Fresh-butter/smai-a4-mlp

SMAI Assignment 4 — MLP on CIFAR-10

• Architecture: 3072 → 512 → 256 → 128 → 10 fully connected; hidden activation ReLU; output = 10 logits (no activation, softmax + cross-entropy loss).

• Weights: weights.npz with W0, b0, W1, b1, ...; W has shape (inputs, outputs), so logits = h @ W + b.

• Preprocessing: flatten the 32×32×3 image (row-major, channels last) to 3072 numbers, divide by 255, then standardise each column with mean and std from preprocessing.npz (computed on the 50,000 CIFAR-10 training images).

• Training: CIFAR-10 train part (true labels, p = 0), plain SGD, lr 0.05, batch 64, 100 epochs, seed 11030.

• Result: train accuracy 0.9576, test accuracy 0.5025 (10,000 test images).