Why More Data Alone Usually Does Not Fix Persistent Underfitting
When a learning curve shows that training error remains above the target because the model underfits, collecting more labeled examples without changing the model or training procedure usually does not solve the problem. Training error may stay similar or rise as the training set grows, and development error is typically no lower than training error. The model must first fit the training data better before additional data can be expected to achieve the target development error.
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
Recognizing High Avoidable Bias from Learning Curves
A Small Train-Dev Gap Can Coexist with High Bias
If training error is still far above the best achievable error, the model has _____.
Interpreting Training and Validation Gaps
Steps for identifying strong underfitting from a learning curve
Recognizing High Bias on a Learning Curve
Diagnosing a Fraud Screening Model's Learning Curve
Meaning of a Small Training-Dev Error Gap
What Does the Reference Performance Represent?
High-Bias Models Usually Fit the Training Set Well
Why More Data Alone Usually Does Not Fix Persistent Underfitting
Learn After
Why More Data May Not Fix a High-Training-Error Model
Adding more training examples by itself can guarantee that dev error will reach a target level even when training error is still above that target.
When the training set gets larger, the training error can only stay the same or _____.
Match each learning-curve element to its role in diagnosing when extra data will not solve the problem.
Order the logic that shows more data by itself will not achieve the target performance.
Why more data alone cannot solve high training error
Decide whether more data alone can close a large training gap.
Why can’t adding more data by itself solve this learning-curve problem?
Which learning-curve pattern most clearly shows that collecting more data by itself will not solve the problem?
If training error does not improve as more labeled examples are added, development error can still be expected to drop below a lower target level from extra data alone.