A summary of MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications

Andrew G. Howard et al. arXiv, 2017 DOI

For the summary of the paper, go to the Summary section of this article.

Table of Contents

• A summary of MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications
  • Table of Contents
  • First Pass
    • Problem
    • Motivation
    • Category
    • Context
    • Contributions
  • Second Pass
    • Background Work
    • Figures, Diagrams, Illustrations, and Graphs
    • Clarity
    • Relevant Work
    • Methodology
    • Author Assumptions
      • Correctness
    • Future Directions
    • Open Questions
    • Author Feedback
  • Summary
  • Summarization Technique
  • Citations

First Pass

Read the title, abstract, introduction, section and sub-section headings, and conclusion

Problem

What is the problem addressed in the paper?

More and more CV models are getting better performance by “going deeper”. However, these models can not be run efficiently on mobile and low powered devices. Therefore, a new class of models (MobileNets) needs to be developed in order to accommodate for these low powered and resource constrained devices.

Motivation

Why should we care about this paper?

We should care about this paper as it introduces a new class of CV models targeted at mobile and low powered devices (MobileNets) as well as a method for creating these models to meet specific latency and size specifications through the usage of two hyper parameters.

Category

What type of paper is this work?

This paper is a deep learning CV paper.

Context

What other types of papers is the work related to?

This work is most similar to papers focusing on CV model architectures and low powered computer vision.

Contributions

What are the author’s main contributions?

The author’s main contributions are a new class of models for low powered devices (MobileNets) and a methodology to generate these models targeting specific sizes and latency requirements through the usage of two hyper parameters.

Second Pass

A proper read through of the paper is required to answer this

Background Work

What has been done prior to this paper?

There has been work in developing small and efficient CV models previously. Additionally, work has been done to develop depth wise separable convolutions which are the backbone of this model class. Furthermore, work has been done in the area of model compression and optimization.

Figures, Diagrams, Illustrations, and Graphs

Are the axes properly labeled? Are results shown with error bars, so that conclusions are statistically significant?

All of the figures are clear. However, not all of the charts are made clear on their own. some of the charts use an unlabeled log scale that is described in the captions of the figure.

Clarity

Is the paper well written?

This paper is well written.

Relevant Work

Mark relevant work for review

The following relevant work can be found in the Citations section of this article.

2. F. N. Iandola, M. W. Moskewicz, K. Ashraf, S. Han, W. J. Dally, and K. Keutzer. Squeezenet: Alexnet-level accuracy with 50x fewer parameters and¡ 1mb model size. arXiv preprint arXiv:1602.07360, 2016.
3. J. Jin, A. Dundar, and E. Culurciello. Flattened convolutional neural networks for feed forward acceleration. arXiv preprint arXiv:1412.5474, 2014.
4. M. Rastegari, V. Ordonez, J. Redmon, and A. Farhadi. Xnornet: Imagenet classification using binary convolutional neural networks. arXiv preprint arXiv:1603.05279, 2016.
5. M. Wang, B. Liu, and H. Foroosh. Factorized convolutional neural networks. arXiv preprint arXiv:1608.04337, 2016.
6. J. Wu, C. Leng, Y. Wang, Q. Hu, and J. Cheng. Quantized convolutional neural networks for mobile devices. arXiv preprint arXiv:1512.06473, 2015.
7. L. Sifre. Rigid-motion scattering for image classification. PhD thesis, Ph. D. thesis, 2014.

Methodology

What methodology did the author’s use to validate their contributions?

The authors trained and tested variations of their MobileNets on tasks such as fine grained recognition, large scale geolocation, face attributes, object detection, and face embeddings. They then evaluated the performance of their models against VGG and Inception V2 and V3. Comparisons were made by measuring the number of billion multiply-adds and million parameters, as well as the mean average precision.

Author Assumptions

What assumptions does the author(s) make? Are they justified assumptions?

The largest assumption that I see is that MobileNets would be applicable for mobile and low powered hardware. The reason I say this is that the paper doesn’t describe a test of running this model on a mobile device.

Correctness

Do the assumptions seem valid?

No they do not. The name seems misleading without a proper test.

Future Directions

My own proposed future directions for the work

I’d like to test variations of MobileNets on modern smartphones, low powered devices (i.e., Raspberry Pis, NVIDIA Jetson Nanos), and more powerful devices (i.e., laptops) in order to see how far I can push a MobileNet to give real-time (minimum 24 FPS) inference without compromising on accuracy.

Open Questions

What open questions do I have about the work?

Why wasn’t a MobileNet variation tested on a mobile device?

Author Feedback

What feedback would I give to the authors?

Overall this was a good paper. I found the description of the depth wise separable convolutions a little confusing. Additionally, the naming of the family models presented (MobileNets) is misleading to me as these models were not tested on mobile devices.

Summary

A summary of the paper

The paper MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications by Andrew G. Howard et al. introduces a new class of models that are designed to be small and fast (low latency) called MobileNets. MobileNets are CNN based DL models that rely on depth wise separable convulsions to inference data. A depth wise separable convolution occurs in two steps:

1. Each channel of the input data structure is analyzed using a 3 x 3 x 1 filter. This is different than traditional CNN building blocks that analyze all of the channels in one pass using a 3 x 3 x N filter (with N being the number of filters).
2. The values from the aforementioned filters are multiples together and summed with a point wise convolution that analyzes all of the channels. This point wise convolution uses a 1 x 1 x N filter (with N being the number of filters).

By using depth wise separable filters, MobileNets decrease their latency due to the fewer number of multiply-add operations that they need to perform on the data.

The authors tested the performance of MobileNets on a variety of tasks against the VGG and Inception class of CV models, and found that MobileNets are both smaller, faster, and nearly as accurate as these much larger models.

Summarization Technique

This paper was summarized using a modified technique proposed by S. Keshav in his work How to Read a Paper [0].

Citations