The paper addresses the issue of confidence calibration in modern neural networks, which is crucial for classification models in various applications. It finds that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, the authors identify depth, width, weight decay, and Batch Normalization as key factors influencing calibration. They evaluate various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. The analysis and experiments reveal that *temperature scaling*, a single-parameter variant of Platt Scaling, is surprisingly effective at calibrating predictions. This method is straightforward to implement and can be easily adopted in practical settings. The paper also provides insights into neural network learning and architectural trends that may cause miscalibration.The paper addresses the issue of confidence calibration in modern neural networks, which is crucial for classification models in various applications. It finds that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, the authors identify depth, width, weight decay, and Batch Normalization as key factors influencing calibration. They evaluate various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. The analysis and experiments reveal that *temperature scaling*, a single-parameter variant of Platt Scaling, is surprisingly effective at calibrating predictions. This method is straightforward to implement and can be easily adopted in practical settings. The paper also provides insights into neural network learning and architectural trends that may cause miscalibration.