0:0014:27

Overcoming Gender Bias in AI

Tahmida Mahmud 14:27 2021

AI learns our gender bias, and in cars and clinics that can be dangerous

Image recognition tags a woman state senator by her appearance and her male colleague as a business person, and health apps tell women with left-arm pain to see a doctor in a couple of days. Deep learning engineer Tahmida Mahmud of Nauto traces how skewed data turns social bias into machine decisions with real costs. If more data is not the cure, what should AI teams actually change?

Learn where gender bias in AI comes from and the practices that help teams reduce it

Watch on YouTube

In short

Deep learning engineer Tahmida Mahmud argues that gender bias in AI comes from training data that mirrors unequal social, historical and economic conditions, so more data alone does not fix it. She walks through examples from image recognition, translation, car safety, health apps, speech technology and hiring, then sets out practices for responsible AI: representative and updated data, accuracy measured per demographic group, and more women on the teams that build AI.

“The model is as good as the data it is trained on.”
“Bias may be an unavoidable fact of life. Let's not make it an unavoidable aspect of technologies.”

What you will learn

  • Recognize how skewed historical data and human labeling introduce gender bias into AI models.
  • Identify real examples of gender bias in image recognition, machine translation, speech technology, health apps and hiring algorithms.
  • Apply data practices that reduce bias: diverse training samples, regular data updates and accuracy measured separately for each demographic group.
  • Ask the right questions during data collection, such as who is collecting and labeling the data and what problem it should solve.
  • Raise bias concerns with your team early, before a product ships.

Key takeaways from the speaker

  • We need to feed meaningful data to our models which represents the level-playing field we want the world to be
  • Data scientists should review the context, limitations and validity of the data which includes exploring what is missing in the dataset and asking the right questions during data collection
  • We need to increase diversity in tech and have more allies for an enabled and equal world
  • AI companies need to lead by example by hiring more women in product design, data science, and engineering since diversity of thought leads to better problem-solving

Details

Level
Beginner
Format
Talk
Technical depth
Introductory technical
For
All career stages
Length
14:27
Published
June 2021

Questions this talk answers

Why isn't more data enough to remove bias from AI?

Data reflects unequal social, historical and economic conditions, so more data is not better when it does not represent the bigger picture. The model is only as good as the data it is trained on.

What causes gender bias in AI systems?

Training data skewed by real-world discrimination (the gender data gap), human labeling that adds another layer of social bias, numbers taken without context, and a lack of diversity in the teams building AI.

Where does gender bias in AI cause real harm?

The speaker cites image tagging, translation, car safety design based on male crash dummies, health apps that misread women's heart attack symptoms, less accurate speech technology for higher voices, and algorithms that screen out women for jobs and loans.

How should you bring up AI bias concerns with your team?

Share your perspective early, because many biases are unintentional and colleagues may simply never have considered them. Fixing problems early avoids adverse impact once the product is finished.

Should ethics and governance practices be set up before building AI models?

Yes, it is of the highest importance. Even a model that is wrong in only 5% of cases needs attention if those errors follow a pattern by gender, race, ethnicity or age.

About this talk

Artificial Intelligence refers to a set of concepts, tools, and techniques that allows machines to simulate human intelligence and mimic human-like tasks through analyzing data and identifying patterns, and insights. Although it may seem like these machines have a mind of their own, they actually mirror human experience and behavior since AI models are trained with real-life data. Being a reflection of human perception, the historical data that AI models use are not free from inherent social biases and more data is not always better specially when it is skewed. In this talk, I will discuss one of the major biases internalized by such models which is gender bias. This is a crucial problem since there is a tendency to think that machines are always perfect, and their decisions are free from human flaws which is clearly not the case here.

Key Takeaways: - Acknowledging the problem is the first steps towards solution - We need to feed meaningful data to our models which represents the level-playing field we want the world to be - Data scientists should review the context, limitations and validity of the data which includes exploring what is missing in the dataset and asking the right questions during data collection - We need to increase diversity in tech and have more allies for an enabled and equal world - AI companies need to lead by example by hiring more women in product design, data science, and engineering since diversity of thought leads to better problem-solving

Topics

Related talks

Browse the library

Impact of AI on Society

Jyotika Singh Director of Data Science, Placemakr

Round-trip 'she's a doctor' through Hungarian and AI makes her the nurse

Beginner AI and data 2022