CNN experiments on NIST SD19 handwritten characters
Modular CNN pipeline for systematic architecture search on the NIST SD19 handwritten character dataset. 0.7835 test accuracy.
Goal
Systematic exploration of CNN architecture decisions on NIST SD19: filter count, kernel size, layer depth, pooling strategy (max vs average), and activation functions (ReLU, sigmoid, tanh).
Setup
- Dataset: NIST SD19 handwritten characters
- Training: 50 epochs, batch size 64, Adam optimizer, categorical cross-entropy
- Evaluation: test accuracy 0.7835, micro/macro/weighted F1, per-class metrics, confusion matrix
Output
Each run saves: confusion matrix heatmap, sample prediction grid, misclassification log, per-class F1 report. Reproducible via fixed seeds.
Takeaway
Max pooling + ReLU reliably outperformed the alternatives on this dataset. The modular structure made ablations cheap to run.