👁️ AIMS Coursework

CNN Architectures (CIFAR-10)

Late 2024 — From SimpleCNN to BN_AsimNet

CNN Batch Normalization Dropout Data Augmentation

About

Systematic exploration of CNN architectures for CIFAR-10 image classification. Investigating the effects of model complexity, regularization (Dropout, Batch Normalization), and data augmentation on classification accuracy.

View on GitHub →

Dataset: CIFAR-10

60K

Total images

10

Classes

32×32

Image size

RGB

Color

Classes: Airplane, Automobile, Bird, Cat, Deer, Dog, Frog, Horse, Ship, Truck

Model Architectures

1. SimpleCNN

2 conv layers + ReLU + 2 max-pool + 3 FC layers

2. ImprovedCNN

3 conv layers with increased filter depth + reduced FC

3. AsimNet

4 conv layers with increasing filter sizes + 512 FC neurons

4. BN_AsimNet ⭐

AsimNet + Batch Normalization after each conv + Dropout (0.25)

Results

Model Val Accuracy Val Loss
SimpleCNN ~51% 1.87
ImprovedCNN ~65% 1.20
AsimNet ~72% 0.98
BN_AsimNet ⭐ ~77% 0.82

Observations:

  • SimpleCNN: Baseline with limited capacity for complex patterns
  • ImprovedCNN: Benefits from deeper architecture and more filters
  • AsimNet: Higher capacity, better generalization but prone to overfitting
  • BN_AsimNet: BatchNorm + Dropout provide best balance of performance and generalization

Key Takeaways

  • Batch Normalization + Dropout together provide best results
  • Deeper networks capture more complex patterns (51% → 77%)
  • Data augmentation (random crop, horizontal flip) improves generalization
  • Regularization is essential for larger models to prevent overfitting