SlowArchon438/smai-a4-models

SMAI Monsoon 2026 - Assignment 4 Models

Trained models on CIFAR-10 classification for Assignment 4.

Preprocessing Details

• Input Dimension: 32 x 32 x 3 -> 3072 flattened pixels.
• Scaling: Scaled to float64 in [0, 1] by dividing by 255.0.
• Standardization: Column-wise zero-mean and unit-variance computed strictly on the training partition statistics; zero-variance columns use divisor 1.

Models

1. Selected NumPy MLP (selected_numpy_mlp.npz)

• Architecture: 3072 -> 128 -> 64 -> 10
• Hidden Activation: ReLU
• Output Activation: Identity (with Categorical Cross-Entropy)
• Parameters: 402,250 weights and biases
• Initialization: He normal (variance 2/n_in) for weights, zeros for biases

2. PyTorch Baseline MLP (pytorch_mlp_p0.npz)

• Architecture: 3072 -> 128 -> 64 -> 10
• Hidden Activation: ReLU
• Training: 100 epochs on clean CIFAR-10 labels (p = 0)

3. PyTorch IOC-NN Clean Labels (pytorch_ioc_p0.npz)

• Architecture: 3072 -> 128 -> 64 -> 10
• Hidden Activation: ELU
• Output Activation: Pre-softmax convex logits
• Constraints: Weights in layers l >= 2 constrained non-negative (W >= 0, reset via e^(w-5) post-update)
• Training: 100 epochs on clean CIFAR-10 labels (p = 0)