate errors
Export models, charts, forecasts, and metrics
What I learned
- Model architecture should follow the structure of the underlying data.
- Evaluation methods must match the prediction task.
- Validation curves and error analysis are as important as headline accuracy.
- Reproducible outputs make experiments easier to compare and communicate.