Social bias in spatial perception: Stochastic Continuous Time Recurrent Neural Network interpretation of a central tendency study

Tsfasman, Maria

Social bias in spatial perception: Stochastic Continuous Time Recurrent Neural Network interpretation of a central tendency study

Files

S1021505_TsfasmanM.pdf (3.53 MB)

Authors

Tsfasman, Maria

Issue Date

2020-07-01

Language

en

Abstract

Human perception is perfect in its adaptability and yet so unreliable in how much it gets altered depending on the situation. We make di erent judgements about the world depending on our experience, the context of the current task, and whether we are alone or surrounded by people. In a recent study Mazzola et al. (2020) asked adults to estimate the lengths of the same set of stimuli in three conditions. In the rst condition, participants had to reproduce lengths of the presented lines on a tablet touch screen. In the second condition, the lengths to reproduce were pointed towards by a robot. In the last condition, the robot not only was showing the stimuli but also behaving in a human-like way: making eye contact, talking, gesticulating. Although the stimuli and the task were the same, the participants were more accurate in the human-like robot condition than in the mechanical robot condition, and generalised towards the mean length of the stimuli the most in the individual tablet condition. This phenomenon can be explained through the theory of Bayesian perception: every time we are about to make a decision about the world we combine our prior experience with the incoming sensory information in a near-optimal way. Depending on the situation and our sensory capabilities the prior information and the direct input weigh more or less in this inference. The hypothesis of Mazzola et al. (2020)'s study is that when interacting with a human (or a more human-like robot) we value our prior experience less than in the individual condition. In order to test this hypothesis and receive insights about the potential di erences between conditions, in this thesis, we trained an arti cial neural network to reproduce Mazzola et al. (2020)'s data. The internal representations of the network did not clearly separate between di erent conditions but when strengthening the network's prior reliance the inter-participant variability increased. This suggests the higher variability between participants from populations with hypothesised stronger prior reliance than in neuro-typical adults - such as younger children and people with schizophrenia. We also demonstrate that the di erences between conditions can be indeed explained by the prior reliance by reproducing the human performance with a network trained on one condition at a time and modifying its prior parameter.