CIFAR-10 Classification on Symmetric and Asymmetric Noise

Overview

This project explores the impact of label noise on the CIFAR-10 dataset and evaluates strategies to mitigate its effects. The study incorporates symmetric and asymmetric noise at varying levels (10%, 30%, 50%, 80%, and 90%) and applies several machine learning methodologies to enhance noise robustness. Three model architectures are developed and compared:

1. Base CNN with Cross Entropy Loss
2. Regularized CNN with Symmetric Cross Entropy Loss
3. Advanced Regularized CNN with Normalized Cross Entropy and Reverse Cross Entropy Loss

Dataset

The CIFAR-10 dataset consists of 60,000 32x32 color images across 10 classes, divided into:

• Training set: 50,000 images
• Test set: 10,000 images

Label noise is introduced to simulate real-world data imperfections:

• Symmetric Noise: Random mislabeling across all classes.
• Asymmetric Noise: Systematic mislabeling of specific classes into predefined others.

Models

Base CNN + Cross Entropy Loss

• A simple convolutional neural network architecture.
• Uses traditional Cross Entropy Loss.
• Designed as a baseline for comparison.

Regularized CNN + Symmetric Cross Entropy Loss

• Enhances the Base CNN by adding batch normalization and dropout layers.
• Implements Symmetric Cross Entropy Loss to balance robustness and accuracy.

Advanced Regularized CNN + NCE and RCE Loss

• Introduces Kaiming and Xavier initialization.
• Employs the Active Passive Loss framework, combining Normalized Cross Entropy (NCE) and Reverse Cross Entropy (RCE).
• Optimized with SGD and fine-tuned hyperparameters.

Methodology

1. Noise Incorporation: Simulate symmetric and asymmetric noise in training data while keeping validation and testing data clean.
2. Data Preprocessing:
   • Convert images to tensors.
   • Normalize pixel values (mean=0.5, standard deviation=0.5).
   • Batch size: 64.
3. Training: Models trained on noisy datasets with varying noise levels, using different loss functions and optimizers.
4. Evaluation: Compare test accuracies under different noise levels for each model.

Results

The models\u2019 performance under symmetric and asymmetric noise conditions:

• Base CNN: Struggles with high noise levels, overfitting noisy labels.
• Regularized CNN: Improved robustness but shows limitations under extreme noise.
• Advanced Regularized CNN: Demonstrates superior resilience, maintaining higher accuracy across noise levels.

Key Findings:

• NCE+RCE Loss outperforms Symmetric Cross Entropy Loss in high-noise scenarios.
• Regularization techniques like dropout and batch normalization enhance model generalization.

Experimentation

• Experiments conducted on Google Colab Pro using NVIDIA Tesla V100 GPU.
• Models trained with 10 random seeds to ensure robustness.
• Training configurations include fixed epochs, learning rate tuning, and noise-specific parameter adjustments.

Dependencies

• Python (>=3.7)
• PyTorch
• NumPy
• Matplotlib

Usage

1. Clone the repository.
2. Install required dependencies.
3. Run the scripts to reproduce experiments:

   python train_model.py --model {base|regularized|advanced} --noise {symmetric|asymmetric} --level {10|30|50|80|90}