Conflict between short-and long-term experiences affects visual perception by modulating sensory or motor response system: evidence from Bayesian inference models

Huge studies have explored the effects of short-and long-term experiences on visual perception, respectively. However, no study investigated whether and how the conflict between the two types of experiences affected our visual perception. To address this question, we adopted a task of estimating simulated self-motion directions (i.e., headings) from optic flow, in which a long-term experience – straight-forward motion is more often than lateral motion – plays an important role. The long-term experience is learned daily or encoded in our brains from birth. The heading directions in the experiment were selected from three different distributions, generating different conflicts between short-and long-term experiences. The results showed that both estimation errors and sizes of serial dependence of previously seen headings on current heading estimates varied when the experience conflict changed. Finally, we developed two Bayesian inference models, assuming that the experience conflict affected visual perception by influencing the sensory representation’s likelihood distribution or motor decision process. We found that both models captured participants’ estimation errors and serial dependences well in three distributions. In conclusion, the current study revealed the effects of the conflict between short-and long-term experiences on visual perception and preliminarily uncovered that Bayesian inference theory could explain the effects. Moreover, the study implied that the experience conflict affected visual perception by modulating our sensory or motor response systems.


44
Learning the statistical regularity of the living environment is vital for human and animal survival.

45
For example, our ancestors discovered that red tomatoes are ripe and edible during the long 46 evolutionary process. As a result, when shopping for tomatoes, we prefer fresh red tomatoes over green ones. Another example is that a person sees a hole in the road when driving to work. The 48 person will slow down in advance when getting off work and passing by the same place. Hence, 49 experiences have important effects on our daily behaviors.

50
According to the duration that the experience is stored, the experience can be categorized as

63
Note that in the studies above, short-term experience means the statistical regularity of features 64 learned in the past few minutes or hours. To differentiate it, we define long-term experience as the PLOS computational biology Short-and long-term experiences affect visual perception 5 65 statistical regularity of features accumulated daily or in the long evolutionary process rather than 66 learned in several days, which is robust and cannot be changed easily. For example, there are more 67 vertical and horizontal orientations than other directions in the natural world [9,10], and straight-68 forward motion is more often than lateral motion. Studies have demonstrated that these long-term observers' self-motion direction (i.e., heading) estimates are systematically compressed toward the 72 straight-ahead direction, showing center bias [15][16][17][18][19]. These all demonstrate the effects of long-term 73 experience on feature perception.
[20] asked participants to complete a color discrimination task (blue vs. purple).

75
In the initial 200 trials, the blue and purple dots each accounted for 50%. In the last 200 trials, the 76 proportion of blue dots were reduced to 6% in the decreasing condition or held constant in the stable 77 condition. They found that participants tended to report more blue colors in the decreasing condition 78 than in the stable condition. They named the phenomenon the "prevalence-induced concept change 79 (PICC)" effect [21]. With their experimental methods, we know that the color distributions of the 80 initial and last 200 trials were learned by participants within hours. Hence, the experiences are short-81 term, indicating that the PICC effect reflects the effects of conflicts between short-term experiences 82 on perception, which led us to ask whether and how the conflict between short-and long-term 83 experiences affected our perception.

84
Moreover, beyond the perceptual errors mentioned above (e.g., central tendency, oblique effect, 85 center bias, PICC effect), researchers have found that our current perception is affected by a single 86 feature that was previously seen several seconds earlier (within ~15 s), known as serial dependence PLOS computational biology Short-and long-term experiences affect visual perception 6 87 (see [22] for a review). For example, Fischer and Whitney [23] were the first to find that orientation 88 estimates of current trials were systematically biased toward the orientations of previous trials.

105
Specifically, the predicted estimates were scaled based on the level of the experience conflict. We

106
found that the two models predicted participants well, implying that the experience conflict affected 107 visual perception by modulating sensory or motor response systems. The current study improves 108 our understanding of the effects of short-and long-term experiences on visual perception.

114
The yellow "+" indicates the simulated heading direction of the current optic flow. The lines and "+" cm, which aimed to reduce the conflicts between the motion parallax and binocular disparity.

141
Participants' viewing direction (i.e., straight-ahead direction) was aligned with the display center.

142
The current study consisted of three blocks, each completed by one group of participants,

160
The trial procedures of the three blocks were the same. On each trial, an optic flow display was 161 presented for 500 ms, followed by a blank display with a horizontal line appearing across the mid-162 screen of the display. Participants were asked to indicate their perceived heading by moving a 163 mouse-controlled probe. No feedback was given in the experiment.

164
Before the commencement of the experiment, participants were asked to conduct 20 practice 165 trials randomly selected from the experimental distribution to get familiarized with the experiment.

166
The experiment lasted for about 50 min. proportions were reduced in the short-term experience. We, therefore, expected that center biases 181 in the flat, hill, and heavy-tail distributions were stronger than in the peak distribution.

182
Next, we tested how the conflict between short-and long-term experiences affected the serial

199
Note that the serial dependence analysis above did not remove the heading error induced by fitted the RHE as a 1 st -derivative Gaussian function of RH (see [23]), given as: in which was the predicted residual heading error, was the two amplitude of curve 208 peaks, was the curve width.

209
For the data in the peak distribution, we fitted the RHE as a linear function of the RH (see

210
[17]), give as: of the distribution was smaller than 0, then or _AH was significantly smaller than 0, 220 indicating a repulsive serial dependence in heading perception.

221
The relative heading (RH) in the above analysis was the difference in the actual heading 222 between the previous and current trials, revealing a serial dependence in the sensory representation 223 stage (i.e., perceptual stage, [23,33,34]). Some previous studies have also pointed out the motor 224 decision process [35]. Next, to examine whether the motor decision process played some role in our 225 serial dependence, we replaced the actual heading of previous trials with their perceived heading.

226
The difference was named perceived relative heading (PRH). We then replaced the RH in Equations

251
Taking the peak distribution as the baseline (blue dots), Figure 2a clearly shows that when the 252 proportions of peripheral headings in the hill and peak distributions (black and red dots, Figure 1b) 253 increase, the slopes gradually decrease, suggesting that the sizes of center bias increase. One-way

326
Specifically, when there were more peripheral headings in the short-term experience than in 327 the long-term experience, observers tended to bias their estimates away from these headings,

328
leading to an increase in center bias in the hill and flat distributions (Figure 2). This finding is process, meaning that the heading estimate is a result of an optimal combination between prior and 345 likelihood distribution (gray shaded area in Figure 6, [17][18][19]). In the current study, we developed 346 two Bayesian inference models based on two assumptions to explain the effects of the conflict 347 between short-and long-term experiences on heading perception from two perspectives.

Model 1 Conflict affects likelihood distributions 363
As mentioned above, the classical Bayesian inference process is an optimal combination of prior 364 ( ( )) and likelihood ( ( | )) distributions (gray shaded area in Figure 6):

368
In the current study, the prior ( ( )) was the long-term experience, following a Gaussian 369 distribution with the straight-ahead direction (0°) serving as the distribution center, given by: in which is a free parameter, indicating the standard deviation of the prior distribution. The 372 smaller the is, the narrower the prior distribution is and the more certainty the prior is. In the 373 model, we assumed that the prior distribution was robust and was not affected by the short-term 374 experience.

PLOS computational biology
Short-and long-term experiences affect visual perception 21 375 In addition, the likelihood ( ( | )) is also a Gaussian distribution with the actual heading of 376 current trial ( ) serving as the distribution center, given by: in which is another free parameter, indicating the standard deviation of the likelihood 379 distribution. Like the , the smaller the is, the narrower the likelihood distribution is and the 380 more certainty the sensory representation of the current heading is.

381
Previous studies have found that when observers are often exposed to one feature (seconds to 382 minutes), neurons of sensory systems selectively responding to the feature would become fatigued.

383
Neurons' firing rates are significantly reduced [36,37]. Many studies have proposed that the

394
As mentioned in the data analysis part of Experiment, the peak distribution severed as the baseline 395 because we proposed that the peak distribution was most closely to our long-term experience.

396
Therefore, we first built a classic Bayesian inference model (gray shaded area in Figure 6 given by: 404 in which 0 is the standard deviation when the actual heading (AH) is 0°; is the increasing 405 factor and must be positive. Therefore, the model included three parameters: , 0 , and .

406
Markov chain Monte Carlo (MCMC) sampling was used to estimate the parameters. We used Next, we repeated the above procedure for the data in the hill and flat distribution. But, the 420 of prior ( ( )) was fixed and from the model of the peak distribution. Therefore, the models in the 421 hill and flat distributions only contained two parameters: 0 , and (red arrows in Figure 6).

422
Finally, we used the ideal models to simulate 12 participants' estimates in each actual heading

461
Together, the patterns of our Model 1's simulated results were similar to participants' patterns 462 (Figures 3 and 4), suggesting that the conflict between short-and long-term experiences may 463 modulate the certainty of our sensory systems' representation to affect heading perception from 464 optic flow.

466
Model 1 assumed that different performances in different distributions were due to the change of 467 likelihood distributions. Some researchers could argue another explanation. In daily life, straight-468 forward motion is more often than lateral motion, leading observers to be reluctant to report the 469 peripheral headings. The conflict between the short-and long-term experiences increased the sense 470 of reluctance, leading to participants bias their estimates toward the center of the prior distribution.

471
That is, the conflict between the short-and long-term experiences affected participants' motor 472 decision system rather than the likelihood distribution.

474
Like Model 1, we also took the peak distribution as the baseline. The predicted performance can be PLOS computational biology Short-and long-term experiences affect visual perception 26 475 given by Equation 4. The modeling procedure was the same as Model 1.

476
After getting the model for the peak distribution, we multiplied a response scaling factor ( )

477
with the posterior estimates ( ( | )), as indicated by the blue arrows in Figure 6. Response scaling 478 factor ( ) can be given by: 480 in which 0 is the standard deviation when the actual heading (AH) is 0°; is the increasing 481 factor and must be positive. Therefore, Model 2 contained two parameters: 0 and .Similar

482
MCMC procedures were adopted to get the response scaling factors in the hill and flat distributions.

483
Lastly, similar procedures were adopted to predict the center bias and serial dependence in the 484 flat and hill distributions (as see the last paragraph in the methods of Model 1).

Results and discussion
486 Figure 8 shows the simulated results of our Model 2. It clearly shows that the center bias increases 487 with increasing the proportion of peripheral headings (peak to hill to flat distributions); before 488 removing center bias, the serial dependence was attractive in three distributions; and after removing 489 center bias, a repulsive serial dependence was in the peak distribution and a very weak attractive

517
General discussion 518 The current study adopted one behavioral experiment and two Bayesian inference models to 519 systematically investigated whether and how the conflict between short-and long-term experiences

603
Moreover, our models explained the potential mechanisms underlying the effects of the 604 experience conflict on heading perception from two perspectives. One possibility is that the 605 experience conflict increased the fatigue of sensory systems, making the likelihood distributions 606 uncertain (red arrows in Figure 6). Hence, participants relied more on the prior distribution and 607 previously present heading, showing stronger center bias and attractive serial dependence, 608 consistent with participants' performance ( Figure 2a). Another possibility is that the experience 609 conflict changed participants' motor systems. Specifically, when there were more peripheral 610 headings in the short-term experiences, participants might avoid reporting peripheral headings and 611 shrink their response ranges. Therefore, our Model 2 worked like the classical Bayesian model (gray 612 shaded area in Figure 6) but multiplied an extra response scaling factor (Equation 7) with the 613 decoded estimates. Model 2 also captured the participants' performances, suggesting that the 614 experience conflict affected heading perception by influencing motor decision systems.

615
Therefore, we developed two Bayesian inference models to explain our findings. Some  626 We know that the size order of s is peak < hill < flat distributions. Hence, with Equation 9, in

638
Future studies 639 The current study proposed that the effects of the conflict between short-and long-term