Learning Confidence for Out-of-Distribution Detection in Neural Networks (1802.04865v1)

Abstract: Modern neural networks are very powerful predictive models, but they are often incapable of recognizing when their predictions may be wrong. Closely related to this is the task of out-of-distribution detection, where a network must determine whether or not an input is outside of the set on which it is expected to safely perform. To jointly address these issues, we propose a method of learning confidence estimates for neural networks that is simple to implement and produces intuitively interpretable outputs. We demonstrate that on the task of out-of-distribution detection, our technique surpasses recently proposed techniques which construct confidence based on the network's output distribution, without requiring any additional labels or access to out-of-distribution examples. Additionally, we address the problem of calibrating out-of-distribution detectors, where we demonstrate that misclassified in-distribution examples can be used as a proxy for out-of-distribution examples.

• The paper introduces a parallel confidence estimation branch that assigns scalar confidence scores to enhance out-of-distribution detection.
• It leverages misclassified in-distribution examples as proxies for out-of-distribution samples, yielding improvements in FPR at 95% TPR and AUROC metrics.
• The method balances task and confidence losses using a budget parameter and input preprocessing to maintain robust neural network performance in sensitive applications.

Confidence Learning for Out-of-Distribution Detection in Neural Networks

The paper presented by DeVries and Taylor introduces a method for confidence estimation in neural networks with the primary objective of enhancing out-of-distribution (OOD) detection. This method is anchored on the inherent necessity for neural networks to not only generate predictions but also to ascertain the reliability of these predictions. Given the increasing deployment of neural networks in sensitive applications, the capacity to identify inputs that fall outside the distribution of the training data is crucial. This paper proposes an approach that provides confidence estimates without requiring additional labels or access to OOD examples during training.

Key Contributions

The authors propose a parallel confidence estimation branch integrated into neural network architectures. This branch functions alongside the prediction branch and generates a scalar confidence score between 0 and 1, indicating the network's belief in the accuracy of its prediction for a given input. This methodology diverges from traditional techniques that adjust the softmax output to derive confidence scores. The innovative aspect of this approach is the use of misclassified in-distribution examples as proxies for OOD examples when calibrating the confidence estimator.

Numerical Results

The paper presents compelling empirical results demonstrating its superiority over existing baseline methods, such as those proposed by Hendrycks & Gimpel and ODIN by Liang et al. The precision of the approach is validated through several metrics, including the false positive rate at 95% true positive rate (FPR at 95% TPR), area under the receiver operating characteristic curve (AUROC), and the area under the precision-recall curve (AUPR).

For example, the paper reports a reduction in the FPR at 95% TPR when applying confidence thresholding compared to baseline methods. Specifically, in experiments with the CIFAR-10 dataset as in-distribution and various datasets as out-of-distribution, this method consistently yielded improvements in distinguishing in-distribution from out-of-distribution data.

Implementation Details

Critical to the method's success are several implementation strategies:

• A budget parameter is introduced to maintain meaningful confidence estimations throughout training by balancing the task and confidence losses.
• The model uses input preprocessing to improve OOD detection, informed by techniques such as the fast gradient sign method (FGSM) for adversarial example generation.
• The approach also incorporates a mechanism to combat excessive regularization that might result from the confidence estimation process, ensuring that the network maintains its capacity to learn complex patterns.

Implications and Future Directions

The implications of this research are both theoretical and practical. The paper addresses the pivotal challenge of OOD detection, enhancing neural network reliability by not only focusing on prediction accuracy but also on the validity of these predictions. This has substantial ramifications for AI safety, a critical aspect as AI systems become more integrated into high-stakes environments.

The paper opens up avenues for extending confidence estimation beyond classification tasks to other domains such as semantic segmentation and natural language understanding. Future work could also explore more complex schemas for the "hints" mechanism, potentially drawing inspiration from human cognitive strategies.

This paper makes a notable stride in addressing neural networks' limitations in handling novel input situations, marking a significant methodological advancement in confidence learning for neural networks.