Symposium: Investigating Within-Trial Timing of Cognitive Processes with EEG

Simple decision processes can be regarded as one of the most important building blocks of behavior. A decision process is here defined as any cognitive process that develops over time and results in a conceptual representation. This includes – but is not limited to – a decision that results in the representation of a course of action, a decision on the content of a perceptual object, and a decision on the meaning of a perceived word. Although current state-of-the-art evidence accumulation models (EAMs) are excellent predictors of behavior that is determined by single decisions, they cannot be used to investigate multiple sequential decisions. It therefore remains unclear how latent decision processes influence subsequent cognitive processes and decisions, and ultimately overt behavior. This has resulted in a lack of understanding of behavior that involves a sequence of decisions – which is imperative, as this is a situation that occurs almost immediately when addressing slightly more complex laboratory tasks, let alone when leaving the lab to investigate real-life situations. In this presentation, we discuss a novel approach in which we first apply machine learning to discover sequential processing stages in EEG data (called Hidden semi-Markov Modeling and Multivariate Pattern Analysis or HMP) and then characterize the duration effects in such identified stages using EAMs. This approach leads to a more fine-grained understanding of decision processes by demarcating the relevant processing stage. Moreover, it allows for understanding multiple decision processes in a single trial that are convoluted in behavior.

This is an in-person presentation on July 21, 2023 (11:00 ~ 11:20 UTC).

My previous work has focused on the possibility that Event-Related Potentials (ERPs) recorded with electroencephalography (EEG) can reflect the beginning of evidence accumulation during speeded decision making. Specifically, N200 peak latencies, negative local peaks in occipital-parietal electrodes around 180 ms, are thought to reflect this onset. I discuss why this matches convergent findings in the literature, and why evidence accumulation is not likely to begin immediately after visual information reaches the occipital cortex in most decision-making tasks. I discuss our recent findings on replication of this work with multiple new sources of data, including a preregistered data set. The replication work was performed with a number of analysis procedures, ranging from: basic regressions of N200 latency estimates compared to various Non-Decision Time (NDT) estimates, to fitting neurocognitive models that predict both choice-response times and single-trial N200 latencies. Multiple datasets were used to assess the theory of N200 latencies reflecting evidence accumulation onsets. The results are presented in the context of the best modeling procedures that generate inference about the truth of N200 latencies, as judged in simulation. We use hierarchical Bayesian modeling, simulation-based Bayesian inference, and computational models of ERPs. While little support was found for a 1 ms to 1 ms correspondence between NDTs and N200 latencies, significant positive relationships were found in most analysis procedures and datasets. I discuss what this means for the theory of the beginning of evidence accumulation during visual tasks. I also discuss the future of understanding ERPs in terms of their computational role in cognition.

This is an in-person presentation on July 21, 2023 (11:20 ~ 11:40 UTC).

Many studies have reported that word frequency influences language processing. For example, the lexical decision (LD) - deciding whether a character string is a word or not - becomes faster and more accurate when word frequency is increased. Yet, our understanding of the precise way continuous changes in frequency impact the different cognitive processes involved in reaching LDs remains limited. To address this, we conducted an EEG LD study in which we manipulated the continuous frequency of Dutch words and non-words (pseudo words and random character strings) and observed the impact on the duration of LD processing stages using a recent machine learning technique. To obtain frequency scores compatible with words and non-words we relied on Google result counts. The trial-level duration estimates of LD processing stages were recovered from EEG using a combination of Hidden semi-Markov models and multivariate pattern recognition (Anderson et al., 2016, Psychol. Rev.). These duration estimates for each processing stage were then analyzed using generalized additive mixed models. We included (potentially nonlinear) effects of frequency and word type as predictors. Confirming previous research, we found evidence for six processing stages. For the first three processing stages (0-70, 70-150, 150-240 ms) stage duration increased for more frequent stimuli. However, already in these earliest stages the effect of frequency differed slightly yet reliably between word types. For the last three processing stages we observed more complex effects of frequency and word type. In contrast to earlier findings, this suggests that frequency has an effect on virtually every process involved in LD, including the earliest ones likely related to visual processing and orthographic encoding.

This is an in-person presentation on July 21, 2023 (11:40 ~ 12:00 UTC).

In the arrow flanker task (Eriksen & Eriksen, 1974), participants have to respond to a central target stimulus while ignoring irrelevant flankers. Mathematical models of conflict processing in the task assume that the response selection process is affected by the processing of interfering flankers until those have been inhibited. For example, the dual-stage two-phase (DSTP) model assumes that attention is in a first stage evenly distributed across the target and flanking stimuli and in the second stage focused solely on the target once irrelevant flankers have been inhibited. In contrast, the shrinking spotlight (SSP) model assumes that attention gradually shifts toward the target stimulus (White et al., 2011). Although both models capture main trends in behavioral data, little is known about how they relate to electrophysiological correlates of conflict resolution and response selection such as the stimulus-locked lateralized readiness potential (sLRP). Using data from 150 participants who completed an arrow flanker task while their EEG was recorded, we integrated parameters of the two mathematical models and electrophysiological correlates of conflict resolution (sLRP peak latencies) in a multi-layer structural equation model framework. Parameters of the mathematical models and electrophysiological correlates of conflict resolution were meaningfully related to each other and demonstrated convergent validity. Both models produced comparable results, but the DSTP model was found to be superior to the SSP model in mapping to ERP measures that reflect sequential processing stages. The findings suggest that both models can help understand how individuals resolve conflicting stimuli, and the DSTP model may be more easily related to electrophysiological measures that capture sequential processing stages.

This is an in-person presentation on July 21, 2023 (12:00 ~ 12:20 UTC).

Neural oscillations at theta frequency (4-8 Hz) are thought to play an important role in guiding behaviour (Cavanagh & Frank, 2014), a phenomenon sometimes labeled as cognitive control. In particular, theta amplitude increases when an unexpected event happens or something goes wrong. However, the computational role of theta has remained unclear. We present a computational model of how theta guides faster (gamma) frequencies, thus to synchronize (functionally, attend to) specific neural modules. In this model, theta amplitude and theta frequency are two dimensions that can be controlled for optimal cognitive control. The model predicts that theta frequency decreases when a more difficult task is upcoming, because slower waves allow more time for competing representations to settle. This decreased frequency should be visible in both neurophysiology (measured via EEG) and behavior (measured via accuracy). In line with model simulations, we find empirically in the EEG spectrum that a cue predicting an upcoming difficult (relative to easy) stimulus, leads to an decreased theta frequency. Similarly, theta frequency measured at the behavioral level is slowed down on such difficult stimuli. This result demonstrates how different aspects of theta oscillations (amplitude and frequency) can be recruited for cognitive control, and how they can be manifested in EEG and behavior.

This is an in-person presentation on July 21, 2023 (12:20 ~ 12:40 UTC).