How much to trust the senses: Likelihood learning

Abstract

Our brain often needs to estimate unknown variables from imperfect information. Our knowledge about the statistical distributions of quantities in our environment (called priors) and currently available information from sensory inputs (called likelihood) are the basis of all Bayesian models of perception and action. While we know that priors are learned, most studies of prior-likelihood integration simply assume that subjects know about the likelihood. However, as the quality of sensory inputs change over time, we also need to learn about new likelihoods. Here, we show that human subjects readily learn the distribution of visual cues (likelihood function) in a way that can be predicted by models of statistically optimal learning. Using a likelihood that depended on color context, we found that a learned likelihood generalized to new priors. Thus, we conclude that subjects learn about likelihood.

Bayesian transfer in a complex spatial localisation task

“…Previous studies have also found that observers give more weight 486 to the sensory cue than is optimal (e.g., Bejjanki et al, 2016); even so, the level of 487 sub-optimality that we observe here is still drastically higher, compared to previous 488 reports. However, Sato and Kording (2014) found better, near-optimal performance 489 in those participants who were told that the sensory information can have one of two 490 levels of variance, and that the variance will sometimes change, compared to those 491 who were not provided with this information. We therefore reasoned that if observers 492 are given additional information about the structure and statistics of the task (e.g., 493…”

“…To our knowledge, only one study has tested Bayesian transfer in the context 134 of sensorimotor learning. Sato and Kording (2014) tested the ability of participants to 135 generalise a newly learnt prior to a previously learnt likelihood. In their task, Sato and 136 Kording (2014) first trained participants to complete the task when only a single 137…”

The Difficulty of Effectively Using Allocentric Prior Information in a Spatial Recall Task

“…Part of this is the use of prior information about the environment. Many studies have found that people rapidly begin to exploit prior information in simple perceptual-motor tasks -for example, when judging the trajectories or speeds of items on a screen [3][4][5] or the locations of hidden objects [6][7][8][9][10][11][12][13][14] . Over a longer timeline, this may also explain prototype effects in spatial recall 15 .…”

“…Of the two, allocentric information is more difficult and noisy but more durable and flexible [48][49][50][51] ; purely egocentric information immediately requires updating or discarding whenever the viewpoint changes, but allocentric information can be used flexibly from any viewpoint. There have already been multiple studies suggesting that prior information in an egocentric frame can be used effectively after some exposure to a novel environment [6][7][8][9][10][11][12][13][14] . It is not yet known if the same is true for allocentric information.…”

An Adaptive Cue Selection Model of Allocentric Spatial Reorientation

“…This further applies in a navigation task where the two cues are vestibular and proprioceptive (Frissen, 2011). Adults also rapidly learn egocentric (sensorimotor) prior distributions and use them in a Bayesian fashion as well (Bejjanki, Knill, & Aslin, 2016;Berniker, Voss, & Kording, 2010;Chambers et al, 2018;Körding & Wolpert, 2004;Kwon & Knill, 2013;Narain, van Beers, Smeets, & Brenner, 2013;Sato & Kording, 2014;Tassinari, Hudson, & Landy, 2006). In practice, this means that they learn where targets tend to be and bias their responses towards the places they tend to be most often.…”