Poster: How Safe Am I Given What I See? Calibrated Prediction of Safety Chances for Image-Controlled Autonomy (2024)
0 likes40 views
The document discusses the development of a calibrated prediction methodology for assessing safety in image-controlled autonomous systems at the University of Florida. It outlines challenges and contributions related to conformal confidence calibration, evaluation results comparing monolithic and composite predictors, and emphasizes the superiority of calibrated predictors. The findings indicate that calibrated predictions outperform uncalibrated ones, demonstrating reliable coverage for conformal calibration.
Poster: How Safe Am I Given What I See? Calibrated Prediction of Safety Chances for Image-Controlled Autonomy (2024)
- 1. Zhenjiang Mao, Carson Sobolewski, Ivan Ruchkin Trustworthy Engineered Autonomy (TEA) Lab Department of Electrical and Computer Engineering, University of Florida How Safe Am I Given What I See? Calibrated Prediction of Safety Chances for Image-Controlled Autonomy APPROACH FUTURE WORK RESULTS: PREDICTION CHALLENGES & CONTRIBUTIONS PROBLEM REFERENCES ■ Given a system with state x, observation y, and a safety property φ(x): ■ Is the system safe now? φ(xi | yi ) ■ What will be the observation after k steps given past observations? yi+k | yi-j:i ■ Will the system be safe after k steps given past observations? φ( xi+k | yi-j:i ) ■ What’s the probability of safety after k steps given a sequence of observations? Pr(φ(xi+k ) | yi-j:i ) Conformal confidence calibration: Draw N observation-state pairs {b1 , b2 , ..., bN } M times from calibration datasets to get M datasets {B1 , B2 , ..., BM }. By implementing Conformal Calibration Algorithm on predictor g given miscoverage level 𝜶 and M datasets, we can get an upper calibration error bound c for a newly drawn BM+1 that: Challenges: ■ Unknown states and dynamical models ■ High-dimensional (1000+) observations ■ Difficult to reliably quantify confidence Contributions: ■ Modular family of online safety predictors ■ Conformal confidence calibration ■ Evaluation on a racing car and a cart pole ■ Conclusion 1: Monolithic (mon) predictors outperform composite (comp) ones. ■ Conclusion 2: Controller-independent (ind) and controller-specific (c-sp) predictors show comparable performance. ■ Adding physical constraints to latent states similar to Neural ODEs (Wen, Wang, and Metaxas 2022) ■ Jointly learning forecasters and evaluators to overcome distribution shifts ■ Using world-model transformer architectures ■ Applying predictors to physical systems Conformal prediction: Vovk, V.; Gammerman, A.; and Shafer, G. 2005. Algorithmic Learning in a Random World. New York: Springer, 2005. ISBN 978-0-387-00152-4. World models: Ha, D.; and Schmidhuber, J. 2018. Recurrent World Models Facilitate Policy Evolution. In Advances in Neural Information Processing Systems, volume 31. Confidence calibration: Guo, C.; Pleiss, G.; Sun, Y.; and Weinberger, K. Q. 2017. On calibration of modern neural networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), Sydney, Australia. RESULTS: CALIBRATION ■ Conclusion 3: Calibrated predictors are superior to uncalibrated ones. ■ Conclusion 4: Conformal calibration coverage is reliable. L to R: (1) uncalibrated; (2): calibrated w/ isotonic regression, conformal bounds for 𝛼=0.05; (3) ECE/MCE over varied k before/after calibration; (4) our conformal bounds for 𝛼=0.05 (blue) contain more than 95% of the true calibration errors (box and whisker plots). L to R: (1) controller-specific (`c-sp') monolithic (`mon') vs. latent composite (`comp') for the racing car; (2) controller-independent (`ind') monolithic vs. latent composite for the racing car; (3) controller-specific monolithic vs. latent composite for the cart pole; (4) controller-independent monolithic vs. latent composite for the cart pole. Performance of safety label predictors over varied horizons Shaded uncertainty shows standard deviation due to different controllers and resampling. where q is the mean true safety and p is the mean safety chance prediction scores: Pr( |qM+1 − pM+1 | ≤ c ) ≥ 1 − α