Learning priors for Bayesian perception in normal and atypical development
Over recent years, biases of sensory magnitude estimations, such as central tendency bias have been thought to reveal the weighting of a Bayesian prior as estimations drifted towards the mean of the stimulus distribution. In temporal estimation specifically, Vierordt's law described the unde...
| Main Author: | |
|---|---|
| Format: | Thesis (University of Nottingham only) |
| Language: | English |
| Published: |
2024
|
| Subjects: | |
| Online Access: | https://eprints.nottingham.ac.uk/78381/ |
| _version_ | 1848801076014219264 |
|---|---|
| author | McKeown, Lucy |
| author_facet | McKeown, Lucy |
| author_sort | McKeown, Lucy |
| building | Nottingham Research Data Repository |
| collection | Online Access |
| description | Over recent years, biases of sensory magnitude estimations, such as central tendency bias have been thought to reveal the weighting of a Bayesian prior as estimations drifted towards the mean of the stimulus distribution.
In temporal estimation specifically, Vierordt's law described the underestimation of long durations and the over estimation of short durations. Vierordt's law has since been interpreted via a Bayesian explanation by Jazayeri and Shadlen (2010). This suggested that statistical regularities were learned and incorporated into a prior, but it did not provide insight into how priors develop over time.
Priors may update through the dynamic estimation of the mean of the stimulus distribution. Or, as suggested more recently, that the prior is updated according to the value of the recent stimulus using a Kalman filter. Glasauer and Shi (2019; 2021b) demonstrated the possible use of the Kalman filter in participants because central tendency effects were reduced for the random walk (gradually changing magnitude) stimuli relative to fully random stimuli. However, the current research suggested that, during a temporal reproduction task, models that updated using the mean of the stimulus distribution provided a much better fit to the data than a simple iterative (Kalman filter) model.
The extent of the central tendency bias in temporal reproduction was also correlated with AQ trait scores and showed a relationship between higher AQ trait scores and lower compression for random walk sequences. Therefore, higher levels of autistic traits may be associated with differences in prior acquisition but a tendency for a static or iterative model according to the number of autistic traits was not found.
Temporal perception is necessary for learning when sensory stimuli appear and how long they persist, in addition to the functional benefits of multisensory integration. However, the possibility that spatial quantities may be more likely to induce iterative strategies than temporal quantities was exemplified by research into path integration (Glasauer & Shi, 2021b; Petzschner & Glasauer, 2011) and visually guided reaching (Verstynen & Sabes, 2011). Using spatial localisation (the prediction of target locations from a cue location), the current research was the first demonstration of flexibility in prior acquisition strategy in line with stimulus features. This was only found in adult participants however, as older children did not show this same level of flexibility and the younger children displayed a type of flexibility that may be a reactive response to the stimuli.
The exploration of bias in a single stimulus has been the focus of most research using the Bayesian observer model. However, some Bayesian model research used the time order error, as a metric for understanding different priors in autistic individuals (Sapey Triomphe et al., 2021). The time order error (TOE) occurs when the first stimulus is perceived with relatively more bias (over and underestimation) than the second stimulus. In the current research, a Bayesian model accounted for a TOE outside the context of a two-stimulus comparison, which was not explained by previous models and may imply that passive memory processes provide a sufficient explanation without the need for processes rooted in interference.
Taken together, these results show that Bayesian models provided a relatively good fit for a range of different prior acquisition strategies across different temporal and spatial tasks. Although the current research found a preference for static models in temporal research and some preference for a previous trial model in spatial research. A degree of flexibility, not yet seen in the literature, was displayed in the spatial task regarding participants prior updating strategy. |
| first_indexed | 2025-11-14T21:01:42Z |
| format | Thesis (University of Nottingham only) |
| id | nottingham-78381 |
| institution | University of Nottingham Malaysia Campus |
| institution_category | Local University |
| language | English |
| last_indexed | 2025-11-14T21:01:42Z |
| publishDate | 2024 |
| recordtype | eprints |
| repository_type | Digital Repository |
| spelling | nottingham-783812024-07-23T04:40:38Z https://eprints.nottingham.ac.uk/78381/ Learning priors for Bayesian perception in normal and atypical development McKeown, Lucy Over recent years, biases of sensory magnitude estimations, such as central tendency bias have been thought to reveal the weighting of a Bayesian prior as estimations drifted towards the mean of the stimulus distribution. In temporal estimation specifically, Vierordt's law described the underestimation of long durations and the over estimation of short durations. Vierordt's law has since been interpreted via a Bayesian explanation by Jazayeri and Shadlen (2010). This suggested that statistical regularities were learned and incorporated into a prior, but it did not provide insight into how priors develop over time. Priors may update through the dynamic estimation of the mean of the stimulus distribution. Or, as suggested more recently, that the prior is updated according to the value of the recent stimulus using a Kalman filter. Glasauer and Shi (2019; 2021b) demonstrated the possible use of the Kalman filter in participants because central tendency effects were reduced for the random walk (gradually changing magnitude) stimuli relative to fully random stimuli. However, the current research suggested that, during a temporal reproduction task, models that updated using the mean of the stimulus distribution provided a much better fit to the data than a simple iterative (Kalman filter) model. The extent of the central tendency bias in temporal reproduction was also correlated with AQ trait scores and showed a relationship between higher AQ trait scores and lower compression for random walk sequences. Therefore, higher levels of autistic traits may be associated with differences in prior acquisition but a tendency for a static or iterative model according to the number of autistic traits was not found. Temporal perception is necessary for learning when sensory stimuli appear and how long they persist, in addition to the functional benefits of multisensory integration. However, the possibility that spatial quantities may be more likely to induce iterative strategies than temporal quantities was exemplified by research into path integration (Glasauer & Shi, 2021b; Petzschner & Glasauer, 2011) and visually guided reaching (Verstynen & Sabes, 2011). Using spatial localisation (the prediction of target locations from a cue location), the current research was the first demonstration of flexibility in prior acquisition strategy in line with stimulus features. This was only found in adult participants however, as older children did not show this same level of flexibility and the younger children displayed a type of flexibility that may be a reactive response to the stimuli. The exploration of bias in a single stimulus has been the focus of most research using the Bayesian observer model. However, some Bayesian model research used the time order error, as a metric for understanding different priors in autistic individuals (Sapey Triomphe et al., 2021). The time order error (TOE) occurs when the first stimulus is perceived with relatively more bias (over and underestimation) than the second stimulus. In the current research, a Bayesian model accounted for a TOE outside the context of a two-stimulus comparison, which was not explained by previous models and may imply that passive memory processes provide a sufficient explanation without the need for processes rooted in interference. Taken together, these results show that Bayesian models provided a relatively good fit for a range of different prior acquisition strategies across different temporal and spatial tasks. Although the current research found a preference for static models in temporal research and some preference for a previous trial model in spatial research. A degree of flexibility, not yet seen in the literature, was displayed in the spatial task regarding participants prior updating strategy. 2024-07-23 Thesis (University of Nottingham only) NonPeerReviewed application/pdf en cc_by https://eprints.nottingham.ac.uk/78381/1/LMcKeown%20thesis%203.pdf McKeown, Lucy (2024) Learning priors for Bayesian perception in normal and atypical development. PhD thesis, University of Nottingham. Bayesian observer model autism autistic perception development perception prior updating prior learning Bayesian priors |
| spellingShingle | Bayesian observer model autism autistic perception development perception prior updating prior learning Bayesian priors McKeown, Lucy Learning priors for Bayesian perception in normal and atypical development |
| title | Learning priors for Bayesian perception in normal and atypical development |
| title_full | Learning priors for Bayesian perception in normal and atypical development |
| title_fullStr | Learning priors for Bayesian perception in normal and atypical development |
| title_full_unstemmed | Learning priors for Bayesian perception in normal and atypical development |
| title_short | Learning priors for Bayesian perception in normal and atypical development |
| title_sort | learning priors for bayesian perception in normal and atypical development |
| topic | Bayesian observer model autism autistic perception development perception prior updating prior learning Bayesian priors |
| url | https://eprints.nottingham.ac.uk/78381/ |