Working Memory Requires a Combination of Transient and Attractor-Dominated Dynamics to Process Unreliably Timed Inputs

Abstract

Working memory stores and processes information received as a stream of continuously incoming stimuli. This requires accurate sequencing and it remains puzzling how this can be reliably achieved by the neuronal system as our perceptual inputs show a high degree of temporal variability. One hypothesis is that accurate timing is achieved by purely transient neuronal dynamics; by contrast a second hypothesis states that the underlying network dynamics are dominated by attractor states. In this study, we resolve this contradiction by theoretically investigating the performance of the system using stimuli with differently accurate timing. Interestingly, only the combination of attractor and transient dynamics enables the network to perform with a low error rate. Further analysis reveals that the transient dynamics of the system are used to process information, while the attractor states store it. The interaction between both types of dynamics yields experimentally testable predictions and we show that this way the system can reliably interact with a timing-unreliable Hebbian-network representing long-term memory. Thus, this study provides a potential solution to the long-standing problem of the basic neuronal dynamics underlying working memory.Humans and animals continuously receive information conveyed by stimuli from the environment. To survive, the brain has to store and process this stream of information which is mainly attributed to the processes of working memory (WM 1, 2 ). These two distinct abilities of WM, to store and to process information, yield a debate about the underlying neuronal network dynamics 3-5 : the network dynamics might either follow (i) attractor or (ii) transient dynamics.Attractor dynamics denotes neuronal network dynamics which is dominated by groups of neurons being persistently active. In general, such a persistent activation is related to an attractor state of the dynamics, with each attractor associated to a specific information content 3, 6-8 . Several experimental and theoretical studies hypothesize that the dynamics underlying WM are dominated by such persistent dynamics 5,[8][9][10] . In contrast to attractor dynamics, neuronal networks with transient dynamics are dominated by an attractor-less continuous flow of neuronal activity across a possibly large neuronal population [11][12][13][14] . This type of dynamics implies a high diversity and complexity which is linked by theoretical studies with a large computational capacity required to process information [15][16][17] . These theoretical studies as well as several pieces of experimental evidence 18-20 yield the hypothesis that the dynamics underlying WM are dominated by transient dynamics 20,21 . Thus, although the two hypotheses -attractor or transient dynamics -seem to contradict each other, experimental and theoretical evidence supports both yielding a debate about the neuronal network dynamics underlying WM 5 .To resolve this contradiction, in this study, we consider the fact that the timing of stimuli received by the WM i...

The interplay of synaptic plasticity and scaling enables self-organized formation and allocation of multiple memory representations

“…Due to the recurrent connections, the memory area should robustly form internal representations of the environmental stimuli, while simultaneously the feed-forward synapses should provide a proper allocation of the stimuli (activity patterns in input area) onto the corresponding representations. In the following, we will show in a step-by-step approach that the interplay of long-term synaptic plasticity (here conventional Hebbian synaptic plasticity) with homeostatic synaptic scaling (see Methods; 7,23,41 ) coordinates synaptic changes such that proper formation and allocation of memory is ensured.…”

“…Similar to previous studies 23,26,41 , we consider here an abstract model to describe the neuronal and synaptic dynamics of the network. Despite the abstract level of description, the model matches experimental in-vivo data of memory allocation.…”

“…By contrast, if a stimulus (here stimulus S1) is repeatedly presented in a given time interval, the neuronal and synaptic dynamics of the network reshapes the pattern of activated neurons in the memory area such that the final pattern consists of a group of interconnected neurons (decrease in ASPL; 2 B, test 1). As shown in our previous studies 23,41 , the combination During the test phases, the resulting network structure is evaluated. Top row: average synaptic weight of feed-forward synapses from input populations I1 and I2 activated by the corresponding stimuli to the groups of neurons which become a CA (CA1 and CA2) and others (RR).…”

A Robust Model of Gated Working Memory

“…However, in many dynamical systems, reservoir included, there exists a trade-off between memory capacity and non-linear computation 343 . This is why some studies have focused on reservoirs with dedicated readout units acting as working memory (WM) units 40,35,47 . These WM units have feedback connections projecting to the reservoir and are trained to store binary values that are input-dependent.…”

Distributed dynamic coding for spatial working memory in hippocampal-prefrontal networks

“…For example, hippocampal pyramidal cells are strongly theta modulated during navigation, and this is strongly correlated with the spatial tuning of firing 16–18 . In SWM, in silico evidence suggests that ongoing oscillatory modulation is critical for building appropriate rate coding dynamics, especially when crucial decision-points are unreliably timed 19 . Beyond oscillatory firing at the single neuron level within the PFC, it is clear that the hippocampal-prefrontal network also has tightly regulated coordination between structures at the level of local field potentials, including coherence in multiple frequency bands 6,20–22 .…”