February 2021 Summaries

Model observability plays a crucial role in detecting, diagnosing, and explaining regressions in deployed machine learning models. Some potential failure modes include Concept Drift, where the underlying task of a model changes over time; Data Drift or Feature Drift, which occurs when the distribution of model inputs changes; Training-prod skew, where the distribution of training data differs from production data; and Cascading Model Failures, which happen when multiple models are interconnected. Additionally, Outliers can be problematic as they may represent edge cases or adversarial attacks on a model. Monitoring tools help identify these issues and enable teams to improve their models after deployment.

Eunice Kokor, the newest member of Arize AI’s Front-end Engineering team, joins from Hearst Magazines where she worked on content display as a Front-end Engineer. Eunice holds a degree in Computer Science from Columbia University and has a passion for technology and problem-solving through code. She envisions an exciting future for fair, responsible, and ethical AI, aiming to help Arize build more tooling to better understand mission-critical machine learning problems. With a background in robotics and hackathons, Eunice is eager to join the team and contribute to Arize AI’s mission.

Arize AI and Spell have partnered to bring model observability to the Spell platform, allowing users to easily transition from research to production and troubleshoot model performance without consuming data science cycles. The integration combines Spell's autoscaling online model APIs with Arize's powerful model monitoring, explainability, and troubleshooting capabilities, enabling teams to build trust between research and end-users. With this partnership, Spell users will have early access to Arize's model observability platform, while Arize users can leverage Spell as a powerful MLOps platform for building and managing machine learning projects.

ML Observability is a platform that enables teams to analyze model degradation and identify the root cause of issues by connecting points across validation and production environments. This allows for a deeper understanding of the "why" behind performance changes, which is different from traditional model monitoring that focuses on aggregates and alerts. The platform provides features such as explainability insights, production feature data analysis, and distribution drift analysis to help teams troubleshoot and improve their models. By using ML Observability, teams can gain confidence in their models' performance, scale their operations, and gain a competitive advantage in the market.