On Calibration of Modern Neural Networks
By Chuan Guo, Geoff Pleiss, Yu Sun, Kilian Q. Weinberger. 2017
8,827 citations recorded by the source corpus.
Confidence calibration -- the problem of predicting probability estimates representative of the true correctness likelihood -- is important for classification models in many applications. We discover that modern neural networks, unlike those from a decade ago, are poorly calibrated. Through extensive experiments, we observe that depth, width, weight decay, and Batch Normalization are important factors influencing calibration. We evaluate the performance of various post-processing calibration methods on state-of-the-art architectures with image and document classification datasets. Our analysis and experiments not only offer insights into neural network learning, but also provide a simple and straightforward recipe for practical settings: on most datasets, temperature scaling -- a single-parameter variant of Platt Scaling -- is surprisingly effective at calibrating predictions.
Subfields: Artificial Intelligence, Safety, Robustness & Interpretability
How researchers see this paper
The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.
In Artificial Intelligence
- Mohit Bansal
- Yoshua Bengio
- Quoc V. Le
- Percy Liang
- Sergey Levine
- Jianfeng Gao
- Luke S. Zettlemoyer
- Ruslan Salakhutdinov
In Safety, Robustness & Interpretability
- Matthias Hein
- Dan Hendrycks
- Percy Liang
- Nicholas Carlini
- Eric Wallace
- Prateek Mittal
- Dawn Song
- Yixuan Li
Citations in the corpus
- Exploiting Cloze-Questions for Few-Shot Text Classification and Natural Language Inference (2020, background): “The underlying intuition is that even without calibration, examples for which labels are predicted with high confidence are typically more likely to be classified correctly (Guo et al., 2017).”
- Measuring Massive Multitask Language Understanding (2020, background): “However, large neural networks are often miscalibrated (Guo et al., 2017), especially under distribution shift (Ovadia et al., 2019).”
- Natural Adversarial Examples (2019, methodology): “r is q E C[(P(Y = YbjC= c) c)2]; or the squared discrepancy between the classifier’s accuracy at a confidence level and the confidence level. We estimate the RMS Calibration Error using adaptive binning [36,22,17]. Here we take the confidence to be the maximum softmax probability [21]. 7 Our second unce”
- Benchmarking Neural Network Robustness to Common Corruptions and Perturbations (2019, background): “However such a measure could make it difficult to compare models since some models tend to be more uncalibrated than others (Guo et al., 2017).”
- Scalable agent alignment via reward modeling: a research direction (2018, background): “Other works aim to calibrate the predictions of neural networks (Guo et al., 2017), so that their subjective uncertainty corresponds with their empirical frequency of mistakes.”
- Large Language Models Can Self-Improve (2022, neutral): “In the following sections, we detail important designs within our method, along with additional approaches for the model to self-improve without supervised data.”
- Calibrate Before Use: Improving Few-Shot Performance of Language Models (2021, methodology): “6 In this paper, we restrict the matrix W to be diagonal, known as vector scaling (Guo et al., 2017), to prevent the parameters from growing quadratically in the size of p̂ (which is ≈ 50, 000 for generation tasks).”
- Model soups: averaging weights of multiple fine-tuned models improves accuracy without increasing inference time (2022, methodology): “The calibrated soup and calibrated ensemble refer to a soup and ensemble composed of models which are calibrated through temperature scaling (Guo et al., 2017).”
- Consistent Accelerated Inference via Confident Adaptive Transformers (2021, methodology): “A large amount of research has been dedicated towards calibrating the model posterior, p θ (ˆ y n +1 | x n +1 ) , such that the accuracy, y n +1 = ˆ y n +1 , is indeed equal to the estimated probability (Niculescu-Mizil and Caruana, 2005; Gal and Ghahramani, 2016; Guo et al., 2017).”
- Dynamic Neural Networks: A Survey (2021, background): “Note that the overconfidence issue in deep models [201], [202] might affect the effective-”
- The Right Tool for the Job: Matching Model and Instance Complexities (2020, methodology): “In this paper we use temperature calibration, which is a simple technique that has been shown to work well in practice (Guo et al., 2017), in particular for BERT fine-tuning (Desai and Durrett, 2020).”
- AugMix: A Simple Data Processing Method to Improve Robustness and Uncertainty (2019, methodology): “(2017); Guo et al. (2019), we use a weight decay of 0.”
- Unsolved Problems in ML Safety (2021, background): “To help models express their domain of competence in a more representative and meaningful way, researchers could further improve model calibration on typical testing data [68, 131, 112, 108, 205, 106, 107, 123], though the greater challenge is calibration on testing data that is unlike the training ”
- Simple and Scalable Predictive Uncertainty Estimation using Deep Ensembles (2016, background): “[20] have also independently shown that non-Bayesian solutions can produce good predictive uncertainty estimates on some tasks.”
- Simple and Principled Uncertainty Estimation with Deterministic Deep Learning via Distance Awareness (2020, methodology): “[29] C.”
- Beyond Individual and Group Fairness (2020, background): “…of recent research work on the topic of fairness and algorithmic bias in machine learning [Dwork et al., 2012, Zemel et al., 2013, Hardt et al., 2016, Kleinberg et al., 2017, Pleiss et al., 2017, Agarwal et al., 2018, Kearns et al., 2018, Gillen et al., 2018, Sharifi-Malvajerdi et al., 2019].”
- Privacy Risks of Securing Machine Learning Models against Adversarial Examples (2019, methodology): “Thus, a straightforward mitigation method is to reduce this difference by applying temperature scaling on logits [17].”