Handwritten Letter CNN
Twenty-six handwritten letters, from a dense baseline to a locked CNN.
- Context
- Deep Learning coursework, individual
- My part
- Everything: data checks, experiments, error analysis
- Stack
- Python, TensorFlow, Keras, CNN
- Year
- 2026
94.95%
Test accuracy across 26 classes, macro F1 0.949
- What it is
- Classifying 28×28 EMNIST-style letters into 26 classes.
- What I did
- I ran data-integrity checks and a dense baseline first, then CNN candidates through controlled validation experiments chosen on macro F1, ending at three conv blocks, heavy dropout and a 128-unit dense layer.
- Result
- What is left clusters on letters people confuse too: I and L, G and Q, U and V. Feature maps and misclassified examples are read as clues, not claimed as explanations.
Up next
All work
