Conference PaperPDF Available

Modelling cognitive bias in safety using Bayesian inference

  • August 2022
  • Conference: 15th World Congress on Computational Mechanics(WCCM-XV)
  • At: Invited talk
Authors:
Hideyoshi Yanagisawa at The University of Tokyo
Hideyoshi Yanagisawa
Download full-text PDF
Read full-text
Download citation

Abstract

In a human centred society, sense of safety perceived by people is important to consider as well as actual safety. The gap between expectations and reality in safety causes social problems. Overestimation of safety provokes dangerous behaviours, while underestimation of safety causes excessive behavioural restraint. For example, in COVID 19 pandemic, the overestimation may lead to the spread of infection due to dangerous behaviours, and the underestimation may lead to slow consumption due to excessive self-control. In this talk, I propose a mathematical framework to model sense of safety using Bayesian inference. Recent neuroscience studies suggest that human brain activities can be explained as a Bayes’ model(Knill & Pouget, 2004). Here, based on Helmholtz’s epistemology, I assume that safety x is inferred as the cause of observation (data: y) and define the sense of safety as Bayesian posterior p(x|y) that is proportional to a product of a prior p(x) and likelihood p(y|x). The prior and the likelihood represent the expectation of safety and the safety based solely on the data (observed safety), respectively. In this model, the prior (safety belief) is updated to the posterior (perception of safety). The difference between the posterior mean and the peak of likelihood represents the gap between perceived and observed safety. This gap is regarded as a cognitive bias termed expectation effect (Yanagisawa, 2016). In our previous study, we modelled the expectation effect as a function of three parameters: a prediction error (difference between prior mean and peak of likelihood), prior precision (inverse variance of the prior), and observation precision (inverse variance of the likelihood). We found that there are two types: assimilation and contrast. Assimilation diminishes the prediction error, and contrast exaggerates the prediction error. The expectation effect explains several psychological biases that lead to abnormal behaviours, such as normalcy bias, excessive anxiety, and overconfidence. Based on the expectation effect model, I discuss the condition of each psychological bias in safety. In addition, I discuss how to model emotions such as anxious and fear in a mathematical manner by applying information theoretic quantities such as surprisal and free energy (Yanagisawa, Kawamata, & Ueda, 2019).
Content uploaded by Hideyoshi Yanagisawa
Author content
All content in this area was uploaded by Hideyoshi Yanagisawa on Sep 19, 2022
Content may be subject to copyright.
15th World Congress on Computational Mechanics (WCCM-XV)
8th Asian Pacific Congress on Computational Mechanics (APCOM-VIII)
31 July – 5 August 2022, Yokohama, Japan
Modelling Cognitive Bias in Safety using Bayesian Inference
Hideyoshi Yanagisawa1*
1 The University of Tokyo
7-3-1 Hongo, Bunkyo, Tokyo, JAPAN, hide@mech.t.u-tokyo.ac.jp
Key Words: Safety, Perception, Cognitive Bias, Bayes, Covid-19
In a human centred society, sense of safety perceived by people is important to consider as well
as actual safety. The gap between expectations and reality in safety causes social problems.
Overestimation of safety provokes dangerous behaviours, while underestimation of safety
causes excessive behavioural restraint. For example, in COVID 19 pandemic, the
overestimation may lead to the spread of infection due to dangerous behaviours, and the
underestimation may lead to slow consumption due to excessive self-control.
In this talk, I propose a mathematical framework to model sense of safety using Bayesian
inference. Recent neuroscience studies suggest that human brain activities can be explained as
a Bayes’ model(Knill & Pouget, 2004). Here, based on Helmholtz’s epistemology, I assume
that safety x is inferred as the cause of observation (data: y) and define the sense of safety as
Bayesian posterior p(x|y) that is proportional to a product of a prior p(x) and likelihood p(y|x).
The prior and the likelihood represent the expectation of safety and the safety based solely on
the data (observed safety), respectively. In this model, the prior (safety belief) is updated to the
posterior (perception of safety).
The difference between the posterior mean and the peak of likelihood represents the gap
between perceived and observed safety. This gap is regarded as a cognitive bias termed
expectation effect (Yanagisawa, 2016). In our previous study, we modelled the expectation
effect as a function of three parameters: a prediction error (difference between prior mean and
peak of likelihood), prior precision (inverse variance of the prior), and observation precision
(inverse variance of the likelihood). We found that there are two types: assimilation and contrast.
Assimilation diminishes the prediction error, and contrast exaggerates the prediction error. The
expectation effect explains several psychological biases that lead to abnormal behaviours, such
as normalcy bias, excessive anxiety, and overconfidence. Based on the expectation effect model,
I discuss the condition of each psychological bias in safety. In addition, I discuss how to model
emotions such as anxious and fear in a mathematical manner by applying information theoretic
quantities such as surprisal and free energy (Yanagisawa, Kawamata, & Ueda, 2019).
REFERENCES
Knill, D. C., & Pouget, A. (2004). The Bayesian brain: the role of uncertainty in neural coding and
computation. Trends in Neurosciences, 27(12), 712-719. doi:10.1016/j.tins.2004.10.007
Yanagisawa, H. (2016). A computational model of perceptual expectation effect based on neural coding
principles. Journal of Sensory Studies, 31(5), 430-439. doi:10.1111/joss.12233
Yanagisawa, H., Kawamata, O., & Ueda, K. (2019). Modeling Emotions Associated With Novelty at
Variable Uncertainty Levels: A Bayesian Approach. Frontiers in Computational Neuroscience,
13(2). doi:10.3389/fncom.2019.00002
ResearchGate has not been able to resolve any citations for this publication.
Modeling Emotions Associated With Novelty at Variable Uncertainty Levels: A Bayesian Approach
Article
Full-text available
  • Jan 2019
Acceptance of novelty depends on the receiver's emotional state. This paper proposes a novel mathematical model for predicting emotions elicited by the novelty of an event under different conditions. It models two emotion dimensions, arousal and valence, and considers different uncertainty levels. A state transition from before experiencing an event to afterwards is assumed, and a Bayesian model estimates a posterior distribution as being proportional to the product of a prior distribution and a likelihood function. Our model uses Kullback-Leibler divergence of the posterior from the prior, which we termed information gain, to represent arousal levels because it corresponds to surprise, a high-arousal emotion, upon experiencing a novel event. Based on Berlyne's hedonic function, we formalized valence as a summation of reward and aversion systems that are modeled as sigmoid functions of information gain. We derived information gain as a function of prediction errors (i.e., differences between the mean of the posterior and the peak likelihood), uncertainty (i.e., variance of the prior that is proportional to prior entropy), and noise (i.e., variance of the likelihood function). This functional model predicted an interaction effect of prediction errors and uncertainty on information gain, which we termed the arousal crossover effect. This effect means that the greater the uncertainty, the greater the information gain for a small prediction error. However, for large prediction errors, greater uncertainty means a smaller information gain. To verify this effect, we conducted an experiment with participants who watched short videos in which different percussion instruments were played. We varied uncertainty levels by using familiar and unfamiliar instruments, and we varied prediction error magnitudes by including congruent or incongruent percussive sounds in the videos. Event-related potential P300 amplitudes and subjective reports of surprise in response to the percussive sounds were used as measures of arousal levels, and the findings supported the hypothesized arousal crossover effect. The concordance between our model's predictions and our experimental results suggests that Bayesian information gain can be decomposed into uncertainty and prediction errors and is a valid measure of emotional arousal. Our model's predictions of arousal may help identify positively accepted novelty.
View
A computational model of perceptual expectation effect based on neural coding principles
Article
Full-text available
Abstract: Prior expectation affects posterior perceptual experience. This contextual bias is called expectation effect. Previous studies have observed two different patterns of expectation effect: contrast and assimilation. Contrast magnifies the perceived incongruity, and assimilation diminishes the incongruity. This study proposes a computational model that explains the conditions of contrast and assimilation based on neural coding principles. This model proposed that prediction error, uncertainty, and external noise affected the expectation effect. Computer simulations with the model show that the pattern of expectation effect shifted from assimilation to contrast as the prediction error increased, uncertainty decreased the extent of the expectation effect, and external noise increased the assimilation. We conducted an experiment on the size–weight illusion (SWI) as a case of the cross-modal expectation effect and discussed correspondence with the simulation. We discovered conditions where the participants perceived bigger object to be heavier than smaller one, which contradicts to conventional SWI. Practical applications: Expectation effect in sensory perception represents a perceptual bias caused by prior expectation, such as illusions and cross-modality. The computational model proposed in this study guides researchers and practitioners who investigate this bias in sensory studies to set a hypothesis with appropriate experimental factors. For example, the model suggests that prediction error can be used as a main factor to identify a condition at which assimilation switches over to contrast. The model provided how expectation uncertainty and noise of stimulus affect the switchover point of prediction error and extent of expectation effect. Uncertainty, which may differ from person to person, can be used as a factor to explain personal differences in the extent of expectation effect.
View
Knill, D.C. & Pouget, A. The Bayesian brain: the role of uncertainty in neural coding and computation. Trends Neurosci. 27, 712-719
Article
To use sensory information efficiently to make judgments and guide action in the world, the brain must represent and use information about uncertainty in its computations for perception and action. Bayesian methods have proven successful in building computational theories for perception and sensorimotor control, and psychophysics is providing a growing body of evidence that human perceptual computations are "Bayes' optimal". This leads to the "Bayesian coding hypothesis": that the brain represents sensory information probabilistically, in the form of probability distributions. Several computational schemes have recently been proposed for how this might be achieved in populations of neurons. Neurophysiological data on the hypothesis, however, is almost non-existent. A major challenge for neuroscientists is to test these ideas experimentally, and so determine whether and how neurons code information about sensory uncertainty.
View