📌 Paper Info

• Title: MobileNetV2: Inverted Residuals and Linear Bottlenecks
• Authors: Mark Sandler, Andrew Howard, Menglong Zhu, Andrey Zhmoginov, Liang-Chieh Chen
• Link: arXiv 1801.04381
• Published: 2018 (Google Research)
• Code: Available in TensorFlow-Slim

🎯 Why I Chose This Paper

This paper is a great starting point for anyone interested in efficient neural network architectures.
It introduces a simple yet powerful design that has become a foundation for many mobile-friendly models.
Since I’m building a habit of reviewing papers and implementing key ideas, this was a perfect entry point.

🧠 Day 1 Review – Overview & Motivation

✅ Abstract Summary

MobileNetV2 proposes a new lightweight CNN architecture optimized for mobile and resource-constrained environments.
It introduces Inverted Residuals and Linear Bottlenecks to reduce computation while preserving expressiveness.
It performs well on classification, object detection (SSDLite), and segmentation tasks with a favorable trade-off between accuracy and efficiency.

✅ Introduction Takeaways

• Modern deep networks perform well, but are too heavy for mobile inference.
• MobileNetV2 solves this by expanding low-dimensional inputs → filtering with depthwise convolution → projecting linearly to low dimensions.
• This avoids large intermediate tensors, which reduces memory usage during inference.

✅ Problem Statement

Modern CNNs achieve high accuracy but are unsuitable for mobile deployment due to computational and memory demands.
This paper addresses that challenge by introducing a module that preserves performance while dramatically reducing resource consumption.
The core idea is to use lightweight depthwise convolutions and avoid non-linearities in narrow layers, which reduces memory footprint.

🗂 GitHub Repository

You can find my detailed review and notes here:
🔗 github.com/hojjang98/Paper-Review

💭 Reflections

This was my first formal paper review and it was a very insightful experience.
I learned that understanding the “motivation” behind a design is as important as understanding the design itself.
Looking forward to digging deeper into the architecture and experiments in the next phase.