Learn Before
Making a development set match the real deployment mix
Case context: A team is building a model that will be deployed across five retail warehouses. For convenience, they consider creating the development set from data taken from only two warehouses.
Question: If the development set includes samples from only two warehouses, what will the team end up improving, and why? What change should they make so that their optimization matches the actual deployment goal?
Sample answer: If the development set is built from only two warehouses, the team will mainly optimize performance on those two locations, because model-building decisions naturally track the development set. That means progress on the other three warehouses may be ignored even though they matter in production. To make the optimization objective match the real task, the development set should be redesigned to include representative data from all five warehouses.
Key points:
- The model-development process follows the development set.
- A narrow development set narrows the optimization target.
- The missing warehouses would not be properly represented.
- The development set should reflect the full deployment environment.
Rubric: The response should state that optimization will be concentrated on the two warehouses included in the development set, and it should recommend revising the development set so that it includes representative data from all five warehouses.
0
1
Tags
Machine Learning
Deep Learning
Machine Learning Strategy
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Yearning @ DeepLearning.AI
Related
Why should the development set track the task you care most about improving?
After the development and test sets are defined, the team usually starts tuning the model to improve development set performance.
Choose the word that completes the statement about dev-set design.
Match each machine learning set or idea to its purpose.
Put the evaluation-set planning steps in the correct order.
A team needs good results in four regions, but its development set samples only two of them. What is the most likely consequence?
If a validation set leaves out important real-world conditions, the team may optimize the wrong problem.
After the data splits are chosen, the team will mainly optimize for _____ performance.
How dev-set coverage changes the direction of optimization
Order the outcomes when a development set leaves out part of the real task.
Why Does the Development Set Shape What the Team Improves?
Making a development set match the real deployment mix
What should a development set be designed to match?