Incorporate Input Sensor Uncertainties in Neural Networks

Abstract

The upcoming release of 6th generation (6G) networks calls for a pervasive usage of artificial intelligence (AI), and in particular neural networks (NNs), to support critical sensing and communication tasks under uncertainty. However, conventional NNs are not designed to ingest per-feature uncertainty, especially in scenarios where sensors provide both measurements and associated confidence levels. This limitation leads to overfitting to noisy inputs and unreliable confidence estimates, which are detrimental in applications such as angle of arrival (AoA) estimation. In this paper, we propose a novel method called gradient-integrated uncertainty network (GIUN), which analytically propagates input uncertainty through deep NNs without relying on Monte Carlo (MC) sampling. GIUN leverages the Jacobian of the network output with respect to the input to compute a closed-form marginal likelihood, enabling calibrated uncertainty estimation with deterministic inference latency. We evaluate GIUN on a wireless sensing task involving AoA estimation using a uniform linear array (ULA) of antennas. Our results show that GIUN achieves up to 17% lower root mean square error (RMSE) and an expected calibration error (ECE) as low as 2% compared to state-of-the-art baselines. These improvements demonstrate the effectiveness of GIUN in producing reliable predictions under uncertainty, while remaining lightweight and scalable for deployment in 6G user equipment.