Testing the predictive coding account of temporal integration in the human visual system : a computational and behavioural study

Aitken, Fraser ; Ales, Justin Michael ; Ales, Justin Michael

University:University of St Andrews

Department:Psychology & Neuroscience (School of)

Others  :  https://research-repository.st-andrews.ac.uk/bitstream/handle/10023/19158/FraserAitkenPhDThesis.pdf?sequence=2&isAllowed=y

来源: DR-NTU

【 摘 要 】

A major goal of vision science is to understand how the visual system maintains behaviourallyrelevant perceptions given the level of uncertainty in the signals it receives. One proposed solution isthat the visual system applies predictive coding to its inputs based on the integration of priorknowledge and current stimulus features. However, support for some vital aspects of predictivecoding in the temporal domain is lacking and simpler accounts of temporal integration also exist. Theaim of this thesis was to test two key attributes of predictive coding in time a) does the visual systemapply adaptive weighting to prediction errors and b) can the visual system apply probabilisticinformation learnt from stimulus sequences when making predictions. In chapters 3 & 4, we testedpredictive coding’s ideas of how prediction errors are weighted under the theoretical guidance of atemporal integration model linked to predictive processing, called the Kalman filter. Here, bothexperiments supported predictive coding. We showed that, consistent with the Kalman filter, visualestimates and the way estimation errors were corrected, adapted to stimulus behaviour and viewingconditions. In chapter 5, we assessed the ability of the visual system to integrate conditionalrelationships present in sequences of stimuli when making predictions. To do this, we inserted astimulus sequence that changed and omitted trials based on Markov transition probabilities that made some transitions more or less probable and assessed reaction times and omission trial responses.Reaction time data was consistent with predictive coding, in that more predictable changes elicitedfaster responses.Omission trials data, was though, less clear. When faced with no stimulus,participants did not apply the conditional probabilities in their decisions optimally, instead applyingnon optimal decision strategies, inconsistent with predictive coding. In summary, this thesis supportsthe predictive coding of temporal integration but questions its application in all situations.