Channel's GitHub page hosting Jupyter Notebook: [ Ссылка ]
In this video, we explore the concept of uncertainty quantification in machine learning predictions and introduce a powerful class of methods called Conformal Predictors. Learn why relying solely on point predictions can be problematic and discover how quantifying uncertainty empowers informed decision-making. We delve into the concept of prediction intervals with attached probabilistic statements, discuss the importance of coverage validity and efficiency, and explore the challenges of achieving finite-sample validity. We also highlight the desirable properties of model-agnostic and distribution-free interval predictors. Join us on this journey to understand how Conformal Predictors provide model-agnostic, distribution-free intervals with finite-sample validity. Enhance your understanding of uncertainty quantification and expand your UQ methods repertoire.
Keywords: uncertainty quantification, Conformal Predictors, prediction intervals, coverage validity, efficiency in prediction intervals, finite-sample validity, model-agnostic interval predictors, distribution-free interval predictors, machine learning predictions, informed decision-making, point predictions, error probability, linear regression, neural network
0:00 - Why Uncertainty Quantification
0:20 - Limitations of Point Predictions
1:20 - Going Beyond Point Predictions
1:40 - Empowering Decision-Making with Uncertainty Quantification
2:02 - Predicting Intervals with Probabilistic Statements
2:43 - Embracing Error and Quantifying Its Probability
3:01 - Importance of Coverage Validity
3:40 - Verifying Claimed Coverage Validity
3:52 - Finite-sample Validity
4:33 - The Role of Efficiency in Reliable Prediction Intervals
5:15 - Beyond Validity and Efficiency: Model-Agnostic and Distribution-Free Interval Predictors
5:54 - The Desirable Properties of Interval Predictors
6:30 - How Conformal Predictors Work
