SYDE 577 CNN for MNIST
SYDE 577 assignment two: get a convolutional net past 99% test accuracy on MNIST. Started from an MLP baseline, then built a configurable CNN with variable kernel size, dilation, and depth, and searched it with Optuna over five pruned trials. Best run: 99.25% validation accuracy with four conv layers, a 7x7 kernel, no pooling.