Title: Advancing Fairness and Inclusion in AI: Google’s Responsible AI Research
Google’s Responsible AI research emphasizes collaboration, bringing together diverse teams and engaging with the community. The Perception Fairness team, with expertise in computer vision and machine learning fairness, plays a crucial role in driving progress. Their mission is to design inclusive AI systems guided by Google’s AI Principles. The team focuses on developing advanced models for various applications, such as classification, captioning, and visual question answering, while prioritizing fairness and inclusion.
Understanding Fairness in AI:
Designing Fair AI: The team works to model human perceptions of demographics, cultures, and social identities responsibly using machine learning. They aim to measure system biases, tackle algorithmic failures, and build more inclusive algorithms.
Reducing Representational Harms: Google’s research delves into analyzing media content and reducing representational harms, such as stereotypes or denigration of certain groups. They employ sociology and social psychology to understand different perceptions and use computational tools to create scalable solutions.
Studying Media Representation: Through the MUSE project, in collaboration with academic researchers and brands, the team examines patterns in mainstream media. Their in-depth analyses provide insights to content creators, advertisers, and contribute to their own research.
Expanding the Scope of Fairness Research: Google aims to develop tools that assess representation in illustrations, humanoid characters, and non-people images. They also pay attention to narrative communication and cultural context, not just depicting the subjects but how they are portrayed.
Analyzing Bias Properties of AI Systems:
Measuring Nuanced System Behavior: While traditional summary statistics simplify bias analysis, Perception Fairness focuses on nuanced system behavior through detailed metrics. Balancing fairness metrics with other product metrics can be challenging but crucial.
Democratizing Fairness Analysis Tooling: The team aims to make fairness analysis tools accessible and widely-used by partnering with organizations and developing benchmarks and test datasets.
Advancing Fairness Analytics: Google partners with product teams to advance novel approaches to fairness analytics. These efforts contribute breakthrough findings and inform the launch strategies of AI systems.
Advancing AI Responsibly:
Improving AI Algorithms: Google not only analyzes model behavior but also works on algorithmic improvements. They have launched upgraded components for Google Photos and Google Images, resulting in more consistent and diversified representation.
Mitigating Failure Modes: The team collaborates with the responsible AI community to develop guardrails that prevent AI systems from failing. They invest in research from pre-training to deployment to ensure the generation of high-quality, inclusive, and controllable outputs.
Ongoing Work and Opportunities: Despite progress, the field of perception fairness in AI is still growing rapidly. Google continues to seek opportunities for technical advancements, bridging the gap between measuring images and understanding human identity and expression.
Google’s Perception Fairness team is committed to advancing fairness and inclusion in AI systems. Through interdisciplinary research and collaboration, they strive to build responsible AI models that accurately represent diverse communities and promote inclusive experiences for users worldwide.