Pretraining with random noise for uncertainty calibration

TL;DR

Pretraining neural networks with random noise and labels improves uncertainty calibration by reducing overconfidence and enhancing the detection of out-of-distribution data, without extra calibration steps.

Contribution

Introducing a simple pretraining method with random noise and labels inspired by neuroscience to improve confidence calibration in neural networks.

Findings

Pretraining with noise reduces overconfidence in untrained networks.

Pretrained networks better identify out-of-distribution inputs.

Calibration is achieved without additional post-processing.

Abstract

Uncertainty calibration is crucial for various machine learning applications, yet it remains challenging. Many models exhibit hallucinations - confident yet inaccurate responses - due to miscalibrated confidence. Here, we show that the common practice of random initialization in deep learning, often considered a standard technique, is an underlying cause of this miscalibration, leading to excessively high confidence in untrained networks. Our method, inspired by developmental neuroscience, addresses this issue by simply pretraining networks with random noise and labels, reducing overconfidence and bringing initial confidence levels closer to chance. This ensures optimal calibration, aligning confidence with accuracy during subsequent data training, without the need for additional pre- or post-processing. Pre-calibrated networks excel at identifying "unknown data," showing low confidence for out-of-distribution inputs, thereby resolving confidence miscalibration.