Learn Before
Explain the rationale and challenges of approximating future dev/test data before a mobile application launch.
Question: Before launching a mobile application, actual user data is not yet available to form dev/test sets. Explain why it is still critical to attempt to approximate future user data, and discuss how this approximation compares to the actual data you expect to face in the future.
Sample answer: Approximating future user data is necessary because without user data, you cannot construct dev/test sets that reflect the target distribution you need to perform well on in the future. By approximating it (e.g., having friends take mobile-phone cat pictures), you create a proxy dataset. However, the limitation is that this approximated data might still fail to perfectly reflect the actual future user data distribution once the app launches, but it is better than not having a representative set.
Key points:
- Dev/test sets must reflect the target distribution the model needs to perform well on in the future.
- Before launch, actual user data is unavailable, requiring approximation.
- Approximation can be done using proxy methods, such as asking friends to take mobile-phone pictures.
- Approximated data is only an approximation and may not perfectly represent final user data distributions.
Rubric: The answer should explain the need to approximate target distributions pre-launch to build dev/test sets, and note that the approximated data may not perfectly represent actual future user data.
0
1
Tags
Machine Learning
Deep Learning
Machine Learning Strategy
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Related
Before a mobile app launches, how can you approximate future dev/test data when no user data is available?
Before a mobile app launches, it is impossible to approximate the future data distribution for dev/test sets.
Before launching a mobile app, you can approximate future user data by asking friends to take _____ pictures of cats and send them to you.
When a mobile app has not yet launched and no user data is available, what is the recommended approach for building dev/test sets?
Before a mobile app launches, it is impossible to construct any dev/test set because real user data does not yet exist.
When no user data exists before app launch, Andrew Ng recommends trying to _____ the future data distribution for dev/test sets.
Match each pre-launch data concept to its correct description.
Order the steps a team should follow to approximate pre-launch dev/test data for a mobile app.
Which specific example does Andrew Ng give in Machine Learning Yearning for approximating pre-launch mobile app data for a cat-detection use case?
Data collected from friends before a mobile app's launch perfectly represents the distribution of future user data.
Andrew Ng suggests asking your _____ to take mobile-phone pictures of cats as one way to approximate pre-launch user data.
Match each pre-launch data approximation term to its correct explanation.
Order the reasoning steps for deciding how to handle the absence of user data when constructing dev/test sets before a mobile app launches.
Explain the rationale and challenges of approximating future dev/test data before a mobile application launch.
Case Study: Designing a cat detection dev/test set before mobile application launch.
What is the purpose of asking friends to take photos when constructing a pre-launch dev/test set?