“Harm and Bias Evaluation and Solution for Adobe Firefly,” a Presentation from Adobe

For the full video of this presentation, please visit: https://www.edge-ai-vision.com/2024/09/harm-and-bias-evaluation-and-solution-for-adobe-firefly-a-presentation-from-adobe/

Rebecca Li, Machine Learning Engineering Manager at Adobe, presents the “Harm and Bias Evaluation and Solution for Adobe Firefly” tutorial at the May 2024 Embedded Vision Summit.

In this talk, Dr. Li explores the comprehensive approach Adobe has taken to mitigate harm and bias for Firefly, Adobe’s groundbreaking AI art generation tool, integrating AI ethics across all development stages. From design to deployment, Firefly prioritizes ethical considerations, embedding AI ethics into every process.

Li showcases strategies for identifying and mitigating potential harm and bias, utilizing advanced AI techniques to ensure content safety. Additionally, she emphasizes her company’s active involvement in shaping the global discourse on technology, highlighting the importance of collective responsibility in fostering ethical AI practices. You’ll learn how to navigate the complex landscape of AI ethics, with insights and strategies to promote responsible innovation and accountability in the digital era.

1. Harm and Bias Evaluation
and Solution for Adobe
Firefly
Dr. Xiaoyang (Rebecca) LI
ML Engineer Manager, Firefly Eval Science
Adobe

2. Harm & Bias
2© 2024 Adobe
Unaddressed Harm & Bias issues can stop models & products
Microsoft Tay, 2016 Meta’s Galactica, 2022

3. Mitigate Harmful Bias & Unsafe Content
Results
from
Adobe
Firefly
?
What is Human Bias?

4. There are many types of representational harms
Footer Goes Here 4
•Missing representations (no result)
•Mislabeled identities / Inaccurate depictions
•Stereotyping
•Over and under representation
•Dehumanization
•Cultural/religious insensitivities

5. Adobe Solution: AI Ethics Program
5© 2024 Adobe

6. Design
6© 2024 Adobe
Before launch, AI ethics committee reviews the
scientific evaluation of the possible intentional &
unintentional generation of:
•Nudity (adult & child)
•Ethnicity & identity (mis)representation
•Violence & gore
•Hate content (signs, hand gestures, etc.)
Interdisciplinary team: go/no-go decision

7. Goal:
•Mitigate bias in AI training sets and commit to preventing the
perpetuation of stereotypes and harm in model output
Approach: Data filtering & enrichment
•Human and automatic moderation removes harmful content
•Smart algorithms enhance data quality with limited resources.
•Data acquisition to close data gaps
Development: Model Training
7© 2024 Adobe

8. Goal:
•Testing of Firefly features and product to mitigate against harmful
biases and stereotypes
Approach:
•Automated testing(ML metrics & models)
•Human Evaluation (semi-automatic pipeline)
Development: Model Testing and Evaluation
8© 2024 Adobe

9. Case Study: Quantifying Harm Types for Text to Image
9

10. Case Study: Quantifying Likelihood of Nudity for GenFill
10© 2024 Adobe

11. Case Study: Evaluating Bias for Generated Humans
11© 2024 Adobe

12. Case Study: Evaluating Bias for Generated Humans
12© 2024 Adobe

13. Result Analysis: Evaluating Bias for Skin Color Diversity
Image Model B
Image Model A
Skin Color

14. Result Analysis: Evaluating Bias for Gender Diversity
Image Model B
Image Model A
Gender

15. Result Analysis: Evaluating Bias for Age Diversity
Age
Image Model B
Image Model A

16. Harm:
•Can’t load harm prompt
•Strict control on unintentional harm
Bias
•Increasing variety during prompt inference
•Customized debiasing strategies for different regions
Deployment
16© 2024 Adobe

17. Feedback by Users
17© 2024 Adobe
harmful and biased content feedback on output quality

18. Address potential harm & bias issues online/in real-time
•Requires guardrails in pre- & post-processing
•Mixture of approaches, interdisciplinary team
Proactive & reactive mitigation
•Source content to improve human representation diversity
•Red teaming, social media watch
Harm & Bias Mitigation Strategy
18© 2024 Adobe

19. Adobe Firefly
https://firefly.adobe.com/
Adobe AI Ethic
https://www.adobe.com/ai/overview/ethics.html
Reducing biased and harmful outcomes in generative AI
https://adobe.design/stories/leading-design/reducing-biased-and-harmful-outcomes-
in-generative-ai
Resource
19© 2024 Adobe

20. Questions?
20© 2024 Adobe