CNS 2019 Workshop (W2): The dynamics and limitations of working memory

July 16 and 17, 2019; Barcelona, Spain
Location: University of Barcelona, Room B7
Chairs: Zack Kilpatrick (University of Colorado Boulder) and Albert Compte (Institut d'investigacions Biomèdiques August Pi i Sunyer)

Description: We use working memory (WM) to hold and manipulate information when making decisions, solving problems, and guiding behavior. Experiments often require subjects to store items in WM for a few seconds, and use this information to make a decision. Since non-human primates and rodents can be trained on such tasks, it has been possible to record some of the neural substrates of working memory during this delay period. Controversy remains as to how items are represented in the brain, and the effect this has on time and resource limitations of WM. Are memoranda represented with stable and persistent activity or dynamic activity? Are WM resources distributed to minimize error or maximize reward? Combined experimental and computational approaches are needed to resolve these current conflicts, and this workshop will discuss open questions and current approaches to answering them. Approaches include data-driven models, human psychophysics, imaging, and neural activity recordings during working memory delay periods.

Speakers: Ila Fiete (Massachusetts Institute of Technology); Rosanne Rademaker (University of California San Diego); Athena Akrami (University College London); Georgia Gregoriou (University of Crete); Aspen Yoo (New York University); Tim Buschman (Princeton University); Paul Bays (University of Cambridge); Dante Wasmuht (University of Oxford); Klaus Wimmer (Centre de Recerca Matemàtica); Albert Compte (IDIBAPS); Zack Kilpatrick (University of Colorado Boulder)

Tuesday July 16
9:30-10:10am: Paul Bays (University of Cambridge): Stochastic sampling: A unifying framework for working memory limits
10:15-10:55am: Aspen Yoo (New York University): The effect of behavioral relevance on working memory representations
11:00-11:25am: Coffee Break
11:30am-12:10pm: Klaus Wimmer (Centre de Recerca Matemàtica): Circuit mechanisms underlying task-triggered changes in population codes in spatial working memory
12:15-12:55pm: Georgia Gregoriou (University of Crete): Encoding and retention of spatial and non-spatial information in the parietal and prefrontal cortices

8:00-10:00pm: Possibly Self-Organized Dinner Group

Wednesday July 17
9:30-10:10am: Tim Buschman (Princeton University): Neural dynamics improve working memory
10:15-10:55am: Dante Wasmuht (University of Oxford): Dynamical multiplexing of information during dual task performance in monkey prefrontal cortex
11:00-11:25am: Coffee Break
11:30am-12:10pm: Rosanne Rademaker (University of California San Diego): Representations in visual cortex during changing memory and sensory demands
12:15-12:55pm: Ila Fiete (Massachusetts Institute of Technology): TBA
1:00-2:45pm: Lunch
2:45-3:25pm: Albert Compte (Institut d'Investigacions Biomèdiques August Pi i Sunyer): Reactivation of prefrontal cortex representations boost serial biases in working memory
3:30-4:10pm: Zack Kilpatrick (University of Colorado Boulder): Neural and synaptic mechanisms of interference in working memory
4:15-4:40pm: Coffee Break
4:45-5:25pm: Open
5:30pm--: Open Discussion Possibly over Drinks
8:00pm--: Workshop Dinner Organized by Albert

Titles and Abstracts
Athena Akrami, University College London
Demarcation of working memory - what's the timescale and content of working memory?

Paul Bays, University of Cambridge
Stochastic sampling: A unifying framework for working memory limits
Recent debate regarding working memory limits has focused on whether memory representations are better characterized as discrete or continuous, with models of each type competing to best capture the errors humans make in recall. I will argue that this dichotomy is misleading, and that the key distinction is between fixed (deterministic) and stochastic mechanisms of WM, with only the latter compatible with observed human performance and the underlying biological system. I will show that reconceptualizing existing models in terms of sampling reveals strong commonalities between supposedly opposing accounts. A probabilistic limit on the number of items successfully recalled from WM is an emergent property of continuous resource models, despite these models having no explicit mechanism to enforce such a limit. Furthermore, adding stochasticity in the number of samples to a discrete model puts its performance on a par with continuous models. A clearer understanding, and also the best fits to data, are obtained by considering biologically plausible implementations of WM based on the inherently stochastic spiking activity of neural populations.

Tim Buschman, Princeton University
Neural dynamics improve working memory
Working memory is our ability to hold things ‘in mind’, acting as a flexible substrate on which thoughts can be placed and manipulated. As such, working memory is critical to cognition. Despite its importance, working memory is surprisingly fragile. Working memory representations degrade over time due to noise and are susceptible to interference from other stimuli. In this talk I will discuss how dynamics in neural activity mitigates these issues. Using a contextual prediction paradigm in mice, we studied how neurons in early sensory cortex represented both the immediate sensory stimulus as well as the memory of recent stimuli. Surprisingly, memory representations in sensory cortex were highly dynamic, similar to observations from prefrontal cortex in monkeys. Furthermore, we found these dynamics were very specific, acting to reduce interference between the memory representation and new sensory stimuli. These results highlight the importance of dynamics in working memory representations, showing how they can protect working memory from interference.

Albert Compte, Institut d'Investigacions Biomèdiques August Pi i Sunyer (Barcelona)
Reactivation of prefrontal cortex representations boost serial biases in working memory
Working memory includes memory processes operating at time scales beyond one trial, as revealed by serial attractive biases in delayed response tasks. By investigating monkey and human electrophysiology data, I will show in this talk that underlying this behavior is the interaction between attractor dynamics within trial, and selective silent subthreshold mechanisms between trials. Reactivation of neural activity selective to the previous trial stimulus in anticipation of the new trial demonstrates how these two mechanisms interact in the prefrontal network in the course of a working memory task. Furthermore, I will present evidence that reactivation of stimulus selectivity prior to a trial is related to the enhancement of serial biases, by using both trial-by-trial correlational analyses and causal perturbation experiments (TMS).

Ila Fiete, Massachusetts Institute of Technology
TBA

Georgia Gregoriou, University of Crete
Encoding and retention of spatial and non-spatial information in the parietal and prefrontal cortices
Electrophysiological recordings in non-human primates have revealed that neurons in both prefrontal and parietal cortices exhibit delay period activity, which is thought to control task relevant information. In this talk, I will present data from simultaneous recordings in the frontal eye fields (FEF), an area within PFC, and the lateral intraparietal area (LIP), in PPC, that aimed to directly compare how behaviorally relevant information about the location and identity of stimuli is represented in the two areas during presentation and memory of a task relevant stimulus. Our results point to differences in spatial and object-related information processing in the two areas and reveal frequency- and epoch-specific synchronous interactions between FEF and LIP.

Zack Kilpatrick, University of Colorado Boulder
Neural and synaptic mechanisms of interference in working memory
Interference occurs in working memory when memoranda interact to produce systematic errors. We develop and analyze neural field models of activity which account for interference within and across trials in visual working memory tasks. Each memorandum is represented by a neural activity bump, which wanders due to stochastic forcing describing voltage and synaptic fluctuations, accounting for normally distributed saccade errors. Human and nonhuman primate studies have shown that, in a sequence of working memory trials, the location of the previous target attractively biases current trial responses. We account for this bias with short term facilitation, showing this synaptic process implements suboptimal inference process in which an observer wrongly assumes the previous trial predicts the next. In addition, we discuss a model of within trial interference in which an observer must store multiple items, each represented by an activity bump. Bump interactions account for commonly observed biases, and we predict an optimal synaptic network strength for a desired item number storage limit.

Rosanne Rademaker, University of California, San Diego
Representations in visual cortex during changing memory and sensory demands
People often remember visual information over brief delays while actively engaging with ongoing inputs from the surrounding visual environment. Depending on the situation, one might prioritize mnemonic contents (i.e. remembering details of a past event), or preferentially attend sensory inputs (i.e. watching traffic while crossing a street). I’ll show that population-level response patterns in early visual cortex can represent the contents of working memory concurrently with passively viewed sensory inputs. On the other hand, trade-offs between memory and sensory representations emerge with changes in attentional priority. Collectively, as behavioral requirements change, memory representations change along with it. This flexibility allows the working memory system to achieve remarkable stability, with recall being only moderately affected by passive or active interference.

Dante Wasmuht, University of Oxford
Dynamical multiplexing of information during dual task performance in monkey prefrontal cortex
Cognitive multitasking is limited by interference between component tasks. Neural substrates for such interference have previously been identified in the lateral prefrontal cortex (lPFC). However, lPFC processes information in a dynamic, context dependent manner, possibly enabling multitasking during complex behaviour. We recorded from lPFC during an overlapping spatial attention and working memory task (dual task). Monkeys learned to perform the dual task with a high success rate and over various levels of experimentally induced inter-task interference. Prefrontal neurons were tuned to unique task features as well as to complex mixtures of features from both tasks and time i.e. switching, linear-mixed and nonlinear-mixed selectivity. This heterogeneity in neural responses resulted in distinct patterns of activity across the population. Population activity patterns transitioned between distinct neural subspaces across task epochs, while protecting task specific representations from interference. This was characterised by orthogonal subspaces for attention and memory locations, as well a common subspace coding task-invariant spatial location. Our data suggests that lPFC multiplexes information within and across tasks as well as time to achieve concurrent but distinct processing on the level of the population.

Klaus Wimmer, Centre de Recerca Matemàtica (Barcelona)
Circuit mechanisms underlying task-triggered changes in population codes in spatial working memory
Neural population activity recorded from primate prefrontal cortex carries information about the remembered stimulus, maintaining a stable representation throughout the delay period of working memory tasks. However, during cue presentation and in the beginning of the delay period the population code is dynamic, and individual prefrontal neurons show heterogeneous activity including strong temporal dynamics in all task phases. Here we set out to investigate how individual neurons support the dynamic and stable population codes and what are the underlying circuit mechanisms. We analyze the dynamics of single neurons and population activity during oculomotor delayed response tasks and show that it can be explained by a network model in which highly dynamic activity during the initial memory storage originates mainly from different neuronal subpopulations. After this long initial transient, the model circuit reaches a stable state, and memory maintenance is achieved through attractor dynamics.

Aspen Yoo, New York University
The effect of behavioral relevance on working memory representations
One of the defining properties of working memory is that it is limited. This raises interesting questions about what information we maintain and use to behave efficiently in our environment. For example, humans allocate working memory resource according to behavioral relevance, resulting in more precise memories for more important items. How might behavioral relevance be represented in the brain? What computations are carried out when deciding how much resource to allocate to each item? How might representing the uncertainty associated with a remembered item aid in decision-making? I have approached these questions with a combination of psychophysical, computational, and neuroimaging methods. I will present results suggesting that behavioral relevance is encoded in some of the same neural populations that encode location, people allocate resource in a way consistent with a loss-minimizing strategy, and people can use memory uncertainty optimally in perceptual decisions.