Computing Bias
Introduced| English |
|---|
| Computing bias/kəmˈpjuːtɪŋ ˈbaɪəs/ |
| training data/ˈtreɪnɪŋ ˈdeɪtə/ |
| diverse/daɪˈvɜːs/ |
Unfairness built into a program
- Computing bias 计算偏见 is unfairness built into a program — sometimes without developers even realising it.
- A program can treat some groups worse than others.
- This is not always deliberate; it can be an accident of how the program was built.
- Bias matters because software increasingly makes real decisions about people.
Computing bias is:
It can treat some groups worse, even by accident.
Bias comes from the data
- A major source is the data used to build the program.
- Many modern programs learn from training data 训练数据 — examples they study to find patterns.
- If that data is biased, the program's results are biased too.
- Biased training data leads to unfair results for the under-represented group.
How bias enters a program
Biased training data leads a program to learn skewed patterns and produce unfair results; diverse testing reveals it, and representative data reduces it.
Programs that learn from examples study their ______ data to find patterns.
Biased training data gives biased results.
A major source of computing bias is:
If the data is skewed, the program's results are skewed.
Testing with diverse users
- Testing with diverse 多样的 users and data helps reveal hidden bias before release.
- If everyone who tests looks the same, the bias stays invisible.
- A varied test set exposes where the program fails different groups.
- This is why diverse testing is a core fix, not an optional extra.
Testing with a diverse set of users helps reveal hidden bias before release.
A varied test set exposes failures on different groups.
A face system trained mostly on light-skinned men makes more errors on others because:
The under-represented groups get worse results.
Developers can reduce bias by:
Representative data and cross-group checks make results fairer.
Reducing bias
- Developers can reduce bias by using representative data and checking results across different groups.
- The goal is a program that works fairly for everyone, not just the majority.
Face recognition. A system trained mostly on light-skinned men works well on them but makes far more mistakes on women and darker-skinned people — it saw fewer such faces while learning. The developers did not intend this; the bias came from unrepresentative training data. Testing with a diverse set of faces would have exposed it.
Computing bias is unfairness built into a program, often unintentionally. A major source is biased training data — if the examples are skewed, so are the results. Testing with diverse users reveals hidden bias before release, and using representative data and checking across groups reduces it.