TY - GEN
T1 - Detecting the doubt effect and subjective beliefs using neural networks and observers’ pupillary responses
AU - Zhu, Xuanying
AU - Qin, Zhenyue
AU - Gedeon, Tom
AU - Jones, Richard
AU - Hossain, Md Zakir
AU - Caldwell, Sabrina
N1 - Publisher Copyright:
© Springer Nature Switzerland AG 2018.
PY - 2018
Y1 - 2018
N2 - We investigated the physiological underpinnings to detect the ‘doubt effect’ – where a presenter’s subjective belief in some information has been manipulated. We constructed stimulus videos in which presenters delivered information that in some cases they were led to doubt, but asked to “present anyway”. We then showed these stimuli to observers and measured their physiological signals (pupillary responses). Neural networks trained with two statistical features reached a higher accuracy in differentiating the doubt/ manipulated-belief compared to the observers’ own veracity judgments, which is overall at chance level. We also trained confirmatory neural networks for the predictability of specific stimuli and extracted significant information on those stimulus presenters. We further showed that a semi-unsupervised training regime can use subjective class labels to achieve similar results to using the ground truth labels, opening the door to much wider applicability of these techniques as expensive ground truth labels (provenance) of stimuli data can be replaced by crowd source evaluations (subjective labels). Overall, we showed that neural networks can be used on subjective data, which includes observer perceptions of the doubt felt by the presenters of information. Our ability to detect this doubt effect is due to our observers’ underlying emotional reactions to what they see, reflected in their physiological signals, and learnt by our neural networks. This kind of technology using physiological signals collected in real time from observers could be used to reflect audience distrust, and perhaps could lead to increased truthfulness in statements presented via the Media.
AB - We investigated the physiological underpinnings to detect the ‘doubt effect’ – where a presenter’s subjective belief in some information has been manipulated. We constructed stimulus videos in which presenters delivered information that in some cases they were led to doubt, but asked to “present anyway”. We then showed these stimuli to observers and measured their physiological signals (pupillary responses). Neural networks trained with two statistical features reached a higher accuracy in differentiating the doubt/ manipulated-belief compared to the observers’ own veracity judgments, which is overall at chance level. We also trained confirmatory neural networks for the predictability of specific stimuli and extracted significant information on those stimulus presenters. We further showed that a semi-unsupervised training regime can use subjective class labels to achieve similar results to using the ground truth labels, opening the door to much wider applicability of these techniques as expensive ground truth labels (provenance) of stimuli data can be replaced by crowd source evaluations (subjective labels). Overall, we showed that neural networks can be used on subjective data, which includes observer perceptions of the doubt felt by the presenters of information. Our ability to detect this doubt effect is due to our observers’ underlying emotional reactions to what they see, reflected in their physiological signals, and learnt by our neural networks. This kind of technology using physiological signals collected in real time from observers could be used to reflect audience distrust, and perhaps could lead to increased truthfulness in statements presented via the Media.
KW - Doubt
KW - Information veracity
KW - Neural networks
KW - Pupillary responses
KW - Semi-unsupervised training
KW - Subjective belief
KW - Trust
UR - http://www.scopus.com/inward/record.url?scp=85059012296&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85059012296&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-04212-7_54
DO - 10.1007/978-3-030-04212-7_54
M3 - Conference contribution
AN - SCOPUS:85059012296
SN - 9783030042110
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 610
EP - 621
BT - Neural Information Processing - 25th International Conference, ICONIP 2018, Proceedings
A2 - Ozawa, Seiichi
A2 - Leung, Andrew Chi Sing
A2 - Cheng, Long
PB - Springer Verlag
T2 - 25th International Conference on Neural Information Processing, ICONIP 2018
Y2 - 13 December 2018 through 16 December 2018
ER -