🧠 Daily Study Log [2025-07-07]
Wasn’t feeling well today, so I kept things light — but still managed to move forward with SCU_Competition (submission 61–65), a short SQL theory session, and a first trial of Neural Style Transfer. Also previewed the EfficientNet paper for upcoming review.

📊 SCU_Competition — Submission 61–65

Focus: Voting weight adjustment, cluster feature ablation, simplified LGBM tests, and extensive feature addition
Goal: Push past Kaggle AUC 0.8973 and validate feature importance balance

Highlights:

• Submission 61: VotingClassifier (LGBM:RF:LR = 6:2:2)
  → CV AUC: 0.8801 / Kaggle AUC: 0.8977 🏆 New Best
• Submission 62: Single LGBM (with same features)
  → CV AUC: 0.8759 / Kaggle AUC: 0.8884
• Submission 63: Removed clustering features from 61
  → CV AUC: 0.8818 / Kaggle AUC: 0.8955
• Submission 64: Simple LGBM w/ clustering features
  → CV AUC: 0.8759 / Kaggle AUC: 0.8884
• Submission 65: Same Voting (6:2:2) + large feature expansion
  → CV ongoing / Kaggle pending

Takeaway:

• Voting (6:2:2) seems optimal so far
• Cluster features help but aren’t essential
• Feature engineering continues to drive gains

🧮 SQL Study — Basic Concepts Review

Focus: Refreshed fundamental database design terms

• Entity vs Attribute: Defined entities as real-world objects; attributes as their descriptive fields
• Primary Key / Foreign Key: Revisited their purpose in relational schemas
• Normalization: Briefly touched on avoiding redundancy (1NF, 2NF concept only)

Reflection:

• Very light session today — focused more on terminology than coding
• Helped reinforce database theory before diving into query practice later

🖼️ CV — Neural Style Transfer (1st Trial)

Task: Implement baseline Neural Style Transfer

• Style image: Starry Night
• Content image: A Cat
• Backbone: VGG19 (layer slicing for style/content losses)

Next Plan:

• Show results & accuracy tomorrow
• Evaluate how well the style was preserved with different weight settings

Reflection:

• Early attempts were noisy, but after tuning style_weight, results improved
• Learned how content/style balance dramatically affects texture transfer

📄 Paper Preview — EfficientNet

Paper: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks (ICML 2019)

• Core Idea: Rather than increasing depth/width/resolution arbitrarily, use a compound scaling method to balance all three
• Architecture: Based on MobileNetV2’s inverted bottlenecks
• Impact: Achieves higher accuracy with fewer parameters than previous models

Preview Summary:

• Today’s focus was just grasping the big picture: EfficientNet proposes a principled way to scale models
• Will review B0 architecture and scaling coefficients in detail next time

✅ TL;DR

📍 SCU: Submission 61 hit a new best (Kaggle AUC 0.8977) — Voting (6:2:2) proves solid
📍 SQL: Light review of entity, attribute, keys — basic DB schema theory
📍 CV: NST first trial done — images improving after tuning, results tomorrow
📍 Paper: EfficientNet concept preview — compound scaling = key innovation