📌 Paper Info

• Title: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
• Authors: Mingxing Tan, Quoc V. Le
• Link: arXiv 1905.11946
• Published: ICML 2019 (Google AI)
• Code: Available in TensorFlow, PyTorch (via timm library)

🌱 Day 0 – Why EfficientNet?

🧩 Context

I recently finished reading the MobileNetV2 paper, which introduced the concept of inverted residual blocks and efficient model design for mobile devices. While reviewing it, I started wondering:

“What’s the next step beyond MobileNetV2? How do modern lightweight models evolve from there?”

That curiosity naturally led me to EfficientNet, which not only builds on MobileNetV2’s architectural ideas, but also introduces a novel model scaling strategy — something I hadn’t explored in depth yet.

🔍 Why I Chose This Paper

• ✅ It’s a logical follow-up to MobileNetV2
• ✅ It introduces the idea of compound scaling, which feels both elegant and practical
• ✅ It has had a major impact on model design in recent years (used in many modern libraries)
• ✅ I wanted to understand how NAS (Neural Architecture Search) can be integrated with scaling strategies

🔜 What’s Next?

In Day 1, I’ll explore the motivation behind EfficientNet, what problem it’s trying to solve, and how its compound scaling approach works.
This will be followed by a breakdown of the architecture and performance benchmarks in Day 2 and beyond.

🧠 Personal Note

I’m not feeling well today, so I decided to keep things light and just share why I picked this paper instead of diving into technical content right away.
Sometimes setting the stage properly helps make the reading experience more meaningful and focused.

📌 Next up: EfficientNet DAY 1 – Abstract, Introduction, and Motivation