PaleoNet
PaleoNet application interface showing dinosaur image classification results
Demo of PaleoNet classifying a dinosaur image with CNN model predictions
Training progress visualization showing model accuracy and loss curves over epochs
Model performance metrics including accuracy, precision, recall, and confusion matrix
A machine learning assignment that grew into a genuine deep learning challenge: train a convolutional neural network to automatically identify dinosaur species from images, starting with a target accuracy of 70% and pushing it further through systematic experimentation.
What I Built
PaleoNet is an end-to-end image classification pipeline built with TensorFlow and Keras, from raw dataset preparation through to a live interactive demo deployed on Streamlit. The model takes an input image and returns a ranked list of predicted dinosaur species with confidence scores.
Technical Highlights
- Designed and iterated on multiple CNN architectures, tuning depth, filter sizes, and pooling strategies to find what generalised best on the limited dataset
- Applied data augmentation techniques — random flips, rotations, and zoom — to artificially expand training variety and reduce overfitting
- Used NumPy and Pandas for dataset wrangling, label encoding, and tracking per-epoch metrics across experiments
- Achieved a final classification accuracy of just above 80%, exceeding the initial 70% target
Architecture
The training pipeline preprocesses images to a fixed resolution, passes them through stacked convolutional and max-pooling blocks, and feeds the flattened output into a dense classifier head with dropout regularisation. Training progress — loss and accuracy curves — was logged each run, making it straightforward to compare the effect of each architectural change.
The finished model is served through a Streamlit web app, where anyone can upload an image and receive a live prediction.
Technologies Used
- AI
- CNN
- TensorFlow
- Streamlit
- Keras
- Python
- NumPy
- Pandas