🧠 Daily Study Log [2025-07-02]
Today’s study was split across three major areas: feature engineering experiments for the SCU_Competition, continuing the MobileNetV2 paper review, and reading about DeepDream and Neural Style Transfer in computer vision.

📊 SCU_Competition — Submission 41–45

Focus: Advanced feature engineering + LGBM tuning
Goal: Push generalization without overfitting

Highlights:

• Submission 41: Added log1p and ratio-based features → solid generalization
• Submission 42: Included price-income gap and wine × campaign mix → minor lift
• Submission 43: Tested early-stage interaction features
• Submission 44: Combined all advanced features → Kaggle AUC 0.8944
• Submission 45: Full stack of features (original + log1p + ratio + diff + interactions) →
  CV AUC: 0.8835 / Kaggle AUC: 0.8902

Takeaways:

• Layering multiple feature types can improve local CV, but diminishing returns appear in public scores
• Slight drop in Kaggle AUC in Submission 45 suggests model may be hitting a generalization ceiling
• Time to prune weaker features or apply SHAP for model-aware selection

📄 Paper Review — MobileNetV2 (Day 2)

Continued: MobileNetV2: Inverted Residuals and Linear Bottlenecks
Today’s Focus: Section 2 – Architecture design and motivation
Repo: GitHub Paper Review

Reflection:

• Depthwise separable convolutions are incredibly efficient for mobile environments
• Inverted residuals + linear bottlenecks minimize memory and computation without hurting accuracy
• Eager to visualize the full architecture block in the next session

🧠 CV Self-Study — DeepDream & Neural Style Transfer

What I Studied:

• DeepDream: Used gradient ascent on feature maps to visualize what CNNs “see”
• Neural Style Transfer (NST): Combines content from one image and style from another using VGG features

Insights:

• DeepDream is more about interpretability than generation
• NST illustrates how CNNs separate low-level style vs high-level structure
• Planning to implement NST soon with personal style + content images

🎯 Next Steps

• SCU: Try SHAP-based pruning on current feature set
• CV: Implement NST with VGG and log results
• Paper: Summarize Section 3 and diagram MobileNetV2 blocks
• Blog: Consider building an auto-generated paper index by tag

✅ TL;DR

📍 SCU: Feature stacking helps, but pruning might now be more effective
📍 CV: DeepDream and NST offer powerful ways to “see” what models learn
📍 Paper: MobileNetV2 continues to impress with its elegant, efficient design