Synthetic Data Can Still Cause Narrow Memorization
A vision model trained on generated bicycle images may look successful during training but still fail on real street scenes. If 80,000 examples come from only about 15 bicycle models, the model can learn those limited designs too well and then struggle with new frames, backgrounds, and lighting conditions.
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
Overfitting from Reused Background Noise
Synthetic Data Can Still Cause Narrow Memorization
What is the main concern when synthetic examples are easy to tell apart from real data?
Representativeness of Synthetic Noise Samples
Synthetic Data Should Reflect the Real Task
Match each synthetic-data case to whether it is representative.
Order the steps for checking whether generated examples are truly representative.
Why synthetic training examples must reflect real variation
Check a Synthesized Audio Data Plan for Hidden Artifacts
How to check whether synthetic training data looks unrepresentative
Which synthetic-data strategy is least likely to create a representativeness problem?
True or False: Synthetic data built from only a small set of source examples can be unrepresentative.
Overfitting from Reused Background Noise
Synthetic Data Can Still Cause Narrow Memorization
Learn After
Why might a detector trained on 120,000 synthetic storefront images built from only 25 shop layouts fail on real streets?
True or False: If a generator makes highly realistic-looking synthetic storefront photos, a classifier will generalize well once enough of those images are produced.
If 80,000 training photos all show only a dozen apartment floor plans, the model may _____.
Match each property of a synthetic road-sign dataset to its description.
Order the reasoning steps showing why a large synthetic robot dataset still causes overfitting.
Why a huge synthetic dataset can still be too narrow
Why a Large Synthetic Bicycle Dataset Failed on Real-World Photos
How much of the real-world distribution do a handful of synthetic bicycle designs cover?
Which change would most directly lower overfitting risk in the simulated bridge-inspection dataset?
True or False: This synthetic-data overfitting risk is limited to one category of object and does not apply elsewhere.