【Day38 文献精读】A Bayesian Perceptual Model Replicates the Cutaneous Rabbit and Other Tactile ..

阅读文献：

Goldreich, D. (2007). A Bayesian perceptual model replicates the cutaneous rabbit and other tactile spatiotemporal illusions. plos one, 2(3), e333-e333. 

文献链接：

A Bayesian Perceptual Model Replicates the Cutaneous Rabbit and Other Tactile Spatiotemporal Illusions

目录

• Abstract
• 1. Introduction
• 2. Results
• 1) Basic Bayesian Observer
• 2) Lambda Variation
• 3) Temporal Order Judgment
• 4) Spatial Attention
• 5) Full Bayesian Observer
• 6）Perceived Velocity
• 3. Discussion

---

Abstract

1) Background: As for rapid sucessive tactile stimuli, the origins of perceptual spatiotemporal distortions (spatial intervals contraction while temporal intervals dilation) are unknown.

2) Methodology: The Bayesian observer replicates the cutaneous rabbit illusion and spatiotemporal illusions.

3) Conclusions: brain automatically incorporates prior expectation for speed in order to overcome spatial and temporal imprecision inherent in the sensorineural signal.

---

1. Introduction

1) Research aim: Explain the perceptual distortions:

① Cutaneous rabbit: a sequence of three or more taps to two skin sites evokes the perception of an object hopping along the skin from the first site to the second, landing in the process on intervening skin that was never touched (Fig. 1A).

② Perceptual length contraction: underestimation of inter-stimulus distance (ISD) (Fig. 1A-
E).

③ Perceptual time dilation: overestimation of inter-stimulus time (IST) (Fig. 1F).

2) The Bayesian observer model described here replicates the spatiotemporal illusions by two assumptions:

① Stimuli separated by small spatial and temporal intervals originate from uniform object motion;

② Objects that contact the skin tend to move slowly.

---

2. Results

1) Basic Bayesian Observer

① Admiting spatial but not temporal imprecision (Fig. 2), this model experiences length contraction but not time dilation.
② Perceived ISD (by posterior) underestimates actual ISD (by maximal likelihood), and perceived velocity underestimates actual velocity(Fig. 2C).

Equations

Define velocity as (slope) m, and first stimulus position as (y-intercept) b in Fig. 2

① prior: trajectory’s prior probability (independent of b)

② likelihood: the probability of the stimulus-evoked neural data given the trajectory

③ posterior:  posterior probability of the candidate trajectory given stimulus-evoked neural data D

④ observer’s perceived velocity, v’ ((the value of m at the mode of the posterior), can be got by setting to zero the partial derivatives of the exponent of Equation 9 with repect to m and b.

Equation 3 predicts that 1) perceived velocity underestimates real velocity, v = l/t. 2) real and perceived velocities will relate non-monotonically when IST is reduced at fixed ISD.
 

⑤  observer’s perceived ISD, l'=v't

 

Equation 1 predicts that perceived ISD will: 1) underestimate actual ISD; 2) asymptotically approach actual ISD as IST increases; and 3) increase linearly with actual ISD, at constant
IST. 4) decrease as σs (spatial uncertainty) increases.

→ Each of these predictions is borne out by the human perceptual data in Fig. 3 A-E

2) Lambda Variation

① A small λ results from strong expectation for slow movement (small σv) and/or poor spatial acuity (large σs), either of which facilitates perceptual length contraction (Equation 1).

② The value of λ varies from one body region to another

→ From Fig. 3F, a linear relation between point localization accuracy and 1/λ suggests that λ variation is caused by variation in σs, and the low-velocity prior, σv, is conserved from one body region to another.

3) Temporal Order Judgment

These points of concordance between human and model TOJ (temporal order judgment) performance suggest that the brain indeed integrates across the full posterior probability distribution.

4) Spatial Attention

If the observer pay more spatial attention to location of the second stimulus:

① From Fig. 5B, the distribution of likilihood became oval-shaped comparing to Fig. 2B→The spatial uncertainty of location of the second stimulus is reduced.

② From Fig. 5C, The midpoint of the perceived trajectory (dashed line) thus shifts towards (higher than the midpoint of the actual trajectory drawed by solid line) the attended location.

5) Full Bayesian Observer

① The kappa effect is reproduced by the full Bayesian observer model, in which tactile sensation suffers from temporal as well as spatial uncertainty (Fig. 6A)

② The full observer experiences perceptual time dilation as well as length contraction (Fig. 6B).

③ Perceptual time (t') as ISD increases at fixed IST (Fig. 6C)→ kappa effect

④ Only where spatial acuity is relatively good (e.g. fingertip; small σs) does time dilation necessarily play a greater role.

⑤ The length contraction equation for the full observer is:

Equation 2, unlike Equation 1, predicts a nonlinear relationship between perceived and real ISD (Fig. 7E, crosses).

6）Perceived Velocity

in basic model, perceived velocity, v':

v' peaks at real velocity, v*, given by

The difference between basic and full models is most pronounced where tactile acuity is greatest
(e.g., the fingertip, lower in Fig. 9A&B).

---

3. Discussion

1) Several theories for explaining the kappa effect and tau effect

Collyer: constant velocity account

Jones and Huang: weighted average model

Brigner: rotation of a perceptual space-time coordinate frame

2) The Bayesian observer model described here provides a coherent explanation for these illusions, suggest that the brain takes advantage of the expectation for slow speed to improve
perception beyond the limits imposed by spatial and temporal uncertainty inherent in the sensorineural signal.

3) Important work related to the model remains to be done.