Divya7106/smai-a4-iocnn

IOC-NN (3072-256-10, ELU)

SMAI Assignment 4, Q1 (CIFAR-10, true labels, p = 0).

• Architecture: 3072-256-10, hidden activation elu, output identity (logits), CCE loss, 789,258 parameters.
• Training: CIFAR-10 train part (50,000 images), SGD, learning rate 0.02, batch size 64, 100 epochs, seed 13010.
• Preprocessing: flatten each 32x32x3 image to 3072 values, divide by 255, then subtract mean and divide by std from preprocessing.npz (train-part statistics).
• Files: model.npz holds W0, b0, W1, b1 with W of shape (n_in, n_out), so logits are elu(x @ W0 + b0) @ W1 + b1.
• Results: train accuracy 0.9476, test accuracy 0.4999.