Why Human Approval Does Not Prove Synthetic Data Works for a Model
Question: Why is it not enough for a person to say that generated transaction records look realistic in order to conclude that they are realistic for a machine learning model?
Sample answer: A person may judge realism by surface appearance, while a model depends on statistical details such as distributions, relationships between variables, and rare-value patterns. Because synthetic data can seem plausible to a person yet still fail to match the patterns a model needs, human approval alone is not sufficient.
Key points:
- People and models evaluate realism using different criteria.
- Data can look plausible to a person but still be unsuitable for a model.
Rubric: The response must explain that human judgment and model judgment differ, and that data can appear realistic to a person without having the statistical properties needed for machine learning.
0
1
Tags
Machine Learning
Deep Learning
Supervised Learning
Dive into Deep Learning @ D2L
Data Science
Machine Learning Strategy
Machine Learning Yearning @ DeepLearning.AI
Related
What makes synthetic data difficult to use effectively?
Human-likeness guarantees machine-likeness in synthetic data
Synthetic Data Can Look Real to a Human First
Match each synthetic-data concept to the description it best fits in the realism dilemma.
Put the checks in a sensible order when deciding whether generated data is usable for training.
A team generates synthetic café-noise clips that human listeners say sound convincing. What should they check before using them for training?
Synthetic examples that seem believable to people are often easier to make than examples that match a model's learned patterns.
A generated image can look ____ to a person while still triggering a detector that says it was synthesized.
Match Each Scenario to the Correct Realism Concept
Order the steps for checking whether synthetic customer-support chats are useful for training.
Why Synthetic Samples Can Fool People but Not Models
Testing Whether Synthetic Data Works for a Model
Why Human Approval Does Not Prove Synthetic Data Works for a Model