Talks

I will present data showing that a simple bump attractor model with weak inhomogeneities and short-term plasticity can link prefrontal single-neuron activity with fine-grained behavioral output and account for the main delay-dependent distortions in spatial working memory: precision, cardinal repulsion biases and serial dependence. I will show that serial dependence is specifically affected in neuropsychiatric disorders and propose affected neural mechanisms based on the model.

The world around us is complex, but at the same time full of meaningful regularities. We can detect, learn and exploit these regularities automatically in an unsupervised manner i.e. without any direct instruction or explicit reward. For example, we effortlessly estimate the average tallness of people in a room, or the boundaries between words in a language. These regularities and prior knowledge, once learned, can affect the way we acquire and interpret new information to build and update our internal model of the world for future decision-making processes. Despite the ubiquity of passively learning from the structured information in the environment, the mechanisms that support learning from real-world experience are largely unknown. By combing sophisticated cognitive tasks in human and rats, neuronal measurements and perturbations in rat and network modelling, we aim to build a multi-level description of how sensory history is utilised in inferring regularities in temporally extended tasks. In this talk, I will specifically focus on a comparative rat and human model to study building and utilising statistical prior distributions in working memory and decision making behaviours.

How does a system deprioritize information held in working memory? Multivariate decoding and encoding analyses of fMRI and EEG datasets have suggested/supported possibilities (e.g., activity-silent representations, priority-based remapping), but alone are inconclusive. After training RNNs to perform a 2-back task, applying demixed (d)PCA to activity patterns from their hidden layers revealed evidence that information about on item ‘n’ transitions to an “unprioritized” subspace while ‘n – 1’ guides behavior, then flips into a “prioritized” subspace prior to its comparison against ‘n + 2’. This provides quantitative hypotheses about the priority-based transformation of information that can be tested in human brains.

I will present a number of somewhat diverse research lines that I am currently pursuing together with colleagues in- and outside my lab – none of which are near a stage of completion. They all reflect the idea that VWM representations are not only OF something, but also FOR something. That is, we use VWM representations for different goals, which may then shape these representations or any associated control processes accordingly. In one project we compared VWM representations used to find targets in a subsequent visual search display to representations used to ignore distractors, to investigate if observers flexibly and proactively suppress distractors. In another project we compare memories of stimuli that are the target for a future action to memories of stimuli that are equally task-relevant, but not the direct target of future action. Recent frameworks have linked activation of VWM representations to action selection, and our experiments allowed us to investigate if actions links strengthen the memory representation. Finally, in a collaborative project with Utrecht University, we look at the difference in multivariate activation patterns when representations are used for an attention task versus when used for a memory task, testing the hypothesis that continued access to the stimulus will actually result in minimal representation.

Visual processing is strongly influenced by recent stimulus history. Prominent theories cast this as a consequence of optimized encoding of visual information by exploiting the temporal statistics of the world. However, this would require the visual system to track the history of individual briefly experienced events, within a stream of visual input, to build up statistical representations over longer timescales. In my talk, I will provide behavioral and neural evidence that neurons in the early visual cortex of the mouse indeed maintain long-term traces of individual past stimuli that persist despite the presentation of several intervening stimuli, leading to long-term and stimulus-specific adaptation over dozens of seconds. Early visual cortex thus maintains concurrent stimulus-specific memory traces of past input, enabling the visual system to build up a statistical representation of the world to optimize the encoding of new information in a changing environment.

Perception is determined by what is attended. In recent years, it has become clear that what is attended is not just determined by our current goals, but strongly biased by past experience grounded in statistical regularities in the environment. In this talk, I will discuss findings from a series of behavioral and EEG studies that examined how such statistical learning affects what we attend to and what we ignore. These findings show that the brain continuously incorporates the statistical structure of the world, affecting how and what information is represented in the mind’s eye.

Working memory plays a central role in cognition, acting as a workspace on which we can place and manipulate our thoughts. As such, cognitive control determines what gets into working memory, and how the contents of working memory are used to guide behavior. In this talk, I will discuss the neural mechanisms that support the control of working memory, focusing on how moving working memory representations between different subspaces can determine what information is processed, how it is transformed, and when one acts.

In our Open Research Area project, we have focused on studying the properties of latent working memory (WM) representations. It has recently been proposed that WM may not only rely on ongoing neural activity, but also on connectivity, which is activity-quiescent or even activity-silent. These quiescent states may play an important role in particular when passive, latent items are being maintained. It is challenging to probe latent representations, because quiescent states are by their nature effectively invisible to standard activity measures in cognitive neuroscience. However, by presenting a visual impulse to perturb the underlying brain network, in combination with multivariate pattern analysis of the resultant impulse response signal, it is possible to illuminate and reveal representations held in quiescent network states. In a series of EEG experiments based on this perturbation technique, we found pervasive evidence for continued maintenance of seemingly useless
information; from previously transformed items, to de-prioritized items, and task-irrelevant properties. It thus seems that WM may hold more than meets the eye, particularly with regard to functionally and
physiologically latent items.

Classical models of working memory propose that contents are either maintained in a semantic or a visuospatial format. Here we applied representational similarity analysis to time-frequency-resolved intracranial EEG data in order to track the maintenance of stimulus-specific representations of natural objects. Using deep neural networks, we then analyzed the representational format of these representations and found that objects were simultaneously maintained in high-level sensory and abstract semantic formats. In a follow-up study, we investigated how sequences of multiple items were maintained and how working memory could be purged from irrelevant contents via top-down control mechanisms.