UAC: Uncertainty-Aware Calibration of Neural Networks for Gesture Detection

TL;DR

The paper addresses reliable gesture detection from IMU data in safety-critical settings by targeting both prediction accuracy and probability calibration under distribution shifts. It introduces Uncertainty-Aware Calibration (UAC), a two-stage framework that first learns uncertainty-aware gesture probabilities from single IMU samples and then calibrates these predictions via Monte Carlo integration and an entropy-weighted, multi-sample aggregation over gesture sequences. Across three public IMU datasets, UAC outperforms temperature scaling, entropy maximization, and Laplace-based baselines in both in-distribution and out-of-distribution scenarios, delivering higher accuracy and better calibration, with notable gains on Wisdm. The work demonstrates practical impact for privacy-preserving gesture recognition in real-world safety-critical applications, offering a scalable approach with competitive inference speed and clear avenues for extension to additional sensor modalities and edge deployment.

Abstract

Artificial intelligence has the potential to impact safety and efficiency in safety-critical domains such as construction, manufacturing, and healthcare. For example, using sensor data from wearable devices, such as inertial measurement units (IMUs), human gestures can be detected while maintaining privacy, thereby ensuring that safety protocols are followed. However, strict safety requirements in these domains have limited the adoption of AI, since accurate calibration of predicted probabilities and robustness against out-of-distribution (OOD) data is necessary. This paper proposes UAC (Uncertainty-Aware Calibration), a novel two-step method to address these challenges in IMU-based gesture recognition. First, we present an uncertainty-aware gesture network architecture that predicts both gesture probabilities and their associated uncertainties from IMU data. This uncertainty is then used to calibrate the probabilities of each potential gesture. Second, an entropy-weighted expectation of predictions over multiple IMU data windows is used to improve accuracy while maintaining correct calibration. Our method is evaluated using three publicly available IMU datasets for gesture detection and is compared to three state-of-the-art calibration methods for neural networks: temperature scaling, entropy maximization, and Laplace approximation. UAC outperforms existing methods, achieving improved accuracy and calibration in both OOD and in-distribution scenarios. Moreover, we find that, unlike our method, none of the state-of-the-art methods significantly improve the calibration of IMU-based gesture recognition models. In conclusion, our work highlights the advantages of uncertainty-aware calibration of neural networks, demonstrating improvements in both calibration and accuracy for gesture detection using IMU data.

Figures (5)

• Figure 1: Flowchart of the 2-step process of our method. 1) The model is trained on the training set to detect gestures and estimate the uncertainty associated with the data and the prediction. 2) The model is used (in out-of or in-distribution scenarios) to aggregate predictions in an uncertainty-aware manner to improve the overall performance while maintaining network calibration.
• Figure 2: Two-step pipeline for entropy-weighted gesture detection. In the first step (\ref{['fig:method:step1']}), an uncertainty-aware classification network $M_u$ is trained on a single sample of IMU data. $M_u$ is then used (\ref{['fig:method:step2']}) to obtain predictions from multiple samples and the final prediction is the entropy-weighted prediction from those samples.
• Figure 3: The Entropy Maximization (EM) max-entropy baseline model is trained to maximize the entropy of wrong predictions through an entropy-maximization loss.
• Figure 4: Temperature Scaling pmlr-v70-guo17a consists of a classification model $M_c$ where the logits are scaled by a factor $T$. $T$ is learned after training $M_c$, to improve the calibration of the model on a validation set.
• Figure 5: Implementation of the last-layer Bayesian Neural Network with Laplacian approximation NEURIPS2021_a7c95857 baseline used in the experiments. The pre-trained model is assumed to be a feature map and the Laplace-approximated posterior is calculated for the last layer only.