About Project
A complete machine learning pipeline for handwritten digit classification using a custom neural network built with TensorFlow and trained on the MNIST dataset. The project includes data preprocessing, model architecture design, training with validation, and an interactive web canvas where users can draw digits and receive real-time predictions. The model achieves 97.8% accuracy on the test set using a carefully tuned architecture.
Key Features
MNIST Training
Custom neural network architecture trained on 60,000 MNIST samples with data augmentation for robust generalization.
Drawing Canvas
Interactive HTML5 canvas with pressure-sensitive drawing, stroke smoothing, and automatic 28x28 preprocessing for prediction.
Real-Time Prediction
Instant digit classification with confidence scores displayed as an animated probability bar chart for all 10 digits.
Model Visualization
Interactive visualization of network layers, weight distributions, and activation heatmaps showing what the network learns.
Batch Processing
Upload multiple handwritten digit images for bulk classification with CSV export of results and accuracy metrics.
Model Export
Save and load trained models in HDF5 format with version tracking and training history logs.
Tech Stack
Project Preview
The web interface features a large drawing canvas on the left with real-time prediction results on the right. Confidence bars animate as the user draws, with the predicted digit displayed prominently. A gallery below shows recent predictions with actual vs. predicted labels.