CIFAR-10 Standard MLP (Clean Labels p=0)
Model submission for SMAI Assignment 4, Question 1 (Section 1.5).
Architecture Details
- Architecture:
[3072, 128, 64, 10] - Parameters: 402,250
- Hidden Activation: ReLU
- Output Activation: Linear Logits (CCE loss)
- Initialization: He Normal ($std = \sqrt{2 / n_{in}}$), biases zero
- Optimizer: SGD ($lr = 0.05$, batch size = 64, epochs = 100)
Preprocessing Details
- Dataset: CIFAR-10 (
uoft-cs/cifar10) - Split: 70% Train (35,000), 15% Val (7,500), 15% Test (7,500)
- Seed:
roll_number % 100000 = 2024101114 % 100000 = 1114 - Flattening: $32 \times 32 \times 3 \rightarrow 3072$
- Standardization: Zero-mean and unit variance per feature computed on train set in
float64.
Performance
- Train Accuracy: 0.8507
- Test Accuracy: 0.4518