tahamajs/Deep_UT/CA6_Generative_Models - Storage Bucket

CA6: Generative Models

This assignment explores generative modeling techniques, implementing GAN-based domain adaptation and Variational Autoencoders (VAEs), demonstrating the power of generative approaches in unsupervised learning and domain transfer.

Overview

The assignment consists of two generative modeling projects:

1. Unsupervised Domain Adaptation with GANs: CycleGAN for domain transfer
2. VAE for Anomaly Detection: Generative modeling for medical imaging

Assignment Snapshot

• Primary Notebooks:
  • Unsupervised_Domain_Adaptation_GAN/code/NNDL_CA6_1.ipynb
  • VAE/code/NNDL_CA6_2.ipynb
• Benchmarks: Adapted classifier accuracy ≈ 87.6% on MNIST-M with FID ≈ 38.7; VAE PSNR ≈ 28.5dB, SSIM ≈ 0.89, ROC AUC ≈ 0.90 for anomaly detection.
• Key Tricks: Cycle-consistency + identity losses for stable stylization, β-VAE control over disentanglement, and reconstruction-based anomaly scoring.

Contents

• Unsupervised_Domain_Adaptation_GAN/: CycleGAN implementation for domain adaptation
• VAE/: Variational Autoencoder for anomaly detection

Each subfolder contains:

• code/: PyTorch implementations
• description/: Assignment specifications
• paper/: Research papers and references
• report/: Detailed analysis and results

Unsupervised Domain Adaptation with GANs

Key Features

• Cycle Consistency: Bidirectional mapping between domains
• Domain Confusion: Adversarial alignment of feature distributions
• Unsupervised Learning: No target domain labels required
• Style Transfer: Realistic transformation of visual appearance

Technical Details

• Generator Networks: U-Net style with residual blocks
• Discriminator Networks: Patch-based discrimination
• Loss Components: Adversarial loss + cycle consistency + identity loss
• Training Strategy: Alternating optimization with careful loss balancing

Results

• Target Accuracy: 87.6% on MNIST-M (vs. 75.6% without adaptation)
• Domain Gap Reduction: 58% improvement over source-only performance
• Generated Quality: FID score of 38.7 indicates realistic samples
• Feature Alignment: t-SNE visualization shows domain-invariant representations

Visual	Description
	A↔B translations showing style adaptation quality.
	Classifier predictions on adapted MNIST-M glyphs.

VAE for Anomaly Detection

Key Features

• Probabilistic Encoding: Amortized variational inference
• Reparameterization Trick: Enables gradient-based optimization
• Anomaly Scoring: Reconstruction error as anomaly indicator
• Medical Application: Polyp detection in gastrointestinal endoscopy

Technical Details

• Encoder: CNN-based recognition network (μ, log σ²)
• Decoder: Transpose CNN for image reconstruction
• ELBO Loss: Reconstruction + KL divergence regularization
• β-VAE Variant: Tunable regularization strength

Results

• Reconstruction Quality: PSNR 28.5dB, SSIM 0.89 on normal images
• Anomaly Detection: AUC 0.90, superior to reconstruction-based methods
• Latent Space: Well-structured manifold for interpolation
• Medical Utility: Reliable polyp detection with low false positive rate

Visual	Description
	Input vs. reconstruction comparison for normal samples.
	Latent interpolation illustrating continuity/disentanglement.

Key Concepts Demonstrated

Generative Adversarial Networks

• Adversarial Training: Generator vs. discriminator optimization
• Cycle Consistency: Ensuring bidirectional mapping quality
• Mode Collapse: Avoiding degenerate solutions
• Training Stability: Techniques for stable GAN training

Variational Autoencoders

• Variational Inference: Approximate posterior distributions
• Reparameterization: Enabling backpropagation through sampling
• ELBO Optimization: Evidence lower bound maximization
• Latent Space Regularization: Controlling latent representations

Domain Adaptation

• Domain Shift: Handling distribution differences between datasets
• Unsupervised Adaptation: Learning without target labels
• Feature Alignment: Matching source and target distributions
• Transfer Learning: Leveraging knowledge from related domains

Anomaly Detection

• Reconstruction-based Methods: Using generative models for anomaly scoring
• One-class Classification: Learning from normal data only
• Threshold Selection: Determining anomaly decision boundaries
• Evaluation Metrics: Precision, recall, F1-score for imbalanced data

Educational Value

This assignment provides comprehensive understanding of:

• Generative Modeling: Creating new data samples from learned distributions
• Adversarial Learning: Training through competition between networks
• Unsupervised Learning: Learning without explicit supervision
• Probabilistic ML: Bayesian approaches to deep learning

Dependencies

• Python 3.8+
• PyTorch 1.9+
• torchvision
• NumPy, Pandas
• Matplotlib, Seaborn
• scikit-learn
• OpenCV (for image processing)

Usage

Domain Adaptation GAN

1. Navigate to Unsupervised_Domain_Adaptation_GAN/code/
2. Prepare source and target domain datasets
3. Train CycleGAN for domain transfer
4. Evaluate adaptation performance in Unsupervised_Domain_Adaptation_GAN/report/

VAE Anomaly Detection

1. Navigate to VAE/code/
2. Train VAE on normal data
3. Compute reconstruction errors on test data
4. Analyze anomaly detection results in VAE/report/

References

• Domain Adaptation Description
• VAE Description
• Research Papers | Research Papers
• Implementation Reports | Implementation Reports

---

Course: Neural Networks and Deep Learning (CA6) Institution: University of Tehran Date: September 2025