Learn Before
Training Performance Is a Baseline for Held-Out Performance
A model’s accuracy on the training set is usually at least as good as its accuracy on dev or test data. If it reaches only 82% on data it has already seen, it is not reasonable to expect 94% on unseen examples.
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
High Training Error Calls for Better Fit, Not More Data
Low Bias and Low Variance Mean Good Performance
A Negative Gap to the Best Achievable Error Usually Signals Memorization
Informal and Technical Uses of Bias and Variance
Training Performance Is a Baseline for Held-Out Performance
Informal Meaning of Model Bias
True or False: In this framework, the term bias refers to a model’s error on the validation or test set.
Informal meaning of bias in a large-sample setting
Informal Meaning of Bias
In this informal usage, bias means the model's error on the training set.
Informally, an algorithm's _____ is the error rate on the training set.
Match each learning-curve term to its meaning.
Order the steps for estimating bias from training error.
Why use a very large training set when estimating bias?
Bias Is Measured on the Holdout Set
In the large-sample limit, bias is typically assessed on the _____ data.
Match each learning-term description with its role in the informal bias definition.
Order the steps for deciding whether bias is the main cause of poor model performance.
How Informal Bias Relates to Training Error
Interpreting Training Error on a Large Dataset
Informal Bias and Training Error
Learn After
What usually has to happen before a model is expected to do well on validation and test data?
True or False: Performance on development and test data is usually lower than performance on the training set.
Training performance comes first
Interpreting Training and Held-Out Performance
Order the steps used to judge whether a better score on unseen data is believable.
Explain why training performance must come before strong validation or test performance.
Diagnose a claim of 72% training accuracy and 89% dev/test accuracy.
Why must training performance be solid before you expect strong validation or test performance?
Which reported result contradicts the training-first rule?
A model that fits the training set poorly can still be expected to do well on dev or test data.