February 2022 Summaries

ML Troubleshooting Is Too Hard Today (But It Doesn’t Have To Be That Way) The stakes for model performance are higher than ever as teams deploy more models into production, and mistakes costlier. A modern approach to ML troubleshooting is needed, shifting from no monitoring to full stack ML observability. Monitoring, at its core, requires data on system performance, which must be made storable, accessible, and displayable. To monitor model performance, one must begin with a prediction and actual, comparing them using the right metric. The correct metric depends on the use case, such as recall, false negative rate, or mean absolute percentage error for fraud models, and mean squared error for demand forecasting models. Establishing thresholds is crucial to determine when a good accuracy rate has become bad enough. Machine learning practitioners must rely on relative metrics and establish a baseline performance to define what is considered good enough. Monitoring alone is not enough, as it's essential to have a modern approach to assessing and troubleshooting model performance, including full stack ML observability with ML performance tracing.

As Director of Engineering and Data Science at Shopify, Wendy Foster leads the development and deployment of sophisticated AI systems that empower millions of merchants worldwide to market and grow their retail businesses. Foster's background in game development and humanities informs her perspective on AI ethics, emphasizing the importance of understanding the impact of technology on users' lives. She prioritizes collaboration with business counterparts to ensure governance goals, responsible AI, and AI risk management. Foster also highlights the need for observability over explainability, arguing that accountability drives operational excellence. Her focus on representation is critical, recognizing that diverse makers and datasets are essential for building a world that technology serves. Ultimately, Foster's work is driven by her passion for empowering entrepreneurs and small businesses to thrive through AI-powered solutions.

Arize AI's recent survey of 945 data scientists, ML engineers, technical executives, and others highlights key challenges faced by MLOps teams. Troubleshooting model issues remains a significant problem for many, with 84.3% of respondents reporting delays in detecting and diagnosing problems at least some of the time. Additionally, communication between ML teams and business executives is often hindered, with over half of data scientists and ML engineers encountering issues with quantifying ROI or explaining machine learning concepts to stakeholders. While explainability remains important, it should not be relied upon solely; instead, a proactive approach to model performance management is recommended.

In the context of machine learning observability, "drift" refers to changes in the statistical properties of data or models over time. Model drift occurs when a model's predictions change without any modification to the underlying model itself. Concept drift is characterized by shifts in the statistical properties of the target variable, while data drift involves changes in the independent variables and their correlations. Upstream drift results from alterations in the data pipeline that can lead to missing values or changes in feature cardinality. Monitoring and diagnosing these various forms of drift are crucial for maintaining optimal model performance and mitigating future performance degradation. Arize is an ML observability platform designed to help teams manage model performance, monitor drift, and troubleshoot issues in production environments.