| Invited Symposia | Symposia

Note: Symposia have not yet been scheduled.

Multiscale Dynamics of Cognitive Processes from M/EEG: Micro-to-Macroscale and Back Again

Chair: Darcy Waller¹; ¹Brown University
Presenters: Atsushi Kikumoto, Bradley Postle, Darcy Waller, Saskia Haegens

From evoked responses to oscillations, signatures of cognition recorded using magneto-/electro-encephalography (M/EEG) have been powerful tools in cognitive neuroscience over decades. However, the challenge of applying these methods to probe mechanisms that underlie dynamic and distributed processes (e.g., network-wide cognitive “representations”) unfolding during cognition is nontrivial. Some multivariate analytic methods (e.g., multivariate decoding) solve the problem of quantifying distributed representations in time, but with results limited to the original scale and modality, reducing their utility for crossing scales. The speakers in this symposium offer solutions to this challenge at multiple scales, probing mechanisms underlying cognition by studying how they are controlled and manipulated in contexts such as inhibitory control, working memory (WM), and decision-making. Altogether, this symposium will illustrate how M/EEG data can be used in diverse ways to provide integrative insights into how cognitive processes emerge across mechanistic scales. At the microscale, Darcy Waller will present biophysical modeling of EEG data to study prefrontal microcircuit mechanisms of reactive and proactive inhibitory motor control. Atsushi Kikumoto will address macroscales, demonstrating how representational similarity analyses reveal distributed conjunctive and competing representations during error monitoring while accounting for regional ERPs. Brad Postle will illustrate how analyses of RNN dynamics can provide new insights into the algorithmic implementations of WM content prioritization that are paralleled in subspaces of EEG data. Finally, Saskia Haegens will demonstrate how the challenge of separating population-level neural activity can be tackled in part by using decomposition and frequency-specific analyses to identify functionally distinct but overlapping oscillatory rhythms.

Presentations

Distinct inhibitory microscale mechanisms of proactive and reactive control from biophysical models of prefrontal EEG

Darcy Waller¹, C.J. Abeshaus², Stephanie Jones¹; ¹Brown University, ²Scripps College

Theories of response inhibition (RI) in the Stop-Signal Task (SST) provide important explanations of cognitive processes, but the underlying neural mechanisms that support stopping through the combination of reactive and proactive control in this task have proven difficult to infer and integrate. Clues into the neural mechanisms of RI arise from EEG signatures that relate to stopping in the SST. For example, components of the Stop-Signal-locked frontocentral event-related potential (FC-ERP) relate systematically to stopping success. Additionally, frontocentral oscillatory signatures that relate to stopping include transient beta (15-29Hz) events, increased during successful stopping. We used Human Neocortical Neurosolver (HNN), a biophysically detailed model of the cortex under external drive, to simulate the microscale generators of FC-ERP components and oscillations related to both proactive and reactive control in the SST. We examined baseline-uncorrected ERPs, finding that contributions of proactive control can be observed in reduced pre-Stop-Signal voltage in Stop trials. This negativity could be reproduced in HNN using local fast and persistent cortical inhibition. In combination with inputs producing beta events and ERP deflections, our models reproduced post-Stop-Signal differences in ERP characteristics across conditions (e.g., smaller P2 in Successful Stops). Close fits to Successful Stop ERPs relied both on increases in proactive control-associated fast inhibition and cue-induced, beta-associated slow inhibition mediating reactive control. These results provide novel predictions about how proactive and reactive control facilitate stopping via distinct, complementary pathways. Future work will consider these mechanisms within a wider network perspective and determine whether they facilitate control over other prefrontal cognitive representations.

Error monitoring emerges from competing conjunctive subspaces

Atsushi Kikumoto^1,2, Kazuhisa Shibata², David Badre³; ¹University of Maryland, ²RIKEN Center for Brain Science, ³Carnegie Mellon University

For behavior to be adaptive, detected errors must be interpreted in light of current goals and contexts. An action that appears mistaken—breaking a window in an office—may be appropriate in another context, such as escaping a fire. How are contexts incorporated into error monitoring, and which representational states inform such metacognitive decisions? Fast and accurate action selection depends on conjunctive neural subspaces integrating contexts, stimuli, and responses within a stable, high-dimensional geometry. We tested whether these subspaces also provide the internal reference for error monitoring, specifically, whether representations corresponding to committed errors and alternative correct actions coexist during evaluation. Participants performed rule-based, context-dependent actions with a response deadline adjusted to calibrate error rates and judged whether their responses were correct or incorrect while EEG was recorded. Using single-trial representational similarity analysis, we recovered task representations associated with committed actions (those erroneously executed) and uncommitted actions (cued but unexecuted alternatives) from distributed EEG patterns at the sub-second scale. During error commissions, committed and uncommitted representations coexisted, with the committed representation ultimately dominating prior to the incorrect response. Yet, during error judgments, uncommitted conjunctions were reactivated and modulated both a low-dimensional decision subspace and domain-general neural signatures of error monitoring, including the error-related negativity (ERN) and error positivity (Pe). Critically, conflict between task representations predicted trial-by-trial error-monitoring accuracy, even after accounting for error-related modulation of the ERN and Pe. These results suggest that error monitoring is directly informed by competing, context-specific task representations.

Harnessing Recurrent Neural Networks to Bridge Levels of Analysis with EEG Data

Bradley Postle¹, Jacqueline Fulvio¹, Jung Woo Hur¹; ¹University of Wisconsin-Madison

It is well established that, although simulation with artificial neural networks can’t “prove” how a task is performed by the brain, they can be powerful tools for generating novel hypotheses about how a task might be performed. This presentation will review our experience simulating performance of a double-serial retrocuing (DSR) working memory (WM) task with simple recurrent neural networks (RNN), then applying dimensionality-reduction techniques to patterns of hidden-layer activity. In this way, we identify latent variables whose dynamics can be interpreted as a normative model for algorithmic bases for controlling WM. One specific example is that demixed principal component analysis (dPCA) of the RNN data reveal that, nonintuitively, prioritization in WM need not be encoded via explicit priority-related changes in the representational geometry of the stimuli themselves. Instead, priority status can be encoded only implicitly, by an item’s location along a (latent) context x priority dimension that emerges with the onset of the prioritization cue. Remarkably, we have observed a similar phenomenon in preliminary analyses of EEG data from humans performing the same task, although in the EEG data stimulus manifolds are segregated by phase-shifted slow oscillation around a (latent) stimulus-value x context x priority axis. Yet to be discovered is how these oscillations in the S-VxCxP-defined subspace may relate to oscillatory dynamics in the scalp-level data. In this way we seek to bridge between latent, population-based mechanisms and the “implementational level” of scalp-level ERPs and oscillations that have been the bread and butter of cognitive electrophysiology for decades.

Disentangling oscillatory dynamics — key to understanding the role of brain oscillations

Saskia Haegens¹; ¹Columbia University

A vast body of research has linked brain oscillations to perception and cognitive function. However, the underlying mechanisms of these population-level dynamics remain elusive. Recent work shows that isolating oscillatory dynamics — that is, separating out aperiodic contributions, and disentangling oscillations that overlap in space, time, and frequency range — is key to addressing this question. Our recent M/EEG work on the alpha and beta rhythms illustrate this approach: 1) Using independent component analysis (ICA) we were able to isolate two functionally distinct posterior alpha rhythms (8–14 Hz), which were differentially modulated during a working-memory task, with opposite relation­ship to behavioral outcomes. Our results support the idea that alpha plays a functionally inhibitory role. Crucially, these distinct alpha dynamics only became evident using a decomposition approach and were not observed using standard sensor-level analysis, emphasizing the necessity of analysis methods that can disentangle overlapping rhythmic dynamics that stem from different sources. 2) In a related line of research focused on the role of the beta rhythm (15–30 Hz) in a categorical decision-making task, we were able to dissociate neural ensembles synchronized at slightly different beta frequencies. We found that beta-band frequency shifts in frontal cortex signal categorical decision outcomes, with activity in these bands predicting the behavioral response. Our results further substantiate the idea that beta reflects the flexible formation of neural ensembles. Uncovering these dynamics in M/EEG data relied on source reconstruction informed by our prior NHP results, and instantaneous frequency analysis to reveal subtle frequency shifts.