Concept: US-based AI technology startup Arize has launched a Bias Tracing tool to detect the root cause of bias in ML pipelines. Used to describe and troubleshoot AI in production, it helps ML practitioners to easily advance models from R&D to production.
Nature of Disruption: Arize’s AI observability platform offers the Bias Tracing tool to track AI performance and identify model drift. Bias Tracing can automatically detect which model inputs and slices contribute the most to bias in production and identify the fundamental cause of that bias. It is linked to Judea Pearl’s work on causal AI, which is claimed to be at the forefront of both explainable AI and AI fairness. Causal AI trains machines to learn cause and effect rather than statistical correlations. The startup employs a fairness metric called recall parity. Recall parity assesses a model’s sensitivity for a specific group compared with another and its ability to correctly forecast true positives. Arize can help an organization recognize a problem and navigate a level deeper to understand where the differential impact is most familiar for specific groups. Arize Bias Tracing is designed to work with classification models.
Outlook: Businesses have been using observability and distributed tracing to improve application performance, debug issues, and identify security risks for a long time. Observability analyzes data logs to keep track of large-scale infrastructure. For complex applications, tracing reassembles a digital twin reflecting the application logic and data flow. Arize is one of a few businesses that has adapted these techniques to improve AI monitoring. Similar techniques are used in the new bias tracing tool to generate a map of AI processing processes that span data sources, feature engineering, training, and deployment. When bias is found, it can help data managers, scientists, and engineers to detect and fix the problem’s root cause. Given the real-world implications in terms of health outcomes and loan decisions, this type of analysis is incredibly useful in domains like healthcare and finance. The startup has raised a total of $23 million in funding since 2019 and aims to expand the design to function with other use cases.