A new perspective paper argues for "measuring data" as a critical task to advance responsible AI development. Just as physical objects can be measured, data used to train AI systems should also be quantitatively analyzed to understand its composition.

The authors propose formalizing data measurement as a core research area. They categorize different measurement types like distance, density, and diversity. Distance metrics like Euclidean distance capture similarities in the data. Density reflects how densely certain concepts are represented. Diversity measures heterogeneity.

The paper summarizes measurements proposed across computer vision, natural language processing, and general data science. For instance, perplexity scores indicate how predictable text data is. Image density models visual patterns. Word mover's distance compares document similarity based on word embeddings.

The authors argue precise data measurements will let practitioners better curate datasets and understand what models learn. Characterizing training data is crucial for responsible AI. Metrics can reveal underrepresentation of groups or embed problematic biases.

For business leaders deploying AI, this paper provides a framework to evaluate data rigorously before training models. Combining multiple measurements will enable deeply understanding datasets. Just as standardized physical measurements enabled advances in science, quantifying data properties will drive progress in AI. Insisting models be trained on measured data is vital for ethical AI.

Sources:

Machine learning (ML) systems are becoming increasingly prevalent, but few business leaders understand how the data used to train these systems is produced. A new study provides insights into the complex processes and people behind ML data work.

The study was conducted by researchers who partnered with two business process outsourcing (BPO) companies that employ human workers to collect and label data used for training machine learning systems. Through interviews and design workshops with data workers, managers, and clients, the researchers aimed to understand how documentation practices could be improved to capture the context and processes involved in producing ML training data.

Key findings:

• Data workers have limited visibility into the overall goals and uses of the projects they work on. Better documentation that explains the client, project timeline, and intended uses of the final dataset could improve coordination and workers’ sense of meaning.
• Data work involves collaboration between many geographically dispersed actors, but communication tends to be top-down. Enabling feedback loops in documentation could foster more participatory, iterative processes between clients, managers, and data workers.
• Data work is dynamic, but documentation practices are often static. Tools like revision histories and timeline tracking could better reflect the evolving nature of projects and datasets.
• Worker needs are not fully represented in current documentation. Including details like task instructions, payment terms, and ethical considerations could make documentation more worker-centric.

The researchers conclude that "data production documentation should travel across multiple stakeholders and facilitate their communication." Viewing documentation as a "boundary object" that adapts to diverse needs while maintaining integrity could enable more participatory and equitable data production processes.

Key implications for business leaders:

• Seek more visibility into how your ML training data is produced and provide input to vendors on documentation needs.
• Recognize and engage data workers as key collaborators, not just low-level laborers. Their feedback can improve dataset quality.
• Set documentation policies that make values like transparency, participation, and fairness explicit. This also mitigates business risks.
• Documentation supports coordination and accountability. Invest in tools and practices that reflect the dynamic reality of data work.

Machine learning models rely heavily on their training datasets, inheriting inherent biases and limitations. Yet unlike other fields like electronics where datasheets meticulously detail components' operating parameters, no standards exist for documenting machine learning datasets. This research proposes "datasheets for datasets" to fill that gap, increasing transparency and mitigating risks.

Datasets impart significant yet often opaque influences on models. Biased data produces biased AI, like résumé screening tools that disadvantaged women. Undocumented datasets also limit reproducibility. The World Economic Forum thus recommends documenting datasets to avoid discrimination.

The authors devised an initial set of questions for dataset creators to reflect on a dataset's motivation, composition, collection, processing, uses, distribution and maintenance. For example, what subpopulations does the data represent? Could it directly or indirectly identify individuals? Might it result in unfair treatment for certain groups?

They refined the questions based on creating sample datasheets and feedback from companies pilots. A legal review led to removing explicit regulatory compliance questions. The goal is to encourage thoughtful dataset creation versus check-box documentation.

Published datasheets have already enabled better model development. At Microsoft, reviewing datasheets surfaced imbalanced gender data. But adding synthetic balanced data produced more equitable models. Datasheets also helped researchers select appropriate datasets.

Challenges remain in adapting datasheets across domains like dynamic datasets. Implementation imposes overhead on creators. And datasheets alone cannot identify every dataset bias or misuse. Combining with expertise like social scientists is recommended when data involves people.

Still, datasheets facilitate communication between creators and users, increasing intentionality in dataset development. Distinguishing datasets via transparency aids accountability and quality. For businesses, documenting training data provides reputational benefits and ensures models reflect corporate values. Proactively auditing algorithms and data is crucial as AI proliferates in customer-facing systems.

More broadly, the paper argues datasheets exemplify responsible data science in service of justice. Much like nutrition labels detail food ingredients for healthier eating, datasheets elucidate datasets' impacts. The authors urge centering inclusion and examining unintended consequences in design.

Responsible innovation requires transcending technical fixes to ask philosophical questions. What societal outcomes do we want these technologies to help achieve? How can AI empower communities historically excluded by “neutral” systems? Keeping diverse perspectives at the heart of development is key.

While datasheets provide a practical tool for dataset creators and consumers, their significance is as much symbolic as practical. They represent taking an intentional stance on transparency and equity in an opaque, high-stakes field. For businesses navigating AI's rise, this study offers inspiration to wield these technologies thoughtfully for human progress.

Sources:

Datasheets for datasets, Timnit Gebru, Jamie Morgenstern, Briana Vecchione, Jennifer Wortman Vaughan, Hanna Wallach, Hal Daumé III, Kate Crawford