🧠 Daily Study Log [2025-07-01]
Today was a well-balanced study day across three key areas: competition experimentation, hands-on GAN practice, and my very first formal paper review!

📊 SCU_Competition — Submission 36–40

Focus: Comparing model types under fixed feature structure
Goal: Find the sweet spot between simplicity and generalization

Highlights:

• Submission 36: Full feature set + plain LGBM (no tuning) → CV AUC 0.8724 / Kaggle AUC 0.8892
• Submission 37: Tuned LGBM (simple structure) → Kaggle AUC 0.8932
• Submission 38: Switched to RandomForest → noticeable drop (AUC 0.8805)
• Submission 39: LogisticRegression baseline → consistent but limited (AUC 0.8736)
• Submission 40: LGBM + RandomSearchCV → robust tuning (AUC 0.8915)

Takeaways:

• Simple LGBM still outperforms both ensemble trees and linear models in this task
• Feature structure is stable; now tuning is the key lever
• RandomSearchCV with constraints yields solid results without overfitting

🧠 Computer Vision Self-Study — GAN Playground

What I Did:

• Played with a basic GAN model using PyTorch
• Generated simple synthetic digits (MNIST-style)
• Adjusted generator/discriminator learning rates and batch sizes

Insights:

• Watching generator loss collapse is oddly satisfying
• Even basic GANs teach a lot about optimization dynamics
• Planning to explore conditional GANs next

📄 Paper Review Kickoff — MobileNetV2

Started: MobileNetV2: Inverted Residuals and Linear Bottlenecks
Today’s Focus: Day 1 – Abstract, Introduction, Problem Statement
Repo: Paper Review – GitHub

Reflection:

• Really enjoyed this first deep-dive
• Realized the importance of mobile constraints in architecture design
• Looking forward to dissecting the inverted residual block tomorrow

🎯 Next Steps

• CV: Try conditional GAN with labels
• SCU: Add SHAP filtering to next submission
• Paper Review: Sketch MobileNetV2 architecture + start Section 3 analysis
• Blog: Consider auto-generated paper index page by category/tag

✅ TL;DR

📍 SCU: Simple LGBM wins again — balance is better than brute force
📍 CV: GANs are super fun and super sensitive to tuning
📍 Paper: Started MobileNetV2 review, great kickoff to the reading habit