🧠 Daily Study Log [2025-09-22]
Explored the idea of scenic bus routes, wrapped up DDPM study with key conclusions, and summarized results from the Garbage Classification project.

💡 Idea

• Proposed Scenic Bus Routes
• Concept: Evaluate bus routes not just for efficiency, but also for beautiful scenery, offering unique value for tourists and daily commuters.

📖 Paper Study

• Continued Denoising Diffusion Probabilistic Models (DDPM) study.
• Day 5 – Discussion & Conclusion
  • Diffusion models = simple, stable, and strong alternative to GANs/VAEs.
  • Strengths: stability, theoretical links to score matching & Langevin dynamics.
  • Weaknesses: slow sampling, heavy compute cost.
  • Future: faster sampling methods (DDIM, latent diffusion), wider domains (audio/video/text), hybrid generative models.
  • Final takeaway: “Add noise → learn to remove it”, foundation for Stable Diffusion, Imagen, DALL·E 2.

🗑️ Personal Project – Garbage Classification

• Goal: Classify waste images (6 categories) with CNN & transfer learning.
• Dataset: Kaggle Garbage Classification (~2,500 images, 6 classes).
• Methods:
  • Baseline CNN → ~60% accuracy
  • MobileNetV2 (frozen) → ~78%
  • MobileNetV2 (fine-tuned) → ~84%
• Findings:
  • Strong gains with transfer learning.
  • Weakest class = trash (recall ~62%).
  • Frequent confusion: plastic ↔ glass.
• Limitations: Small dataset, class imbalance, webcam input less reliable.
• Next Steps: More augmentation, background removal, test stronger models (EfficientNetV2, ViT).

✅ TL;DR

📍 New idea: Scenic Bus Routes for tourists & commuters
📍 Finished DDPM: strong alternative to GANs, foundation for modern diffusion models
📍 Garbage Classification: MobileNetV2 achieved ~84%, but needs better handling of class imbalance