🧠 Daily Study Log [2025-07-09]
Feeling better today and finally picked up some solid momentum.
Made five new submissions for the SCU_Competition, reviewed SQL normalization theory, cleaned up my CV code for Chapter 2, and completed Day 1 of the EfficientNet paper study.

📊 SCU_Competition — Submission Marathon

Focus: Rebalanced ensemble weights and tested feature interaction sets

• ✅ Submitted 5 new experiments — mostly variations of VotingClassifier weights (e.g., 6:2:2, 7:1.5:1.5, 8:1:1)
• ✅ Best Kaggle AUC: 0.9002 — finally crossed the 0.9 milestone 🎉
• ✅ Used a combination of strong feature engineering and light clustering
• Also cleaned up noisy features and re-validated interaction terms

Reflection:

• VotingClassifier remains the most stable approach
• Overweighting a single model can hurt generalization
• The key is not complexity, but well-structured feature design

🧠 SQL Study — Normalization & Denormalization

Topics Covered:

• Reviewed 1NF, 2NF, 3NF with examples
• Compared normalization vs. denormalization: trade-offs and use cases
• Discussed how denormalization may help query performance in practice

Reflection:

• The concepts were familiar but felt more grounded this time
• Realized that database design is not about correctness alone — it’s about context and intent

📘 Paper Study — EfficientNet (Day 1 Completed)

Focus:

• Finished reading and summarizing Abstract, Introduction, and Motivation
• Deep-dived into the concept of compound scaling
• Cleaned up the LaTeX-style math to be GitHub Pages compatible using $$...$$ syntax

Reflection:

• Having already read MobileNetV2, this paper felt like a natural next step
• Looking forward to Day 2: exploring architecture details and scaling coefficients

👁️ CV Study — DL4VS Chapter 2 (Backprop, Basic Layers)

Tasks Completed:

• Read through Chapter 2: activation functions, FC/Conv layers
• Implemented manual backpropagation and verified gradients
• Wrote minimal code blocks to better understand forward/backward passes

Reflection:

• Much more practical than Chapter 1
• Feels like I’m building foundational blocks for later experiments

🔍 Tomorrow’s Plan

• ✅ Push SCU_Competition toward submission #80
• ✅ EfficientNet Day 2: architectural breakdown and B0–B7 scaling
• ✅ Finish Chapter 2 code and launch initial experiment
• ✅ Explore real-world SQL schema patterns (JOIN optimization, denormalized designs)

✅ TL;DR

📍 SCU: 5 new submissions — hit 0.9002 AUC for the first time!
📍 SQL: Reviewed normalization vs. denormalization from a practical angle
📍 Paper: EfficientNet Day 1 done — compound scaling makes sense now
📍 CV: Chapter 2 progress — implemented backprop manually
📍 Tomorrow: Another day of experiments, theory, and code