Core Concepts
Conformal prediction can be used to detect if deep learning models are out-of-calibration during onboard processing.
Abstract
The paper explores the relationship between conformal prediction and model uncertainty, and exploits this relationship to perform onboard out-of-calibration detection for deep learning models.
Key highlights:
- Conformal prediction provides finite sample coverage guarantees in the form of a prediction set that is guaranteed to contain the true class within a user-defined error rate.
- The average size of the conformal prediction set is related to the uncertainty of the deep learning model. Uncertain models tend to have larger prediction sets, while overconfident models have smaller prediction sets.
- Under noisy scenarios, the outputs of uncertain models like ResNet50 become untrustworthy, leading to an increase in the average prediction set size. This can be used to detect if the model is out-of-calibration.
- Overconfident models like InceptionV3 and DenseNet161 cannot be easily detected as out-of-calibration using the prediction set size alone, as their outputs remain overconfident even under noise.
- The paper demonstrates the out-of-calibration detection procedure using popular classification models like ResNet50, DenseNet161, InceptionV3, and MobileNetV2 on the EuroSAT remote sensing dataset.
Stats
The average normalized softmax entropy increases with increasing noise severity for the ResNet50 model, indicating higher uncertainty.
The average normalized softmax entropy remains constant for the InceptionV3 and DenseNet161 models, indicating overconfidence.
The average prediction set size increases significantly for the ResNet50 model under noisy conditions, while the increase is negligible for the InceptionV3 and DenseNet161 models.
Quotes
"If exchangeability is violated, e.g., due to covariate shift [13] or noise, then (1) does not hold. How do we detect, autonomously, that a covariate shift has taken place and the model is out-of-calibration?"
"An overconfident model tends to place importance on a single class irrespective of whether the prediction is correct or not. A distribution of the softmax outputs of such a network will peak near higher values. However, if a model is uncertain, then the model tries to output a flat softmax distribution that reflects the uncertainty in its predictions."