Case Study

Architecting a self-driving vehicle with a restricted dataset.

Case context: You are leading a project to build an autonomous vehicle, but your dataset of driving footage and corresponding controls is very small. You are deciding between two machine learning approaches: a purely end-to-end model that maps inputs directly to driving controls, or a pipeline model that first detects cars, then detects pedestrians, and finally plans a path.

Question: Based on the concept of function complexity and data requirements, which design path should you choose to ensure the system can be trained effectively with your limited data? Explain your decision.

Sample answer: You should choose the pipeline approach. An autonomous driving pipeline encodes three simpler subtasks: detecting other cars, detecting pedestrians, and planning a path. Because each of these steps is a relatively simpler function compared to mapping raw input to final path/controls directly, they can be learned with less training data than a purely end-to-end approach, making it the viable option for a restricted dataset.

Key points:

  • Choose the pipeline approach over the end-to-end approach.
  • The pipeline encodes three specific steps: detect cars, detect pedestrians, and plan a path.
  • Each step in the pipeline represents a simpler function.
  • Simpler functions require less data to learn, making it suitable for limited datasets.

Rubric: The answer must choose the pipeline approach and justify it by explaining that the three steps (detecting cars, detecting pedestrians, and planning a path) are simpler functions, which allows them to be learned with less data than a purely end-to-end approach.

0

1

Updated 2026-05-26

Contributors are:

Who are from:

Tags

Machine Learning

Deep Learning

Supervised Learning

Dive into Deep Learning @ D2L

Data Science

Machine Learning Strategy

Related