Advanced level · Module 5: Responsible Deep Learning · Lesson 1 of 4
Bias in Vision and Language Models
Recognise how training data shapes what deep models do well and badly, and measure it.
In this lesson you will
- Identify sources of bias in image and text data
- Measure performance across conditions, groups and languages
Key concepts
representation in training dataskin tone and lighting in imageslanguage coveragedisaggregated evaluation
The lesson
Large image and text data sets over-represent some places, people and languages. Models trained on them often perform worse for under-represented groups — darker skin tones in poor lighting, African languages, rural scenes.
Measure performance separately across conditions and groups relevant to the application: lighting, device type, region, language.
Gaps call for targeted data collection, evaluation on local data and honest statements of limits.
Step by step
- List groups and conditions relevant to each provided model.
- Evaluate the models separately for each.
- Identify the largest gaps.
- Propose data collection to address them.
Practice activity
Complete the disaggregated audit of the two provided models.
Expected result: An audit with per-group results and targeted recommendations.
Check your understanding
-
Why do many models perform worse on African languages?
Show answer
Those languages are under-represented in training data.
-
What is disaggregated evaluation?
Show answer
Measuring performance separately for groups and conditions.
Taught live by an instructor — online or in person. Free course lessons are for self-study and do not lead to a certificate — see registration options.