🧠 Daily Study Log [2025-07-14]
Another productive day — wrapped up key topics across SQL, CV, paper study, and Kaggle exploration.
Found a new competition to challenge myself with and improved CNN results on CIFAR-10!

🧪 SQL Study — DML / DCL / TCL

Focus:

• Studied Data Manipulation Language (DML), Data Control Language (DCL), and Transaction Control Language (TCL)
• Reviewed key commands like INSERT, COMMIT, ROLLBACK, GRANT, and more

Reflection:

• Clearer understanding of database workflows
• Useful for managing structured data in data pipelines and analytics tasks

🏁 Kaggle — New Competition Exploration

Focus:

• Found a promising new competition: 2nd WEAR Dataset Challenge
• Focused on wearable sensor data and human activity recognition

Reflection:

• Great opportunity to explore time-series + sensor modeling
• Planning to set up a baseline this week and iterate with feature extraction and LSTM/Transformer models

🧠 CV Study — Chapter 4: CIFAR-10 Classification

Topics Covered:

• Followed the CNN-based implementation for CIFAR-10
• Improved upon the book’s base model with additional techniques:
  • Data augmentation
  • Dropout, BatchNorm
  • L2 Regularization
  • Tuned learning rate (Adam with 1e-4)

Result:

• Final Test Accuracy: ~67.2% (book baseline was lower)
• Validation Accuracy: ~66.7%

Reflection:

• Strong improvement from baseline → real gain from regularization and augmentation
• Ready to try more complex architectures next (e.g., ResNet)

📘 Paper Study — EfficientNet (Day 4)

Focus:

• Continued reviewing Section 4.3 — Scaling Results & Table 2
• Began interpreting performance improvements across EfficientNet-B1 to B7
• Reviewed key ablation study findings and model comparisons

Reflection:

• The compound scaling strategy scales better than naive methods
• Planning to wrap up study within 2 days, then post final summary

✅ TL;DR

📍 SQL: Studied DML/DCL/TCL — important for DB & ETL workflows
📍 Kaggle: Found new HAR competition — preparing first baseline
📍 CV: Finished CIFAR-10 (Ch.4) — achieved ~67.2% test accuracy
📍 Paper: EfficientNet Day 4 — analyzing scaling results & wrap-up soon
📍 Tomorrow: Competition baseline, Paper Day 5, maybe some huggingface fun?