Unlike the other two, chapter four aims to manipulate binary choice outcomes through changing the visual saliency distribution under SAM. Experimental results show that visual salience can increase choice correction rates when the outcome is more rewarding.
INTRODUCTION
- General introduction
- The concept of salience
- Bottom-up and top-down salience
- A computer vision approach for visual saliency modeling
- The attentional Drift-Diffusion-Model (aDDM)
- The direction of this dissertation
Extensive work in psychology and neuroscience contributes to our understanding of the visual attention system. It is a particular application of the general drift diffusion model (Ratcliff and McKoon, 2008; Ratcliff, Smith, et al., 2016) in the realm of perceptual decision making.
SALIENCY IN EXPERIMENTAL GAMES
Introduction
Schelling reported that seven of the eight people he asked chose to meet at the bridge. The algorithm predicts that the bridge is prominent because it has the most dark/light contrast and is close to the center of the map.
Background on focality in games
In SCH, “the process that brings one of the labels to the player's mind”—its salience—is predicted ex ante by the SAM model. Our replications of the Caltech and UCLA subjects found much lower rates of third choice.
The Saliency Attentive Model (SAM) algorithm
In psychology, attention-grabbing characteristics are divided into two categories: bottom-up and top-down. Bottom-up features are based on sensory physical features such as color, contrast, and orientation.
Experimental procedure
The SAM salience map, in which the brightness indicates salience level. the lightest point is the most noticeable spot). This area is generated from the ranking of all saliency values of each pixel. d): the original image with the saliency heatmap overlaid ("warmer" red colors indicate higher saliency).
Analysis and Results
An intuitive hypothesis is that the degree of matching in an image should be affected by the spread of saliencies in the image. In Figure 2.7d, the peak of the concealer selection distribution no longer falls in the most prominent area.
A Saliency-perturbed Cognitive Hierarchy Model (SCH)
Except for level zero players, all other levels of players behave the same as in CHC. Training data is shown in the upper figure 2.8ab and test data in the lower figure 2.8cd. In the hider case (b), level 0 players choose strategies that increase with salience, and all higher-level types choose less salient locations more often.
Quantal response equilibrium on saliency
When 𝜆 is greater than the cutoff, the probability distribution of the hider's strategy over all actions monotonically decreases as saliency increases. QRE further predicts that searchers' advantage occurs only when 𝜆 is small, which also means that subjects choose more randomly than trying to win the game. If they want to win the game and play equilibrium strategies, we should have observed an underdog's advantage, but we never did.
Another two applications of SAM in economics
When 𝜆 is less than a cut-off threshold, hiders also prefer saliency that the probability distribution over all actions monotonically increases as saliency increases. Remember the realized strategy we plot in Figure 2.8, where hider's strategy has a u-shaped property that it first falls and then sharply increases. We can compare these actual gaze maps with the SAM predictions.22 Figure 2.12 shows an example and summary statistics.
Visual saliency in investment experiments
For example, in Figure 2.13(b), the SAM salience is highest in the center of the price time series, where there is a large positive return. The figure shows two examples of ground truth density maps from our experiment on the left and heat maps predicted by SAM where we also show participants of the underlying price paths. The authors find that about 20% of the variance in salience values can be explained by such statistics.
Conclusion and discussion
Their use of the term salience almost always refers to hidden (and unmeasured) variation in likely salience inferred from observable choices (e.g., Level-k auctions: Can a non-equilibrium model of strategic thinking avoid the winner's curse and outbidding in private value do auctions explain?” In:Econometrica75(6), pp. Subjective patterns of randomness and choice: some consequences of collective responses.” In:Journal of Experimental Psychology: Human Perception and Performance35(1), p.
APPENDIX
2.A Results from no-feedback trials
The x-axis is the range of salient values and the y-axis is the probability density. However, there are a disproportionate number of high importance site choices (ie, the density increases sharply at the right end of the scale). There is a slightly disproportionate tendency to select the lowest salient locations (near zero at the left end of the scale), especially for concealers.
2.B Model comparison
Nor is it the case that zero-level behavior is the only source driving fit. The Level-k 6 model is almost equally accurate in terms of AIC and BIC, so we have commented on what we can learn from it in the text. Each point on the graph indicated what percentage of choices were made for locations within images based on the importance of those locations. selection data and model prediction in the training data set search condition.
2.C Details of SCH model fitting and SCH converging properties
The left figure simulates search win rates with the best fit𝜇and𝜆 relative to the different ones. As can be seen, the salience bias will only affect the search win rate when the level is low. The search profit rate will converge around the equilibrium level of 0.05 when the strategic level is high.
2.D Quantal response equilibrium prediction on saliency
Although profit-seeking converges, higher-order behavior does not actually converge, meaning that the convergence in profit-seeking is actually due to the mixing of more types. In the first two equations, the denominator will not change with n, so they can be considered constant for studying seeker advantage. In an extreme case, when 𝜆 approaches zero, people will choose salient locations, which will generate an advantage for the seeker due to the correlation between two roles.
MODELING CHOICE-PROCESS DATA ON STRATEGIC GAMES – AN APPLICATION OF HIDDEN MARKOV MODEL
HMM)
- Introduction
- Discrete Model (HMM)
- An application of the Gaussian hidden Markov model (gHMM)
- A novel way for level identification
- Extended continuous-time model (ctHMM)
- Prediction of time-pressure condition in games
- Discussion
The order of fixations is indicated by discrete numbers (the first fixation is marked with "1", the second with "2", etc.). A sample prediction looks like: at time 2s, with probability 0.7 the decision maker is considering the hidden strategy 𝑠𝑖 and with. An output, in ctgHMM is the salient value of the fixation location in the discrete model.
3.A Eye tracking procedure
3.B HMM model estimation
At any moment𝑡, the particles (decision makers) are in the high state, low states, or those who have already left the game (i.e., are in the decision state). For those who enter the decision state, we take the high or low states experienced before the decision state as the decision score. When we add an exogenous time pressure to time 𝑡, the result of the simulation will be a mixture of people who naturally finished their thinking by 𝑡 (those who reached the decision state before time 𝑡) and people who did not reach the decision state D in time.
MANIPULATING VISUAL SALIENCY IN SIMPLE CHOICE PROBLEM
Introduction
The posterior way provides a better interpretation of the information aspect where the information costs lie. arises from forming the attention strategy.1. Since such a signaling process is not optimal, after forming the salience-induced posterior, the remainder of the salience-refined rational inattention model must be solved in the same way as before, replacing the prior probability of being in each state with a different posterior probability than priors. We show that the result of the experiment fits well with the prediction of the srRI model, with people making more errors when the most salient item is the less rewarding item.
Saliency in Rational Inattention Model
What we can learn from the result of Equation 4.6 is that salience affects the decision by distorting people's underlying beliefs about different payoff states. The bending curves in Figure 4.2.1 showed how the rate of correction changes as the posterior belief about being in state 1 changes. The correction rate will not be affected much by the posterior belief if the value is large enough to make the correct choice or the learning cost is small.
Experiment Design and Data
Such salience biases will increase correct rates when the salient item is also the rewarding item and impair correction rates when they are opposite. Subjects first experienced an introductory session in which the basic tasks were explained, followed by a test session that asked questions about the rules. They then experienced an unlimited time session with N=5 images that will count towards payment, but only for training purposes.
Set-up
Balanced locations of saliency centers: All selected images have saliency centers that are evenly located on the left or right side. Half of the images have a salience center on the left, and half on the right. Eleven images have more fruit on the left and nine images have more fruit on the right.
Results
Models (3) and (4) regress the correctness of a choice on the absolute value difference between two alternatives, congruence property, and controls. Model (1) regressive response time, (2) regressive normalized response time9, both on the absolute value difference between two stacks (abs(ValueDiff)), the absolute difference in fruit numbers (abs(ItemsMore)), whether it is congruent (Congruence), interaction between value difference and congruence, and controls. Such an effect becomes weaker as the value difference increases (positive interaction coefficient and negative congruence coefficient).
Results in a more complicated environment
Conclusion and discussion
The result fits well with selection theory, which is currently trending in cognitive science (Failing and Theeuwes, 2018; Awh, Belopolsky and Theeuwes, 2012), questioning the traditional dichotomy approach by introducing 'bottom-up' and 'top -down' to separate. ” (Itti and Koch, 2001; Frintrop, Rome and Christensen, 2010; Veale, Hafed and Yoshida, 2017). This task in itself will be a suitable paradigm to test the top-down task relevance and compare the result with the bottom-up driven salience. For example, the real attention patterns may migrate from the bottom-up mode to the search mode, where subjects search “top-down” for the most valuable item in each trial.
DISCUSSION AND FUTURE DIRECTIONS
Conclusion
Limitation
Future directions
That is why we often see that the information about the former case is written in the focal place while the information about the latter is often hidden in obscured places.
BIBLIOGRAPHY
