Generative Modeling & Representation Learning Experiments

This repository contains a collection of ongoing research experiments focused on representation learning, generative modeling, and visual understanding. The work primarily explores Variational Autoencoders (VAEs), Conditional VAEs (CVAEs), and their extensions toward invariant learning and multimodal vision systems.

Note: These are research-in-progress notebooks and do not represent final results.

---

Research Areas

1. CVAE Optimization (Posterior Collapse Mitigation)

• Implementation of Conditional Variational Autoencoder (CVAE)
• Focus on improving latent space utilization
• Custom loss with β-weighted KL divergence
• Experiments with long training schedules to stabilize learning

📁 CVAE_KLD_Optimization.ipynb

---

2. Learning Invariances in Visual Representations

• Study of how neural networks learn invariances under:
  • Rotation
  • Translation
  • Scaling
  • Shearing
• Conducted on MNIST dataset
• Custom CNN architecture with controlled data sampling
• Evaluation using classification metrics (accuracy, F1, etc.)

📁 learning_Invariances_MNIST.ipynb

---

3. MedSigLIP + Deformable Image Registration

• Experimental integration of:
  • SigLIP-based vision encoders
  • VoxelMorph-style deformable registration
• Pairwise image alignment using learned transformations
• Custom registration losses and displacement analysis
• Early exploration toward medical image alignment with representation learning

📁 MedSigLip_+_VoxelMorph.ipynb

---

Tech Stack

• Python
• PyTorch
• Hugging Face Transformers (SigLIP)
• Torchvision
• NumPy, Matplotlib
• Scikit-learn

---

Notes

These notebooks serve as prototypes and experimental logs.
Future work includes:

• Extending CVAE applications
• Deeper study of covariance-aware representations
• Integration with Vision-Language Models (e.g., CLIP, SAM-style systems)

---