Vincent, Peter; (2024) Bias and variance in the perception and recall of orientation. Doctoral thesis (Ph.D), UCL (University College London).

Preview

Text Vincent_10191693_thesis_sig-removed.pdf Download (16MB) | Preview

Abstract

This thesis studies the computational and theoretical accounts of bias in orientation recall and reconstruction experiments and explores how the choice of stimuli affects the variance of the response distribution. My investigation of the bias and variance falls into two categories - analysis of the response distribution under different experimental conditions, and modelling of representation and decision making in orientation reconstruction. Firstly, I characterise the response distribution and show that, under appropriate assumptions, participants take the mean of their beliefs over the likelihood, without incorporating a pro-cardinal prior. I follow this with an analysis of the variance in responses to understand the impact of task features such as variable working memory delay durations, different stimuli and the relevance of recent responses and distractor stimuli. Contrasting the presence of serial and distractor effects with the observation that the mean of likelihoods explains participant responses suggests that participants accurately marginalise the serial and distractor effects, which I will use to propose why participants do not learn the environmental statistics, and to discuss the relevance of our results to the mechanisms the brain uses to perceive orientation. Finally, I will study models based on the principle of efficient coding, proposing how these models may be generalised without losing their interpretability. I emphasise the failure modes of these models and propose how these may be resolved, arguing for a new class of model that generates responses by matching the stored representation with the representation of the manipulandum. Additionally I will suggest that the representation extends beyond the orientation and uncertainty to include a belief over the cardinality of the stimulus. By jointly considering our results and previous literature on similar tasks, I will challenge the narrative that the bias and the structure of the variance in orientation reconstruction arises from a strictly pro cardinal prior.

Download activity - last month

Export as CSVJSONXML

Download activity - last 12 months

Export as CSVXMLJSON

Downloads by country - last 12 months

Export as CSVXMLJSON

Archive Staff Only

View Item