AIF-C01 ยท Responsible and explainable AI
Datasets, bias and variance
Inclusive, diverse, balanced and curated datasets, and the effects of bias and variance such as underfitting, overfitting and harm to demographic groups.
Key points
- Good training data is inclusive and diverse, balanced across groups and classes, and curated from trusted sources.
- Bias (model too simple) causes underfitting: poor results even on training data.
- Variance (model too sensitive) causes overfitting: great on training data, poor on new data.
- Unrepresentative data can also make a model perform worse for some demographic groups.
Practice questions
A model scores 99% on its training data but only 70% on unseen data. What is the most likely problem?
- Underfitting
- Overfitting
- Class balance
- Label encryption
Show answer
Answer: B. Overfitting
A large gap between training and new-data performance indicates overfitting (high variance). Underfitting would show poor scores on both.
Which characteristics describe a dataset suited to responsible AI?
- Diverse and inclusive of the people it will affect
- Collected from a single convenient source only
- Balanced across relevant groups and classes
- Curated from vetted sources with quality checks
- Scraped in bulk with no review
Show answer
Answer: A. Diverse and inclusive of the people it will affect, C. Balanced across relevant groups and classes, D. Curated from vetted sources with quality checks
Diverse, balanced and curated data reduce harm to groups. Scraping unchecked data or using one convenient source bakes in gaps and errors.
True or false: high overall accuracy proves a model is fair to every demographic group.
Show answer
Answer: False
Overall accuracy can hide poor performance on a small subgroup; fairness needs per-group evaluation.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
