FRINK Linked Data Fragments server

Matches in SemOpenAlex for { <https://semopenalex.org/work/W2950376067> ?p ?o ?g. }

• W2950376067 endingPage "160.e8" @default.
• W2950376067 startingPage "147" @default.
• W2950376067 abstract "•Random recurrent connections can support flexible working memory•Overlap of connections causes interference between memories, limiting capacity•Model captures many behavioral and physiological characteristics of working memory•Structured sensory networks can constrain high-dimensional random representations Working memory is fundamental to cognition, allowing one to hold information “in mind.” A defining characteristic of working memory is its flexibility: we can hold anything in mind. However, typical models of working memory rely on finely tuned, content-specific attractors to persistently maintain neural activity and therefore do not allow for the flexibility observed in behavior. Here, we present a flexible model of working memory that maintains representations through random recurrent connections between two layers of neurons: a structured “sensory” layer and a randomly connected, unstructured layer. As the interactions are untuned with respect to the content being stored, the network maintains any arbitrary input. However, in our model, this flexibility comes at a cost: the random connections overlap, leading to interference between representations and limiting the memory capacity of the network. Additionally, our model captures several other key behavioral and neurophysiological characteristics of working memory. Working memory is fundamental to cognition, allowing one to hold information “in mind.” A defining characteristic of working memory is its flexibility: we can hold anything in mind. However, typical models of working memory rely on finely tuned, content-specific attractors to persistently maintain neural activity and therefore do not allow for the flexibility observed in behavior. Here, we present a flexible model of working memory that maintains representations through random recurrent connections between two layers of neurons: a structured “sensory” layer and a randomly connected, unstructured layer. As the interactions are untuned with respect to the content being stored, the network maintains any arbitrary input. However, in our model, this flexibility comes at a cost: the random connections overlap, leading to interference between representations and limiting the memory capacity of the network. Additionally, our model captures several other key behavioral and neurophysiological characteristics of working memory. Working memory is the ability to hold information “in mind.” It acts as a workspace on which information can be held, manipulated, and then used to guide behavior. In this way, it plays a critical role in cognition, decoupling behavior from the immediate sensory world. However, the circuit mechanisms that support working memory remain unclear. In particular, existing models fail to capture key behavioral and neural characteristics of working memory. Working memory has two defining behavioral characteristics. First, it is highly flexible: one can hold anything in mind and can do it from the first experience. This provides cognition with its versatility, allowing us to think and learn about anything. Second, working memory has a severely limited capacity. Humans and monkeys are able to maintain only 3 or 4 objects at once (Luck and Vogel, 1997Luck S.J. Vogel E.K. The capacity of visual working memory for features and conjunctions.Nature. 1997; 390: 279-281Crossref PubMed Scopus (2848) Google Scholar, Cowan, 2010Cowan N. The magical mystery four: how is working memory capacity limited, and why?.Curr. Dir. Psychol. Sci. 2010; 19: 51-57Crossref PubMed Scopus (637) Google Scholar, Buschman et al., 2011Buschman T.J. Siegel M. Roy J.E. Miller E.K. Neural substrates of cognitive capacity limitations.Proc. Natl. Acad. Sci. USA. 2011; 108: 11252-11255Crossref PubMed Scopus (189) Google Scholar). In other words, although one can hold anything in mind, one can only hold a few of them at a time. In addition to these behavioral characteristics, previous work has identified several neural characteristics of working memory. First, the contents of working memory are thought to be represented in both the persistent activity of neurons (Funahashi et al., 1989Funahashi S. Bruce C.J. Goldman-Rakic P.S. Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortex.J. Neurophysiol. 1989; 61: 331-349Crossref PubMed Scopus (1966) Google Scholar, Fuster, 1999Fuster J.M. Memory in the Cerebral Cortex: An Empirical Approach to Neural Networks in the Human and Nonhuman Primate. MIT Press, 1999Google Scholar, Romo et al., 2002Romo R. Hernández A. Zainos A. Lemus L. Brody C.D. Neuronal correlates of decision-making in secondary somatosensory cortex.Nat. Neurosci. 2002; 5: 1217-1225Crossref PubMed Scopus (279) Google Scholar) and in the dynamic evolution of neural activity over time (Murray et al., 2017Murray J.D. Bernacchia A. Roy N.A. Constantinidis C. Romo R. Wang X.-J. Stable population coding for working memory coexists with heterogeneous neural dynamics in prefrontal cortex.Proc. Natl. Acad. Sci. USA. 2017; 114: 394-399Crossref PubMed Scopus (160) Google Scholar, Stokes, 2015Stokes M.G. ‘Activity-silent’ working memory in prefrontal cortex: a dynamic coding framework.Trends Cogn. Sci. 2015; 19: 394-405Abstract Full Text Full Text PDF PubMed Scopus (370) Google Scholar). Second, working memory representations are distributed across the brain: they have been observed in prefrontal, parietal, and sensory cortex (for review, see Christophel et al., 2017Christophel T.B. Klink P.C. Spitzer B. Roelfsema P.R. Haynes J.D. The distributed nature of working memory.Trends Cogn. Sci. 2017; 21: 111-124Abstract Full Text Full Text PDF PubMed Scopus (316) Google Scholar). Third, increasing the number of items held in working memory (the “memory load”) increases the overall activity in these brain regions (up to an individual capacity limit; Curtis and D’Esposito, 2003Curtis C.E. D’Esposito M. Persistent activity in the prefrontal cortex during working memory.Trends Cogn. Sci. 2003; 7: 415-423Abstract Full Text Full Text PDF PubMed Scopus (1369) Google Scholar, Ma et al., 2014Ma W.J. Husain M. Bays P.M. Changing concepts of working memory.Nat. Neurosci. 2014; 17: 347-356Crossref PubMed Scopus (640) Google Scholar). However, increasing memory load also reduces the selectivity of individual neurons in a divisive-normalization-like manner; the firing rate of neurons selective for one item is decreased with the addition of other items. This normalization is thought to lead to the reduced memory performance and accuracy at high memory loads (Buschman et al., 2011Buschman T.J. Siegel M. Roy J.E. Miller E.K. Neural substrates of cognitive capacity limitations.Proc. Natl. Acad. Sci. USA. 2011; 108: 11252-11255Crossref PubMed Scopus (189) Google Scholar, Sprague et al., 2014Sprague T.C. Ester E.F. Serences J.T. Reconstructions of information in visual spatial working memory degrade with memory load.Curr. Biol. 2014; 24: 2174-2180Abstract Full Text Full Text PDF PubMed Scopus (105) Google Scholar). Theoretical models have captured some, but not all, of these characteristics. The dominant model of working memory is that recurrent network interactions, either within or between brain regions, give rise to persistent neural activity (Wang, 2001Wang X.J. Synaptic reverberation underlying mnemonic persistent activity.Trends Neurosci. 2001; 24: 455-463Abstract Full Text Full Text PDF PubMed Scopus (754) Google Scholar, Barak and Tsodyks, 2014Barak O. Tsodyks M. Working models of working memory.Curr. Opin. Neurobiol. 2014; 25: 20-24Crossref PubMed Scopus (122) Google Scholar). These models have a limited capacity, as lateral inhibition limits the number of simultaneous patterns of activity that can be maintained (Edin et al., 2009Edin F. Klingberg T. Johansson P. McNab F. Tegnér J. Compte A. Mechanism for top-down control of working memory capacity.Proc. Natl. Acad. Sci. USA. 2009; 106: 6802-6807Crossref PubMed Scopus (247) Google Scholar, Swan and Wyble, 2014Swan G. Wyble B. The binding pool: a model of shared neural resources for distinct items in visual working memory.Atten. Percept. Psychophys. 2014; 76: 2136-2157Crossref PubMed Scopus (65) Google Scholar). However, these models are inflexible. They rely on fine-tuning of connections to embed stable fixed points in the network dynamics specific to the content being stored. These connections must be hardwired or learned for each type of information, and so the network cannot flexibly represent novel, unexpected stimuli. Indeed, networks of this type in the brain seem to encode ecologically relevant information, such as heading direction (Kim et al., 2017Kim S.S. Rouault H. Druckmann S. Jayaraman V. Ring attractor dynamics in the Drosophila central brain.Science. 2017; 356: 849-853Crossref PubMed Scopus (156) Google Scholar). Models that represent working memory as a result of transient dynamics in neural activity are similarly inflexible. They require learning to embed the dynamics, to decode the temporally evolving representations, or to ensure the dynamics are orthogonal to mnemonic representations (Vogels et al., 2005Vogels T.P. Rajan K. Abbott L.F. Neural network dynamics.Annu. Rev. Neurosci. 2005; 28: 357-376Crossref PubMed Scopus (345) Google Scholar, Druckmann and Chklovskii, 2012Druckmann S. Chklovskii D.B. Neuronal circuits underlying persistent representations despite time varying activity.Curr. Biol. 2012; 22: 2095-2103Abstract Full Text Full Text PDF PubMed Scopus (98) Google Scholar, Jaeger, 2002Jaeger H. Short term memory in echo state networks.2002http://www.faculty.jacobs-university.de/hjaeger/pubs/STMEchoStatesTechRep.pdfGoogle Scholar). Other models capture the flexibility of working memory, such as those that hypothesize working memory representations, are encoded in short-term synaptic plasticity or changes in single-cell biophysics (Loewenstein and Sompolinsky, 2003Loewenstein Y. Sompolinsky H. Temporal integration by calcium dynamics in a model neuron.Nat. Neurosci. 2003; 6: 961-967Crossref PubMed Scopus (111) Google Scholar, Hasselmo and Stern, 2006Hasselmo M.E. Stern C.E. Mechanisms underlying working memory for novel information.Trends Cogn. Sci. 2006; 10: 487-493Abstract Full Text Full Text PDF PubMed Scopus (190) Google Scholar, Mongillo et al., 2008Mongillo G. Barak O. Tsodyks M. Synaptic theory of working memory.Science. 2008; 319: 1543-1546Crossref PubMed Scopus (707) Google Scholar). However, these models do not directly explain the limited capacity of working memory and do not capture many of the neurophysiological characteristics of working memory, such as the coexistence of persistent and dynamic representations seen in neural data. Here, we propose a flexible model of working memory that relies on random reciprocal connections to generate persistent activity. As the connections are random, they are inherently untuned with respect to the content being stored and do not need to be learned, allowing the network to maintain any representation. However, this flexibility comes at a cost—when multiple memories are stored in the network, they begin to interfere, resulting in a divisive-normalization-like reduction of responses and imposing a capacity limit on the network. Thus, our model provides a mechanistic explanation for the limited capacity of working memory; it is a necessary trade-off for its flexibility. We model a simplified two-layer network of Poisson spiking neurons (Figure 1A; see STAR Methods for a detailed description). The first layer is the “sensory network” and consists of 8 independent ring-like sub-networks (each with 512 neurons). These sub-networks mimic simplified sensory networks and can be thought of as encoding the identity of independent stimuli at different locations in space. Therefore, we can vary working memory load by varying the number of sensory sub-networks receiving inputs. Neurons within each sensory sub-network are arranged topographically according to selectivity. Position around the ring corresponds to specific values of an encoded feature, such as color or orientation. Consistent with biological observations, connections within a sensory sub-network have a center-surround structure: neurons with similar selectivity share excitatory connections although inhibition is broader (Figure 1A, inset; Kiyonaga and Egner, 2016Kiyonaga A. Egner T. Center-surround inhibition in working memory.Curr. Biol. 2016; 26: 64-68Abstract Full Text Full Text PDF PubMed Scopus (37) Google Scholar, Kim et al., 2017Kim S.S. Rouault H. Druckmann S. Jayaraman V. Ring attractor dynamics in the Drosophila central brain.Science. 2017; 356: 849-853Crossref PubMed Scopus (156) Google Scholar). However, recurrent excitation within each sub-network is too low to maintain memories alone. For simplicity, we first consider a network without connections between sensory sub-networks, although this constraint is relaxed in later models. The second layer is the “random network” (1,024 neurons, a four-fold compression from the sensory network). Neurons in this layer are randomly and reciprocally connected to neurons in the sensory network. Each neuron in the random network has bi-directional excitatory connections with a random subset of neurons in the sensory network (with likelihood γ; here, 0.35). Importantly, all sensory neurons converge onto the same random network. The connections between the sensory and random networks are balanced such that individual neurons receive an equal amount of excitatory and inhibitory drive (i.e., the summed input weight to each neuron from the other network is zero). To achieve this, all pairs of random and sensory neurons without excitatory connections have direct, weak, inhibitory connections (see STAR Methods for details). Such excitation-inhibition balance is consistent with neurophysiological findings (Vogels et al., 2011Vogels T.P. Sprekeler H. Zenke F. Clopath C. Gerstner W. Inhibitory plasticity balances excitation and inhibition in sensory pathways and memory networks.Science. 2011; 334: 1569-1573Crossref PubMed Scopus (379) Google Scholar, Vogels and Abbott, 2005Vogels T.P. Abbott L.F. Signal propagation and logic gating in networks of integrate-and-fire neurons.J. Neurosci. 2005; 25: 10786-10795Crossref PubMed Scopus (310) Google Scholar, Mariño et al., 2005Mariño J. Schummers J. Lyon D.C. Schwabe L. Beck O. Wiesing P. Obermayer K. Sur M. Invariant computations in local cortical networks with balanced excitation and inhibition.Nat. Neurosci. 2005; 8: 194-201Crossref PubMed Scopus (233) Google Scholar). Despite this simple architecture, the network is able to maintain stimulus inputs over extended memory delays (Figure 1B). This is due to the bi-directional and reciprocal connections between the sensory and random networks. Activity in the sensory network feeds forward into the random network, activating a random subset of neurons (Figure 1B, bottom). In turn, neurons from the random network feed back into the sensory network, maintaining activity after the stimulus input is removed (Figure 1B, top; sensory sub-networks 1–4). The reciprocal nature of the connections ensures the synaptic drive fed back into a sensory sub-network from the random network closely matches its own representation (Figure S1A). In this way, the network can flexibly maintain the representation of any input into the sensory network. Neurons from the random network project back to multiple sensory sub-networks. However, activity in the random network does not lead to spuriously sustained representations in other, unstimulated sensory sub-networks (Figure 1B; sub-networks 7 and 8). As feedback connections are random and balanced, they destructively interfere at other locations. In other words, the feedback input from the random network to other sensory sub-networks is orthogonal to their encoding space. Neurons in the sensory network show physiologically realistic tuning curves due to their center-surround architecture (Figures 2A and 2B ). This tuning is effectively inherited by the random network, although with greater complexity (Figures 2C and 2D), matching neurophysiological findings (Funahashi et al., 1989Funahashi S. Bruce C.J. Goldman-Rakic P.S. Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortex.J. Neurophysiol. 1989; 61: 331-349Crossref PubMed Scopus (1966) Google Scholar, Zaksas and Pasternak, 2006Zaksas D. Pasternak T. Directional signals in the prefrontal cortex and in area MT during a working memory for visual motion task.J. Neurosci. 2006; 26: 11726-11742Crossref PubMed Scopus (165) Google Scholar, Mendoza-Halliday et al., 2014Mendoza-Halliday D. Torres S. Martinez-Trujillo J.C. Sharp emergence of feature-selective sustained activity along the dorsal visual pathway.Nat. Neurosci. 2014; 17: 1255-1262Crossref PubMed Scopus (139) Google Scholar). However, as connectivity is random, the tuning of neurons in the random network is not consistent across inputs to different sensory networks (Figure 2E). This leads to neurons in the random network showing “linear conjunctive” coding (Figure 2F), preferring different input values for different sensory sub-networks (e.g., different colors at different locations) and responding to unique combinations of inputs into multiple sensory sub-networks. Such linear conjunctive representations are consistent with experimental observations in prefrontal cortex (Fusi et al., 2016Fusi S. Miller E.K. Rigotti M. Why neurons mix: high dimensionality for higher cognition.Curr. Opin. Neurobiol. 2016; 37: 66-74Crossref PubMed Scopus (281) Google Scholar, Lindsay et al., 2017Lindsay G.W. Rigotti M. Warden M.R. Miller E.K. Fusi S. Hebbian learning in a random network captures selectivity properties of the prefrontal cortex.J. Neurosci. 2017; 37: 11021-11036Crossref PubMed Scopus (25) Google Scholar). Multiple memories can be stored in the network simultaneously (Figure 1B). For a few items (typically ≤3), memories do not significantly interfere—there is sufficient space in the high-dimensional random network for maintaining multiple patterns. However, as the number of sensory inputs is increased, interference in the random network also increases, eventually causing memory failures (Figure 1B; sub-networks 5 and 6). Figure 3A shows the percentage of correct memories at the end of the delay period as a function of load (i.e., number of items presented; see STAR Methods for details). This closely matches behavioral results in both humans and monkeys (Buschman et al., 2011Buschman T.J. Siegel M. Roy J.E. Miller E.K. Neural substrates of cognitive capacity limitations.Proc. Natl. Acad. Sci. USA. 2011; 108: 11252-11255Crossref PubMed Scopus (189) Google Scholar, Cowan, 2010Cowan N. The magical mystery four: how is working memory capacity limited, and why?.Curr. Dir. Psychol. Sci. 2010; 19: 51-57Crossref PubMed Scopus (637) Google Scholar). Indeed, the decrease in performance with working memory load in the model was highly correlated with experiments (r = 0.97; p = 0.0054; to experimental data from Luck and Vogel, 1997Luck S.J. Vogel E.K. The capacity of visual working memory for features and conjunctions.Nature. 1997; 390: 279-281Crossref PubMed Scopus (2848) Google Scholar; see STAR Methods for details). Also consistent with behavior, the speed of forgetting during the delay period increased with load (Figure 3B). As with all of our results, this is not due to tuning of network parameters; parameters were set to maximize the total number of items remembered across all loads, not to match any behavioral or electrophysiological results (see STAR Methods). Errors in working memory increased with the number of items held in memory (Figures 3C and S1B). Errors also accumulated more quickly at higher loads (Figure 3D). The increase in error was not just due to forgetting of inputs—even when an input was successfully maintained, memory error increased with memory load (Figures 3E and 3F). This is consistent with behavioral studies that have shown increases in memory error, even below working memory capacity limits (Rademaker et al., 2018Rademaker R.L. Park Y.E. Sack A.T. Tong F. Evidence of gradual loss of precision for simple features and complex objects in visual working memory.J. Exp. Psychol. Hum. Percept. Perform. 2018; 44: 925-940Crossref PubMed Scopus (26) Google Scholar, Bays et al., 2009Bays P.M. Catalao R.F. Husain M. The precision of visual working memory is set by allocation of a shared resource.J. Vis. 2009; 9: 1-11Crossref PubMed Scopus (622) Google Scholar, Adam et al., 2017Adam K.C.S. Vogel E.K. Awh E. Clear evidence for item limits in visual working memory.Cognit. Psychol. 2017; 97: 79-97Crossref PubMed Scopus (99) Google Scholar; but see Pertzov et al., 2017Pertzov Y. Manohar S. Husain M. Rapid forgetting results from competition over time between items in visual working memory.J. Exp. Psychol. Learn. Mem. Cogn. 2017; 43: 528-536Crossref PubMed Scopus (56) Google Scholar). Indeed, the load-dependent decrease in memory accuracy was highly correlated between the model and experiments (r = 0.997; p = 0.00205; experimental data from Ma et al., 2014Ma W.J. Husain M. Bays P.M. Changing concepts of working memory.Nat. Neurosci. 2014; 17: 347-356Crossref PubMed Scopus (640) Google Scholar). Together, these results show how our model bridges the gap between “discrete” models, where memories are completely forgotten, and “continuous” models, where interference between memories decreases their accuracy. To directly test the model’s ability to generalize across these results, we fit the model parameters to match either memory performance or memory accuracy (see STAR Methods for details). Models fit on one dataset generalized to capture the other dataset (Figures S1D and S1E; note, this is the only model fit directly to experimental data). Our model provides a simple mechanistic explanation for the limited capacity of working memory: it is due to interference in neural representations in the shared random network. This is a natural consequence of the convergent, random connectivity between the two non-linear networks; as multiple inputs are presented to the sensory networks, their representations interfere in the random network, disrupting maintenance. This is an unavoidable consequence of the convergence and does not depend on network parameters (as we show below). In this way, our model suggests capacity limits are a necessary trade-off for the flexibility of working memory. The model makes specific predictions about how neural activity should change as more items are held in working memory. First, increasing the number of stimulus inputs increases the overall average firing rate in the random network, saturating at the capacity limit of the network (∼3 or 4 items; Figure 4A). This is consistent with experimental observations of gross activity levels in prefrontal and parietal cortex; both blood-oxygen-level-dependent (BOLD) and evoked potentials increase with working memory load, saturating at an individual’s capacity limit (Curtis and D’Esposito, 2003Curtis C.E. D’Esposito M. Persistent activity in the prefrontal cortex during working memory.Trends Cogn. Sci. 2003; 7: 415-423Abstract Full Text Full Text PDF PubMed Scopus (1369) Google Scholar, Ma et al., 2014Ma W.J. Husain M. Bays P.M. Changing concepts of working memory.Nat. Neurosci. 2014; 17: 347-356Crossref PubMed Scopus (640) Google Scholar). Again, the model was highly correlated with experiments (r = 0.998; p = 1.97 × 10−6; experimental data from Ma et al., 2014Ma W.J. Husain M. Bays P.M. Changing concepts of working memory.Nat. Neurosci. 2014; 17: 347-356Crossref PubMed Scopus (640) Google Scholar). Second, the model predicts maintaining multiple memories will reduce the response of selective neurons in the random network (Figure 4B). This is consistent with experimental observations, which have shown divisive-normalization-like regularization of mnemonic responses in single neurons and across the population (Buschman et al., 2011Buschman T.J. Siegel M. Roy J.E. Miller E.K. Neural substrates of cognitive capacity limitations.Proc. Natl. Acad. Sci. USA. 2011; 108: 11252-11255Crossref PubMed Scopus (189) Google Scholar, Sprague et al., 2014Sprague T.C. Ester E.F. Serences J.T. Reconstructions of information in visual spatial working memory degrade with memory load.Curr. Biol. 2014; 24: 2174-2180Abstract Full Text Full Text PDF PubMed Scopus (105) Google Scholar). Our model provides a potential circuit mechanism for such divisive-normalization-like regularization, suggesting it is the result of balanced excitation-inhibition between networks. The low fraction of connectivity (γ) and balanced excitation-inhibition means that a neuron in the random network that is selective for one stimulus is more likely to be inhibited than excited by a second stimulus (see STAR Methods). Thus, the response of selective neurons is reduced as items are added to memory (Figure 4B). This effect can be seen across the population. Figure 4C shows the relative response of neurons in the random network to two stimuli either presented separately (x axis) or together (y axis). As noted above, the overall average response increases when two stimuli are presented. However, the response to the pair of stimuli was not a summation of the response to each stimulus alone; rather, it was a sublinear mixture (slope is ∼0.50). This is consistent with electrophysiological results during perception (Reynolds et al., 1999Reynolds J.H. Chelazzi L. Desimone R. Competitive mechanisms subserve attention in macaque areas V2 and V4.J. Neurosci. 1999; 19: 1736-1753Crossref PubMed Google Scholar) and with divisive normalization models that predict the response to two stimuli should be the average of the response to each stimulus alone (resulting in a slope of 0.5). Divisive normalization has been observed in many cognitive domains (Carandini and Heeger, 1994Carandini M. Heeger D.J. Summation and division by neurons in primate visual cortex.Science. 1994; 264: 1333-1336Crossref PubMed Scopus (459) Google Scholar, Carandini and Heeger, 2011Carandini M. Heeger D.J. Normalization as a canonical neural computation.Nat. Rev. Neurosci. 2011; 13: 51-62Crossref PubMed Scopus (1082) Google Scholar). Indeed, in working memory, divisive-normalization-like regularization may explain the observed decrease in stimulus information with memory load: it reduces selectivity, which reduces information (Figure 4D; Buschman et al., 2011Buschman T.J. Siegel M. Roy J.E. Miller E.K. Neural substrates of cognitive capacity limitations.Proc. Natl. Acad. Sci. USA. 2011; 108: 11252-11255Crossref PubMed Scopus (189) Google Scholar, Sprague et al., 2014Sprague T.C. Ester E.F. Serences J.T. Reconstructions of information in visual spatial working memory degrade with memory load.Curr. Biol. 2014; 24: 2174-2180Abstract Full Text Full Text PDF PubMed Scopus (105) Google Scholar). More broadly, our model suggests excitation-inhibition balance within a convergent, non-linear network could be a circuit mechanism for implementing divisive normalization. Interference between memory representations also provides a simple mechanistic explanation for the reduced memory precision at higher memory loads (Ma et al., 2014Ma W.J. Husain M. Bays P.M. Changing concepts of working memory.Nat. Neurosci. 2014; 17: 347-356Crossref PubMed Scopus (640) Google Scholar, Bays and Husain, 2008Bays P.M. Husain M. Dynamic shifts of limited working memory resources in human vision.Science. 2008; 321: 851-854Crossref PubMed Scopus (788) Google Scholar). In our model, memory representations drift over time, due to the accumulation of noise from Poisson variability in neural spiking (Burak and Fiete, 2012Burak Y. Fiete I.R. Fundamental limits on persistent activity in networks of noisy neurons.Proc. Natl. Acad. Sci. USA. 2012; 109: 17645-17650Crossref PubMed Scopus (61) Google Scholar). The increased interference in the random network at higher memory loads leads to weaker feedback, reducing the representation in the sensory network. This reduction both allows noise to have a greater impact, causing an increase in drift (Figures 3C–3F), and impairs decoding of memory representations (Bays, 2014Bays P.M. Noise in neural populations accounts for errors in working memory.J. Neurosci. 2014; 34: 3632-3645Crossref PubMed Scopus (130) Google Scholar, Bays, 2015Bays P.M. Spikes not slots: noise in neural populations limits working memory.Trends Cogn. Sci. 2015; 19: 431-438Abstract Full Text Full Text PDF PubMed Scopus (81) Google Scholar). Given the model’s prediction that interference impairs memory performance, reducing interference between memories should improve working memory performance and accuracy. Indeed, how much two stimulus representations overlapped in the random network strongly determined the ability to accurately maintain both memories. Figure 5A shows memory performance was impaired when two memories were less correlated, reflecting greater interference between the memories in the random network (see also Figure S1C). Conversely, more correlated memories were better remembered. Surprisingly, errors in memory accumulated over time in a way that reduced interference between memories (Figure 5B). Both of these predictions are testable hypotheses that could be addressed with future electrophysiology or imaging experiments. The model captures several more key electrophysiological findings related to working memory. First, we observe persistent mnemonic activity, consistent with electrophysiological results in both monkeys and humans (Funahashi et al., 1989Funahashi S. Bruce C.J. Goldman-Rakic P.S. Mnemonic coding of visual space in the monkey’s dorsolateral prefrontal cortex.J. Neurophysiol. 1989; 61: 331-349Crossref PubMed Scopus (1966) Google Scholar, Fuster, 1973Fuster J.M. Unit activity in prefront" @default.
• W2950376067 created "2019-06-27" @default.
• W2950376067 creator A5028333846 @default.
• W2950376067 creator A5055391576 @default.
• W2950376067 date "2019-07-01" @default.
• W2950376067 modified "2023-10-06" @default.
• W2950376067 title "A Flexible Model of Working Memory" @default.
• W2950376067 cites W1495865139 @default.
• W2950376067 cites W1495874871 @default.
• W2950376067 cites W1503270516 @default.
• W2950376067 cites W1503602982 @default.
• W2950376067 cites W1635648989 @default.
• W2950376067 cites W1747565446 @default.
• W2950376067 cites W1797750125 @default.
• W2950376067 cites W1877812147 @default.
• W2950376067 cites W1934970405 @default.
• W2950376067 cites W1939596644 @default.
• W2950376067 cites W1967950674 @default.
• W2950376067 cites W1971300091 @default.
• W2950376067 cites W1973327463 @default.
• W2950376067 cites W1985392528 @default.
• W2950376067 cites W1986028388 @default.
• W2950376067 cites W1988013646 @default.
• W2950376067 cites W1989039239 @default.
• W2950376067 cites W1996242427 @default.
• W2950376067 cites W1998570540 @default.
• W2950376067 cites W2001711642 @default.
• W2950376067 cites W2001753487 @default.
• W2950376067 cites W2005605159 @default.
• W2950376067 cites W2009759904 @default.
• W2950376067 cites W2013142139 @default.
• W2950376067 cites W2013993313 @default.
• W2950376067 cites W2014350866 @default.
• W2950376067 cites W2014542273 @default.
• W2950376067 cites W2018111368 @default.
• W2950376067 cites W2033708181 @default.
• W2950376067 cites W2038718814 @default.
• W2950376067 cites W2041601616 @default.
• W2950376067 cites W2049517487 @default.
• W2950376067 cites W2054768221 @default.
• W2950376067 cites W2065286706 @default.
• W2950376067 cites W2068949971 @default.
• W2950376067 cites W2071461458 @default.
• W2950376067 cites W2072804750 @default.
• W2950376067 cites W2076919983 @default.
• W2950376067 cites W2089242029 @default.
• W2950376067 cites W2092348893 @default.
• W2950376067 cites W2099231026 @default.
• W2950376067 cites W2100800904 @default.
• W2950376067 cites W2101199693 @default.
• W2950376067 cites W2103179919 @default.
• W2950376067 cites W2103780005 @default.
• W2950376067 cites W2107150897 @default.
• W2950376067 cites W2108722540 @default.
• W2950376067 cites W2118706537 @default.
• W2950376067 cites W2123745704 @default.
• W2950376067 cites W2124196462 @default.
• W2950376067 cites W2125728075 @default.
• W2950376067 cites W2132637814 @default.
• W2950376067 cites W2134598164 @default.
• W2950376067 cites W2135955567 @default.
• W2950376067 cites W2151137320 @default.
• W2950376067 cites W2152868521 @default.
• W2950376067 cites W2154934134 @default.
• W2950376067 cites W2159749299 @default.
• W2950376067 cites W2159927097 @default.
• W2950376067 cites W2166507263 @default.
• W2950376067 cites W2167809052 @default.
• W2950376067 cites W2168462263 @default.
• W2950376067 cites W2170319235 @default.
• W2950376067 cites W2170478537 @default.
• W2950376067 cites W2176033094 @default.
• W2950376067 cites W2195449819 @default.
• W2950376067 cites W2199584684 @default.
• W2950376067 cites W2212384750 @default.
• W2950376067 cites W2255581062 @default.
• W2950376067 cites W2284461153 @default.
• W2950376067 cites W2297363866 @default.
• W2950376067 cites W2400660831 @default.
• W2950376067 cites W2526554353 @default.
• W2950376067 cites W2562471898 @default.
• W2950376067 cites W2568841898 @default.
• W2950376067 cites W2569798719 @default.
• W2950376067 cites W2610861664 @default.
• W2950376067 cites W2618001251 @default.
• W2950376067 cites W2618100165 @default.
• W2950376067 cites W2738942526 @default.
• W2950376067 cites W2793490979 @default.
• W2950376067 cites W2809542406 @default.
• W2950376067 cites W2952820172 @default.
• W2950376067 cites W2953125824 @default.
• W2950376067 doi "https://doi.org/10.1016/j.neuron.2019.04.020" @default.
• W2950376067 hasPubMedCentralId "https://www.ncbi.nlm.nih.gov/pmc/articles/6613943" @default.
• W2950376067 hasPubMedId "https://pubmed.ncbi.nlm.nih.gov/31103359" @default.
• W2950376067 hasPublicationYear "2019" @default.
• W2950376067 type Work @default.
• W2950376067 sameAs 2950376067 @default.
• W2950376067 citedByCount "100" @default.

• next