Case Study

Deciding the Architecture for a New Self-Driving Startup

Case context: A new autonomous vehicle startup is deciding how to build its perception and control stack. One team proposes a single deep neural network that takes in camera feeds and directly outputs steering angles. Another team proposes a modular pipeline that first detects cars and pedestrians, and then plans the path. They currently have limited full end-to-end driving logs but a lot of isolated bounding box data.

Question: Based on Machine Learning Yearning, which approach should the startup choose and why?

Sample answer: The startup should choose the modular, non-end-to-end approach. According to Machine Learning Yearning, an end-to-end approach is less promising for autonomous driving until more end-to-end data becomes available. Because they have limited end-to-end driving logs but abundant isolated data, a non-end-to-end architecture better matches their current data availability.

Key points:

  • Recommend the modular/non-end-to-end approach.
  • Identify that end-to-end data is currently limited.
  • State that non-end-to-end architecture better matches the available data.

Rubric: The student must recommend the non-end-to-end (modular) approach and justify it by stating that its architecture better matches the availability of data.

0

1

Updated 2026-05-27

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Machine Learning Yearning @ DeepLearning.AI