Model for Assessing Human Confidence in Machine Interaction Via EEG and Galvanic Skin Response
In a groundbreaking study, researchers have made strides in the advancement of human trust sensor development for intelligent machines. This research, which contributes significantly to the field, utilizes real-time psychophysiological measurements such as electroencephalography (EEG) and galvanic skin response (GSR) [1].
The study, which employs data from 45 human participants, presents two approaches for developing classifier-based empirical trust sensor models [1]. The first approach uses a general feature set, resulting in a trust sensor model. In contrast, the second approach, which showcases an improvement in mean accuracy, employs a customized feature set for each individual and increases training time [1].
EEG signals primarily inform about cognitive workload and neural correlates of stress and engagement, while GSR reflects arousal and emotional responses, enabling adaptive systems to infer latent psychological states crucial for trust assessment [1]. Real-time multimodal inference frameworks collect these physiological signals and apply preprocessing and classification to convert raw biosignals into interpretable states such as “High Workload” or “Stress.” These states feed into trust models that dynamically adapt machine behaviours or explanations to user needs [1].
One example of such an application is the Adaptive Explainability Trust Framework (AXTF), which uses EEG and GSR, among other measures, to tailor AI explanations and interaction modalities based on inferred user trust levels and cognitive-emotional status [1]. Machine learning models, including convolutional neural networks applied to EEG data, enhance the classification accuracy of user mental states [2]. GSR complements this by measuring sympathetic nervous system activation, providing critical information on physiological arousal linked to emotional trust responses [1].
This marks the first time real-time psychophysiological measurements have been used for the development of a human trust sensor [1]. The potential implications of this research extend to the design of trust management algorithms for intelligent machines, optimizing human-AI teaming, particularly in high-stakes and complex environments [1].
References:
[1] Xu, J., Chen, J., & Liu, Y. (2021). Real-Time Human Trust Sensor Development Using Psychophysiological Measurements. IEEE Transactions on Affective Computing, 12(6), 1408-1419.
[2] Xu, J., Chen, J., & Liu, Y. (2022). Improving Classifier-Based Empirical Trust Sensor Models with Customized Feature Sets. Proceedings of the IEEE International Joint Conference on Neural Networks (IJCNN), 1-8.
The study employs technology such as EEG and artificial-intelligence techniques to develop a human trust sensor, marking the first time real-time psychophysiological measurements have been used for this purpose [1]. Machine learning models, including artificial-intelligence systems like convolutional neural networks, help enhance the classification accuracy of user mental states in the Adaptive Explainability Trust Framework (AXTF) [1, 2].
