Accuracy Is Not Enough: Building Trustworthy AI with Conformal Prediction
summary
Conformal prediction converts point outputs into prediction sets with a user‑specified coverage (examples use 95% coverage) under the exchangeability assumption, using a calibration set and nonconformity scores to guarantee marginal coverage. Libraries and code demonstrated include TorchCP and a GitHub implementation of the API‑is‑enough sampling method that combines frequency, normalized entropy and semantic similarity to handle closed‑source LLMs. Open‑model sampling relies on logits required for open models. Demos in healthcare, NLP and robotics show detection of hallucinations, calibrated intervals, and robots asking for help.
AI-written summary · about
chapters (8)
- 0:00Trustworthy AI intro
- 2:15Machine learning in high-stakes healthcare applications
- 12:18Fixing model calibration
- 16:43NLP trustworthy AI
- 27:28Robotics safety conformal
- 30:55Conformal tools
- 32:40Key takeaways
- 34:26Q&A
related talks
31mUncertainty quantification: How much can you trust your machine learning model?
36mQuantifying uncertainty in Machine Learning predictions
35mUncertainty Unleashed
30mBoosting AI Reliability- Uncertainty Quantification with MAPIE
49mBoosting AI Reliability: Uncertainty Quantification with MAPIE
01h 24mBoosting AI Reliability: Uncertainty Quantification with MAPIE
33mAdaptive Prediction Intervals
32mLeveraging Conformal Prediction for Calibrated Probabilistic Time Series Forecasts