Uncertainty Quantification in Deep Neural Networks through Statistical Inference on Latent Space
This work addresses the issue of overconfident predictions in deep learning, which is critical for applications requiring reliable uncertainty estimates, though it appears incremental as it builds on existing latent-space approaches.
The paper tackles the problem of overconfidence in uncertainty quantification for deep neural network classifiers by developing a method that uses latent-space representations to assess prediction accuracy. The result is a statistical model that can detect out-of-distribution data points as inaccurately predicted, aiding in outlier detection.
Uncertainty-quantification methods are applied to estimate the confidence of deep-neural-networks classifiers over their predictions. However, most widely used methods are known to be overconfident. We address this problem by developing an algorithm that exploits the latent-space representation of data points fed into the network, to assess the accuracy of their prediction. Using the latent-space representation generated by the fraction of training set that the network classifies correctly, we build a statistical model that is able to capture the likelihood of a given prediction. We show on a synthetic dataset that commonly used methods are mostly overconfident. Overconfidence occurs also for predictions made on data points that are outside the distribution that generated the training data. In contrast, our method can detect such out-of-distribution data points as inaccurately predicted, thus aiding in the automatic detection of outliers.