Google

For autonomous driving, machine learning can be used to detect cars and pedestrians, and labeled car or pedestrian image data is relatively easy to obtain. By contrast, a pure end-to-end approach would require a large dataset of image and steering-direction pairs that is time-consuming and expensive to collect, making it difficult to train. This data availability supports considering intermediate car and pedestrian detectors in a multi-stage, non-end-to-end pipeline.

Autonomous Driving Data Availability Favors Intermediate Detectors

Machine Learning Yearning, a free ebook from Andrew Ng, teaches you how to structure Machine Learning projects.
https://www.deeplearning.ai/machine-learning-yearning/

Note: The content of the book is aligned with the Coursera Deeplearning.ai specialization.  https://www.deeplearning.ai/deep-learning-specialization/ 

Machine Learning Yearning (Deeplearning.ai)

Why does data availability favor intermediate detectors over a pure end-to-end approach for autonomous driving?

True or False: Collecting (Image, Steering Direction) pairs for end-to-end autonomous driving is just as easy as collecting labeled car and pedestrian images.

For a pure end-to-end autonomous driving system, each training example must be an (Image, _____) pair.

Why is it easier to collect training data for a car or pedestrian detector than for a pure end-to-end autonomous driving system?

A pure end-to-end autonomous driving system is easy to train because (Image, Steering Direction) pairs are inexpensive to collect.

To train a pure end-to-end autonomous driving system, you need a large dataset of _____ pairs.

Match each pipeline-related concept to its data characteristic in the autonomous driving context from Machine Learning Yearning.

Order the reasoning steps for deciding whether to use a multi-stage pipeline over a pure end-to-end approach in autonomous driving.

According to Machine Learning Yearning, which outcome does the scarcity of (Image, Steering Direction) pairs most directly support?

Machine Learning Yearning states that it is difficult to find labeled car and pedestrian images for training intermediate detectors.

Machine Learning Yearning advises that if there is a lot of data for training _____ of a pipeline, you might consider using a pipeline with multiple stages.

Match each data-related concept to its pipeline design implication in Machine Learning Yearning's autonomous driving example.

Order the steps that build Machine Learning Yearning's data-availability argument for preferring a multi-stage pipeline in autonomous driving.

Question:
Compare the data collection challenges and costs of a pure end-to-end autonomous driving system versus a multi-stage pipeline that uses intermediate detectors. Explain how data availability shapes the decision to use one system design over the other.

Sample answer:
A pure end-to-end autonomous driving system requires a large dataset of image and steering-direction pairs. This data is time-consuming and expensive to collect because it requires people to drive cars around and record their steering direction, making the system difficult to train. In contrast, intermediate modules like car and pedestrian detectors can use existing computer vision datasets that contain large numbers of labeled cars and pedestrians, which are relatively easy and inexpensive to obtain. Since there is high data availability for these intermediate detectors but scarce/expensive data for end-to-end steering mapping, developers are favored to choose a multi-stage, non-end-to-end pipeline over a pure end-to-end system.

Key points:
- End-to-end systems require (Image, Steering Direction) pairs which are expensive and time-consuming to collect via physical driving.
- Intermediate detectors (car/pedestrian) can leverage abundant, easily obtainable labeled image datasets.
- Abundant data for intermediate modules favors using a multi-stage, non-end-to-end pipeline.

Rubric:
The answer must compare the collection costs of image/steering-direction pairs with labeled car/pedestrian images, and explain how this difference favors intermediate modules in a multi-stage pipeline.

Analyze how data collection costs influence autonomous driving pipeline design choices

Case context:
A new autonomous driving startup has a limited budget and timeline. They are choosing between building a pure end-to-end steering system and a multi-stage pipeline that utilizes intermediate car and pedestrian detectors. They notice that public computer vision datasets contain a massive number of labeled car and pedestrian images, but they do not have pre-collected datasets of driver steering logs mapped to front-facing camera footage.

Question:
Based on the concept of data availability for autonomous driving, which architectural approach should the startup select and why? Detail the specific data constraints that influence this decision.

Sample answer:
The startup should select a multi-stage, non-end-to-end pipeline that utilizes intermediate car and pedestrian detectors. A pure end-to-end approach requires a large dataset of image and steering-direction pairs, which is expensive and time-consuming to collect since it requires recording manual driving. Because labeled car and pedestrian image data is already relatively easy to obtain from existing computer vision datasets, building intermediate detectors is much more feasible under their budget and timeline constraint. Therefore, the availability of intermediate data favors a multi-stage pipeline design.

Key points:
- Recommend a multi-stage, non-end-to-end pipeline with intermediate detectors.
- Contrast the high cost/time of collecting image and steering-direction pairs against the ease of obtaining labeled car and pedestrian images.
- Explain that high availability of data for intermediate modules favors a multi-stage approach.

Rubric:
The student should identify the multi-stage pipeline (with intermediate detectors) as the better choice, noting that image-steering pairs are expensive/time-consuming to collect whereas labeled car/pedestrian images are easy to obtain.

Evaluating pipeline architecture for a new self-driving car startup

Question:
What specific data collection constraint makes a pure end-to-end autonomous driving system difficult to train compared to intermediate detectors?

Sample answer:
A pure end-to-end autonomous driving system is difficult to train because it requires a large dataset of image and steering-direction pairs, which is time-consuming and expensive to collect. In contrast, training data for intermediate detectors (such as labeled car and pedestrian images) is relatively easy to obtain.

Key points:
- End-to-end training requires image and steering-direction pairs which are expensive and time-consuming to collect.
- Intermediate detectors utilize labeled car and pedestrian images which are relatively easy to obtain.

Rubric:
The answer must identify that collecting image and steering-direction pairs for the end-to-end system is expensive and time-consuming, unlike the relatively easy-to-obtain labeled car/pedestrian data for intermediate modules.

Learn Before

Related