http://yiziruifang.com/tags/robust-ml/Robust MLHugo Blox Builder (https://hugoblox.com)en-usSat, 01 Jun 2024 00:00:00 +0000http://yiziruifang.com/media/icon_hu7729264130191091259.pnghttp://yiziruifang.com/tags/robust-ml/http://yiziruifang.com/project/conformal-predictors/Sat, 01 Jun 2024 00:00:00 +0000http://yiziruifang.com/project/conformal-predictors/<h2 id="hiring-manager-view">Hiring-manager view</h2> <p>This paper is a direct uncertainty-quantification signal: it studies the data and calibration choices behind conformal prediction rather than treating uncertainty estimates as a black box.</p> <h2 id="scientific-problem">Scientific problem</h2> <p>Inductive conformal predictors rely on data splits and calibration sets to produce uncertainty-aware prediction sets. The practical question is how data usage decisions affect validity, efficiency, and downstream model behavior.</p> <h2 id="method">Method</h2> <ul> <li>Investigated how data allocation choices influence inductive conformal prediction.</li> <li>Focused on calibration behavior, data efficiency, and prediction-set quality.</li> <li>Connected empirical model behavior to uncertainty guarantees relevant to safety-sensitive ML systems.</li> </ul> <h2 id="evaluation-signal">Evaluation signal</h2> <p>The evaluation centers on how calibration and data usage choices change uncertainty quality and reliability, especially when model outputs must support downstream decisions.</p>