Deep Learning (DL) · Intermediate level

Intermediate level · Module 2: Modern CNN Architectures · Lesson 3 of 3

Efficient Architectures for Mobile

Use architectures designed for phones and low-power devices, and trade accuracy against size and speed.

In this lesson you will

  • Explain depthwise separable convolutions
  • Compare efficient and standard architectures on real constraints

Key concepts

depthwise separable convolutionMobileNetparameters and FLOPslatency

The lesson

MobileNet and similar architectures replace standard convolutions with depthwise separable ones, which use far fewer calculations and parameters with a modest loss of accuracy.

For field use on modest smartphones, model size, memory and speed often matter more than the last percent of accuracy. Measure inference time on realistic hardware.

Choose based on agreed targets — for example at least 90% accuracy, under 20 MB and under half a second per photo on a mid-range phone.

Step by step

  1. Fine-tune a MobileNet on the project data.
  2. Measure model size and CPU inference time for both models.
  3. Compare accuracy.
  4. Recommend a model against the targets.

Practice activity

Complete the architecture comparison from the practical activity.

Expected result: A measured comparison with a recommendation tied to the deployment targets.

Check your understanding

  1. Why are depthwise separable convolutions efficient?

    Show answer

    They split a convolution into cheaper steps, using far fewer calculations and parameters.

  2. What matters most for models used on modest phones?

    Show answer

    Size, memory use and speed, alongside acceptable accuracy.

Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.