Nothing but lies: Constructing specific neural predictors of deception
Poster No:
1862
Submission Type:
Abstract Submission
Authors:
Sangil Lee1, Runxuan Niu2, Lusha Zhu2, Andrew Kayser3, Ming Hsu1
Institutions:
1University of California, Berkeley, Berkeley, CA, 2Peking University, Beijing, Beijing, 3University of California, San Francisco, San Francisco, CA
First Author:
Co-Author(s):
Introduction:
A fundamental challenge in brain-based lie detection is to distinguish between signals of deception and confounding signals from co-occurring processes. Absent this distinction, a lie detector may well flag innocent but confounded behaviors as deceptive. First, we show that existing methods for constructing neural predictors of deception can lead to overgeneralization, in which the predictors detect not only lies but also innocent behaviors. Second, we identify a general approach to construct a statistical predictor that, by explicitly accounting for a control task, removes confounding signals while predicting the task of interest. Third, we show that this approach can generate a neural predictor of deception that does not overgeneralize, supporting the possibility that signals intrinsic to deception may exist.
Methods:
We designed a pair of isomorphic signaling games that differed only in the extent to which players' actions could be assigned a truth value. In both games the participant (the sender) is presented with two potential allocations of monetary gains for themselves and a counterpart (the receiver). On each trial, one outcome provides a larger payoff to the sender, while the other provides a larger payoff to the receiver. Critically, in the deception condition, senders must choose between two messages that are verifiably truthful or false (e.g., "Option A will earn you more money than Option B"). In contrast, in the preference condition, senders and receivers are faced with identical payoffs, but the senders' messages do not have a truth value (e.g., "I prefer that you choose Option A"). Participants are aware that as the sender, they can see the options but cannot make the choice between them, while the receiver cannot see the options but is responsible for the choice.
To establish whether the deceptive choices could be predicted by neural data, we used a leave-one-subject-out cross-validation using the T-PLS whole-brain algorithm to train neural predictors of deception on n-1 subjects and to predict the choices of the left-out subject. Using the criterion for discriminant validity, we also tested the cross-task generalizability of the neural predictor derived from the deception task. Specifically, the predictor that had previously been trained on n-1 subjects in the deception condition was tested on the left-out subject's preference condition.
To establish whether the deceptive choices could be predicted by neural data, we used a leave-one-subject-out cross-validation using the T-PLS whole-brain algorithm to train neural predictors of deception on n-1 subjects and to predict the choices of the left-out subject. Using the criterion for discriminant validity, we also tested the cross-task generalizability of the neural predictor derived from the deception task. Specifically, the predictor that had previously been trained on n-1 subjects in the deception condition was tested on the left-out subject's preference condition.
Results:
Using whole-brain MVPA, the average prediction accuracy was 75.8% for deception, which was significantly above chance at 50% (t(32) = 3.40, p = 0.0018). However, we also found strong evidence for overgeneralization, as the neural predictor of deception was also able to predict honest but selfish behavior in the preference condition at 72.7% (t(32) = 2.89, p = 0.0069).
Accordingly, we propose a dual-goal tuning approach that eliminates this overgeneralization in expectation. Our approach uses a Gram-Schmidt procedure to orthogonalize the predictor with regards to the nuisance signal such that, critically, the inner product between the predictor and the nuisance signal map is expected to be zero for out-of-sample rather than in-sample prediction. To achieve this goal, we dedicate a separate hyperparameter that is tuned to achieve out-of-sample orthogonality, while the prediction algorithm's hyperparameters are tuned to maximize prediction performances, thereby implementing a dual-goal approach.
Using the dual-goal tuning approach, our new prediction performances are specific to deception as it can significantly predict deceptive choices at 69.7% (t(32) = 2.42, p = 0.021), but not selfish choices in the preference condition (45.5 %, t(32) = -0.52, p = 0.61).
Accordingly, we propose a dual-goal tuning approach that eliminates this overgeneralization in expectation. Our approach uses a Gram-Schmidt procedure to orthogonalize the predictor with regards to the nuisance signal such that, critically, the inner product between the predictor and the nuisance signal map is expected to be zero for out-of-sample rather than in-sample prediction. To achieve this goal, we dedicate a separate hyperparameter that is tuned to achieve out-of-sample orthogonality, while the prediction algorithm's hyperparameters are tuned to maximize prediction performances, thereby implementing a dual-goal approach.
Using the dual-goal tuning approach, our new prediction performances are specific to deception as it can significantly predict deceptive choices at 69.7% (t(32) = 2.42, p = 0.021), but not selfish choices in the preference condition (45.5 %, t(32) = -0.52, p = 0.61).
Conclusions:
Beyond lie detection, studies have argued for the distinctiveness of mental processes by showing an absence of overgeneralization across datasets with naïve MVPA predictors. However, our results suggest that even when there is a considerable amount of overgeneralization, the underlying mental processes may be distinguishable.
Emotion, Motivation and Social Neuroscience:
Social Interaction
Higher Cognitive Functions:
Decision Making
Modeling and Analysis Methods:
Classification and Predictive Modeling
Methods Development 1
Multivariate Approaches 2
Keywords:
FUNCTIONAL MRI
Multivariate
Social Interactions
Statistical Methods
1|2Indicates the priority used for review