CIFAR-100 Image Classification with CNNs

Exploring CNN architectures, noise robustness, and transfer learning for 100-class image recognition

View the report ->

---

Highlights

Model	Clean Accuracy	Noisy Accuracy	Accuracy Drop	Parameters
CNN (scratch)	69.1%	1.2%	67.8%	~2.4M
CNN (noise-augmented)	64.9%	47.3%	17.6%	~2.4M
VGG-16 + MLP	63.3%	1.5%	61.8%	138M + 157K

---

Project Structure

├── CIFAR100_cnn_image_classification.ipynb       # Main experiment notebook
├── index.html       # Interactive HTML report
├── README.md       # Interactive HTML report
├── checkpoints/
│   ├── a1_cnn_best.pth        # Best CNN model (clean training)
│   ├── a3_cnn_noisy_best.pth  # Best noise-augmented CNN
│   └── a4_mlp_best.pth        # VGG-16 MLP classifier
├── images/                    # All training plots & visualizations
└── requirements.txt

---

Overview

This project investigates CNN-based approaches for classifying images from CIFAR-100 — a challenging benchmark with 100 fine-grained categories and only 500 training images per class. We address three key questions:

1. How well can a moderate CNN perform when trained from scratch?
2. How robust are learned features to input noise, and can we improve robustness?
3. Does transfer learning from ImageNet provide better features?

---

Architecture

The CNN uses three convolutional blocks with progressive regularization:

Block 1: Conv(3→64) → BN → ReLU → Conv(64→64) → BN → ReLU → MaxPool → Dropout(0.2)
Block 2: Conv(64→128) → BN → ReLU → Conv(128→128) → BN → ReLU → MaxPool → Dropout(0.3)
Block 3: Conv(128→256) → BN → ReLU → Conv(256→256) → BN → ReLU → MaxPool → Dropout(0.4)
Classifier: AdaptiveAvgPool(1,1) → FC(256→512) → ReLU → Dropout(0.5) → FC(512→100)

Design choices:

• Batch normalization after each convolution for stable training
• Increasing dropout rates (0.2 → 0.5) for progressive regularization
• Kaiming He initialization for proper gradient flow
• ~2.4M parameters — moderate and efficient

---

Experiments

CNN Trained from Scratch

Metric	Value
Training Accuracy	79.5%
Validation Accuracy	69.0%
Test Accuracy	69.1%

Training: Adam (lr=1e-3, weight_decay=5e-4), CosineAnnealingLR, 100 epochs, batch size 128.

Noise Robustness

We inject additive Gaussian noise (σ²=0.05) to evaluate model robustness:

The standard CNN drops from 69.1% → 1.2% under noise — a catastrophic 67.8% degradation. By training with 30% noisy samples, the noise-augmented model achieves:

• 64.9% clean accuracy (only −4.2% trade-off)
• 47.3% noisy accuracy (vs 1.2% without augmentation)
• 17.6% accuracy drop (vs 67.8%)

Transfer Learning (VGG-16)

Using frozen VGG-16 features with a lightweight MLP classifier (157K trainable params):

Condition	Accuracy
Clean Test	63.3%
Noisy Test	1.5%

VGG-16 achieves competitive accuracy with minimal training, but is equally vulnerable to noise.

Full Comparison

---

Key Findings

• Best clean accuracy: CNN from scratch (69.1%) — small and effective for CIFAR-100
• Best robustness: Noise-augmented training reduces accuracy drop by 4× (67.8% → 17.6%)
• Transfer learning trade-off: VGG-16 provides strong results with minimal training, but no inherent noise robustness
• Noise augmentation works: A simple 70/30 clean-to-noisy training ratio dramatically improves resilience

---

Quick Start

pip install -r requirements.txt
jupyter notebook CIFAR100_cnn_image_classification.ipynb

---

References

• Krizhevsky, A. Learning Multiple Layers of Features from Tiny Images
• Simonyan, K. & Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition (ICLR 2015)

---

Author

• Name: Kushal Ghosh
• Email: gkushalg01@gmail.com
• GitHub: @gkushalg01
• LinkedIn: linkedin.com/in/kushal-ghosh