My mind's playing tricks on me: Understanding event integration in rapid stimulus streams.

Humans generally perceive the external world in a coherent manner. This perceptual coherence exists in most individuals despite the supposedly overwhelming number of individual bits of perceptual information available at any given moment. A substantial body of scientific work has been dedicated to understanding the cognitive reasons underlying the coherence of human perception. A key notion in this context was put forth almost a century ago by Gestalt psychologists (Von Ehrenfels, 1937; Wagemans et al., 2012), stating that humans group and categorise sensory input into objects because the human mind is dispositioned to perceive patterns. Grouping sensory input into distinct objects is achieved via perceptual integration. Integration can occur either temporally, spatially, as well as multimodally, allowing us to perceive the world as trees, humans, and melodies, thereby preventing sensory overload. The study of human perception often analyses difficult perceptual tasks in which processes, such as integration, regularly fail. The idea underlying this approach is that insight about the circumstances leading a system, such as that of human perception, to reach its limits and fail simultaneously informs about how the system functions when it functions well.

This thesis will study the limits of perceptual integration in the time domain. We will specifically analyse two cases in which visual stimuli presented in rapid sequence are integrated erroneously, leading to perceptions that were not presented as such in the physical world. The distractor intrusion phenomenon, in which multi-dimensional stimulus features are bound together wrongly, will be investigated first. The 2-feature Simultaneous Type/ Serial Token neural model (2f-ST2 model) will be presented to account for distractor intrusions across a range of different empirical paradigms studied in humans. We will provide behavioural as well as virtual electrophysiological (EEG) results generated by the 2f-ST2 model that qualitatively match those found with humans, providing evidence in favour of the 2f-ST2 model’s validity. Besides, we will provide a series of empirical analyses that further elucidate the cognitive mechanisms underlying distractor intrusions. In essence, these results suggest that whether integration occurs correctly or erroneously depends largely on the timing with which transient attentional enhancement (TAE) impacts relevant cognitive processes.

Temporal event integration describes the cognitive process that binds two rapidly and successively presented visual stimuli into a single percept, given that there is a perceptually meaningful way of doing so. We analysed temporal event integration applying a machine learning approach to EEG data. Our results replicate previous findings on a whole-brain basis and further suggest that temporal event integration occurs about 300 – 450 ms after stimuli were presented specifically, and, more generally, with characteristics that are in line with those of distractor intrusions and the dynamics proposed by the 2f-ST2 model. We finally provide a series of simulation analyses that demonstrate the easily observable and dire methodological risk of overhyping when applying machine learning algorithms to neuroimaging datasets, such as those adopted by ourselves when investigating temporal event integration. Overhyping describes that choosing classification hyperparameters based on which generate the most desirable results (e.g., maximal classification accuracy) can threaten the external validity (i.e., generalisability) of an analysis if the necessary precautions are not taken. In this context we demonstrate that overhyping can lead to classifiers generating spurious above chance-level accuracies even though no signal was present in the data before providing effective safeguards that limit the risk of overhyping.