Brodski-Guerniero, Alla (2018)
Neural correlates of the use of prior knowledge in predictive coding.
Technische Universität Darmstadt
Ph.D. Thesis, Primary publication
|
Text
AllaBrodskiGuerniero_PhDThesis_uploadULB_edited2.pdf Copyright Information: In Copyright. Download (3MB) | Preview |
Item Type: | Ph.D. Thesis | ||||
---|---|---|---|---|---|
Type of entry: | Primary publication | ||||
Title: | Neural correlates of the use of prior knowledge in predictive coding | ||||
Language: | English | ||||
Referees: | Galuske, Prof.Dr. Ralf ; Matthias, PD Munk ; Michael, Prof.Dr. Wibral | ||||
Date: | 2018 | ||||
Place of Publication: | Darmstadt | ||||
Date of oral examination: | 26 January 2018 | ||||
Abstract: | Every day, we use our sensory organs to perceive the environment around us. However, our perception not only depends on sensory information, but also on information already present in our brains, i.e. prior knowledge acquired by previous experience. The idea that prior knowledge is required for efficient perception goes back to Hermann von Helmholtz (1867). He raised the hypothesis that perception is a knowledge-driven inference process, in which prior knowledge allows to infer the (uncertain) causes of our sensory inputs. According to the currently very prominent “predictive coding theory” (e. g. Rao and Ballard, 1999; Friston, 2005, 2010; Hawkins and Blakeslee, 2005; Clark, 2012; Hohwy, 2013) this inference process is realized in our brains by using prior knowledge to build internal predictions for incoming information. Despite the increasing popularity of predictive coding theory in the last decade (see Clark, 2012 and comments to his article), previous research in the field has left out several important aspects: 1. The neural correlates of the use of prior knowledge are still widely unexplored; 2. Neurophysiological evidence for the neural implementation of predictive coding is limited and 3. Assumption-free approaches to study predictive coding mechanism are missing. In the present work, I try to fill these gaps using three studies with magnetoencephalographic (MEG) recordings in human participants: Study 1 (n = 48) investigates how prior knowledge from life-long experience influences perception. The results demonstrate that prediction errors induced by the violation of predictions based on life-long experience with faces are reflected in increased high-frequency gamma band activity (> 68 Hz). For studies 2 and 3, neurophysiological analysis is combined with information-theoretic analysis methods. These allow investigating the neural correlates of predictive coding with only few prior assumptions. In particular, the information-theoretic measure active information storage (AIS; Lizier et al., 2012; Wibral et al., 2014) can quantify how much information is maintained in neural activity (predictable information). I use AIS in order to study the neural correlates of activated prior knowledge in study 2 and 3. Study 2 (n = 52) assesses how prior knowledge is pre-activated in task relevant states to become usable for predictions. I find that pre-activation of prior knowledge for predictions about faces increases alpha and beta band related predictable information as measured by AIS in content specific brain areas. Study 3 (n patients = 19; n controls = 19) explores whether predictive coding related mechanism are impaired in autism spectrum disorder (ASD). The results show that alpha and beta band related predictable information is reduced in the brain of ASD patients, in particular in the posterior part of the default mode network. These findings indicate reduced use or precision of prior knowledge in ASD. In summary, the results presented in the present work illustrate the neural correlates of the use of prior knowledge in the predictive coding framework. They provide neurophysiological evidence for the link of prediction errors and fast neural activity (study 1, gamma band) as well as predictions and slower neural activity (study 2 and 3, alpha and beta band). These findings are in line with a theoretical proposal for the neural implementation of predictive coding theory (Bastos et al., 2012). Further, by application of AIS analysis (study 2 and 3) the present work introduces the largely assumption-free usage of information-theoretic measures to study the neural correlates of predictive coding in the human brain. In future, analysis of predictable information as measured by AIS may be applied to a broad variety of experiments studying predictive coding and also for research on neuropsychiatric disorders as has been demonstrated for ASD. |
||||
Alternative Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-72273 | ||||
Classification DDC: | 500 Science and mathematics > 500 Science 500 Science and mathematics > 570 Life sciences, biology |
||||
Divisions: | 10 Department of Biology > Neurophysiology and Neurosensory Systems 10 Department of Biology > Systems Neurophysiology |
||||
Date Deposited: | 15 Feb 2018 14:02 | ||||
Last Modified: | 09 Jul 2020 02:01 | ||||
URI: | https://tuprints.ulb.tu-darmstadt.de/id/eprint/7227 | ||||
PPN: | 426506278 | ||||
Export: |
View Item |