2021/09/28 by Dhaivat Bhatt, Bhatt, Dhaivat, Kaustubh Mani +9
Computer Science · #Advanced Neural Network Applications #Adversarial Robustness in Machine Learning #Computer Vision and Pattern Recognition (cs.CV) #FOS: Computer and information sciences #Machine Learning (cs.LG) #Machine Learning and Data Classification #Robotics (cs.RO)
paper · pdf · doi:10.48550/arxiv.2109.13913
openalex publication_date 2021/09/28 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
While modern deep neural networks are performant perception modules,\nperformance (accuracy) alone is insufficient, particularly for safety-critical\nrobotic applications such as self-driving vehicles. Robot autonomy stacks also\nrequire these otherwise blackbox models to produce reliable and calibrated\nmeasures of confidence on their predictions. Existing approaches estimate\nuncertainty from these neural network perception stacks by modifying network\narchitectures, inference procedure, or loss functions. However, in general,\nthese methods lack calibration, meaning that the predictive uncertainties do\nnot faithfully represent the true underlying uncertainties (process noise). Our\nkey insight is that calibration is only achieved by imposing constraints across\nmultiple examples, such as those in a mini-batch; as opposed to existing\napproaches which only impose constraints per-sample, often leading to\noverconfident (thus miscalibrated) uncertainty estimates. By enforcing the\ndistribution of outputs of a neural network to resemble a target distribution\nby minimizing an f-divergence, we obtain significantly better-calibrated\nmodels compared to prior approaches. Our approach, f-Cal, outperforms\nexisting uncertainty calibration approaches on robot perception tasks such as\nobject detection and monocular depth estimation over multiple real-world\nbenchmarks.\n