Digit Recognizer (MNIST)

Classify handwritten digits (0-9) from 28x28 grayscale images.

Computer Vision Accuracy Simple CNN

Feature Importance

Conv1 0.25 Conv2 0.22 BatchNorm 0.15 MaxPool 0.12 Dropout 0.10 Dense1 0.08 Softmax 0.08

Key Insights

Simple CNN achieves 99.5%+. Batch normalization after each conv layer is essential. Data augmentation (rotation, shift) gives minor but meaningful gains.