bayesian models for motion perception Flashcards
1
Q
Welchman (2008)
A
- when viewing 3D motion, viewers overestimate angular trajectories of moving object so they report that object will miss them when it is actually on a collision course with their head
- motion-in-depth is estimated to be slower than equivalent lateral motion (differentially reliable), which explains biased estimates of trajectory when these signals are contrasted
- biased reports of 3D motion due to brain’s use of prior probabilities which favour slow real-world motion: prior has differential effect on lateral motion and motion-in-depth as they differ in reliability
- so motion-in-depth / approaching objects rely more on prior assumptions and are perceived as slower
2
Q
Champion (2007)
A
- perceived direction of moving plaids w/unequal component contrasts biased towards component w/highest contrast
- higher speed = increased bias to higher contrast component
- except at very low speeds!
- speed + contrast interaction
- inconsistent with bayesian model of speed perception
3
Q
Freeman (2010)
A
- smooth pursuit eye movement illusions: pursued objects appear slower (Aubert-Fleishl illusion), stationary objects appear to move (Filehne illusion) etc
- presence or absence of eye movement is paramount in driving differences in precision of estimates
- extra-retinal info (proprioception, efference copies of eye speed etc) is less reliable than retinal info, so more influenced by prior assumption that movement is zero
- so speed of pursued stimuli perceived as slower
- so two bayes estimates needed: one for pursued stimuli and one for fixated stimuli
4
Q
Weiss (2002)
A
- Bayesian model states visual system predicts near zero net movement in environment, with the influence of this prior expectation depending on reliability of sensory information
- thin, low contrast rhombus appears to have diagonal motion
- less reliable so posterior (perception) relies more heavily on prior assumption of no movement
5
Q
Hammett (2007)
A
- perception of speed can be distorted to look faster or slower depending on luminance, contrast, temporal frequency etc
- bayesian model can explain why perceived speed is often reduced at low contrast and why speed in peripheral vision may be lower than central vision
- but prior cannot explain circumstances where perceived speed is increased in conditions of reduced info about speed, e.g. at low contrast, perceived speed is reduced for slow rates of movement, but it can be overestimated at faster rates (Thompson 2006). should be that less speed info = more reliance on prior (zero)
- found that low luminance = increased perceived speed
- but low luminance causing increased speed perception is not consistent with bayesian model that postulates a slow speed prior
- whereas, ratio model of speed encoding does predict data (little effect of luminance below 4 Hz but sharp increase in perceived speed for frequencies above this)
- MT cells’ responses are consistent with increases in perceived speed due to low luminance (Pack 2005)
- so speed perception can be computed without the need for a prior
- also generally argue that bayesian approach only has value when prior is known - can’t make predictions otherwise
6
Q
Hassan + Hammett (2015)
A
- bayesian model does not predict reversals in perceptual bias
- perceived slowing of low contrast patterns is greater at high luminance
- challenges bayesian model as it would predict that low contrast patterns should have same relative perceived speed at either luminance
- ratio model of speed encoding that incorporates known luminance-induced changes in gain (e.g. Purpura 1988) predicts that perceptual bias will be greater at slow speeds at high luminance, and greater at fast speeds at low luminance due to stimulus encoding properties of cells in visual cortex
7
Q
Purpura 1988
A
ratio model of speed encoding that incorporates known luminance-induced changes in gain
8
Q
Stocker + Simoncelli (2006)
A
- bayesian model of speed perception
- does not predict reversal in perceptual bias shown by the Thompson effect at high speeds
- state their model “would be able to fit these behaviours with a prior that increases at high speeds” (but this has no predictive power?)
9
Q
Sotiropoulos (2014)
A
- well-known that perceived speed of moving objects depends on image contrast
- lower contrast = lower perceived speed (thompson effect)
- explained by bayesian model involving slow speed prior
- however, some evidence that thompson effect is attenuated or even reversed at high speeds or low luminances
- found that at high speeds, thompson effect was attenuated (i.e. grating appeared to move faster) by both low AND high contrasts
- bayesian model can’t account for interaction so needs to be modified
- Thompson’s (2006) ratio model fits data better but doesn’t explain low speeds (where speed decreases with contrast)
- therefore incorporated the bayesian model and thompson model (both 2006) to provide better description of data. partially accounts for interaction
- interaction would be better explained by model in which perceived speed decreases with contrast at low speeds but increases with contrast at high speeds
10
Q
Thompson (2006)
A
- ratio coding model
- ratio of low-pass and band-pass temporal filters (functions of temporal frequency and contrast)
- attempts to explain speed-dependent effect of contrast on perceived speed
- nonlinear
- but can’t account for increase in perceived speed with high contrast
- biologically plausible way of achieving variable speed tuning in MT neurons by using a small number of V1 neurons tuned not to speed but to a limited range of temporal and spatial frequencies
11
Q
Gibson
A
perception is entirely bottom-up process
12
Q
Necker cube
A
bistability of percept can’t be driven solely by sensory info, requires top-down explanation