Report

Synaptic Theory of Working Memory

See allHide authors and affiliations

Science  14 Mar 2008:
Vol. 319, Issue 5869, pp. 1543-1546
DOI: 10.1126/science.1150769

Abstract

It is usually assumed that enhanced spiking activity in the form of persistent reverberation for several seconds is the neural correlate of working memory. Here, we propose that working memory is sustained by calcium-mediated synaptic facilitation in the recurrent connections of neocortical networks. In this account, the presynaptic residual calcium is used as a buffer that is loaded, refreshed, and read out by spiking activity. Because of the long time constants of calcium kinetics, the refresh rate can be low, resulting in a mechanism that is metabolically efficient and robust. The duration and stability of working memory can be regulated by modulating the spontaneous activity in the network.

Working memory (WM) enables the temporary holding of information for processing purposes, playing a crucial role in the execution of a wide range of cognitive tasks (1). In the delayed-response paradigm, a stimulus that is briefly presented to an animal has to be kept for several seconds until the execution of a task. Enhanced, stimulus-specific spiking activity has been observed during this delay period and is considered to be a neuronal correlate of WM (25). The current theoretical framework holds that delay activity emerges either from intrinsic cell properties (6, 7) or as persistent reverberations in selective neural populations coding for different memories (812). These populations are formed during learning via long-term synaptic modifications (13). However, electrophysiological studies have shown that the delay activity increase can be very modest (14, 15), sometimes disappearing completely during part of the delay period (16). These observations suggest that WM might not reside entirely in the spiking activity. Furthermore, holding information in a spiking form is energetically expensive because of the high metabolic cost of action potentials (17). Here, we present an alternative account based on properties of excitatory synaptic transmission in the prefrontal cortex (PFC) (18). The PFC is a cortical area implicated in WM (4), and excitatory synaptic transmission in this area can be markedly facilitatory, unlike sensory areas where it is mostly depressing (19, 20). We therefore propose that an item is maintained in the WM state by short-term synaptic facilitation mediated by increased residual calcium levels at the presynaptic terminals of the neurons that code for this item (21). Because removal of residual calcium from presynaptic terminals is a relatively slow process (22, 23), the memory can be transiently held for about 1 s without enhanced spiking activity.

We implemented this mechanism with a recurrent network of integrate-and-fire neurons (24). The network encodes a set of memories (items) by randomly composed selective populations of excitatory neurons (Fig. 1B). Connections between the neurons coding for the same memory are stronger than connections between different populations, mimicking the result of prior long-term Hebbian learning (25) or intrinsic clustering of recurrent connections (26). Inhibitory neurons are connected to the excitatory ones in a nonstructured way, resulting in competition between different memories [see supporting online material (SOM)]. All excitatory-to-excitatory connections display facilitating transmission, described by a phenomenological model of short-term plasticity (20, 27). Synaptic efficacy is modulated by the amount of available resources (x, normalized so that 0 < x < 1) and the utilization parameter (u) that defines the fraction of resources used by each spike, reflecting the residual calcium level (22, 23) (Fig. 1A and SOM). Upon a spike, an amount ux of the available resources is used to produce the postsynaptic current, thus reducing x. This process mimics neurotransmitter depletion. The spike also increases u, mimicking calcium influx into the presynaptic terminal and its effects on release probability. Between spikes, x and u recover to their baseline levels (x = 1 and u = U) with time constants τD (depressing) and τF (facilitating), respectively. The phenomenological model reproduces the behavior of cortical synapses, both depressing (τD > τF) and facilitating (τF > τD) (27). For PFC facilitating excitatory connections, the experimental fit reports τF ≫ τD (18), with τF on the order of 1 s.

Fig. 1.

Physiology and anatomy of the network. (A) Short-term synaptic plasticity model. (Left) Kinetic scheme with the corresponding equations for synaptic variables. δ, Dirac delta function; tsp, time of presynaptic spike. (Right) Example of the postsynaptic response to a train of presynaptic action potentials in the case of a facilitating connection. During the train, u increases (facilitation) and x decreases (depression). Synaptic efficacy is modulated by the product ux. vm, membrane potential. (B) Network architecture. Colored triangles are excitatory neurons that code for different memories. Black open triangles are nonselective excitatory neurons. Black circles are inhibitory neurons with nonstructured connections to the entire network.

The simulations begin with loading one item into WM by providing transient external excitation to the corresponding neural population (Fig. 2A). The population activity increases for the duration of the input, changing the internal state of the synaptic connections. The connections are both depressed (reduced x) and facilitated (increased u), with depression dominant on the time scale of τD and facilitation dominant on the time scale of τF (where τD = 0.2 s and τF = 1.5 s; see SOM for all parameter values). As long as the synapses remain facilitated, the memory can be reactivated by presenting a weak nonspecific excitatory input to the whole network (gray shading), even though the neural activity is at the spontaneous level. Reactivation is expressed as a short epoch of synchronized activity [“population spike” (PS)], where almost every neuron in the population fires a spike within an interval of about 20 ms (28, 29). Even though the reactivating signal is nonspecific (that is, it uniformly targets all the neurons), the network response is memory-specific: The neurons coding for the loaded item produce a PS; the others stay at baseline activity level. The PS also refreshes the memory by producing additional facilitation, thus enabling subsequent memory reactivations. In the absence of reactivating signals, the memory fades away over a time scale on the order of τF.

Fig. 2.

Memory maintenance with synaptic facilitation. (A) The item is loaded into the memory by activating the corresponding population at time t = 0 (dark shading). A nonspecific read-out signal is applied to the entire excitatory network as explained in the text (gray shading), leading to the selective reactivation of the target population via the PS. Black and green dots, spike rasters from a subset (10%) of neurons from the targeted and one of the nontargeted populations, respectively; red curve, average value of x in the synaptic connections of the target population; blue curve, average value of u for the same synapses. (B) Same as (A), with an increased background input. The target population reactivates spontaneously with a set of PSs. Red arrow indicates termination of excitation. (C) A further increase in background input leads to WM with asynchronous elevated firing in the target population. (Right) Corresponding histograms for (A) to (C) of the difference in firing rate between the delay period and the spontaneous state for different neurons in the target population [note the different x axis in (A)]. The delay period is defined in (A) as the interval after the termination of the selective input until the onset of the read-out signal and in (B) and (C) as the period until the decrease of external excitation.

In the above scenario, the network has a single stable activity state corresponding to the spontaneous activity, thus appropriately timed external signals are required to extract the memory from synaptic to spiking form. A more persistent form of WM requires the selective population to exhibit a bistable activity regime, where the spontaneous state coexists with another stable state (8). Our network can be forced into this regime by increasing spontaneous activity by means of a global nonspecific background input (see SOM for the mathematical analysis). Accordingly, we simulated the network for increasing levels of background input. In the bistable regime, PSs become persistent without reactivating inputs (Fig. 2B). Each reactivation increases u and decreases x, the latter terminating the PS. The time between subsequent PSs is controlled by the recovery from synaptic depression so that the PSs tend to occur with a period on the order of τD. With a τD compatible with (18), this would correspond to cortical oscillations in the theta-range, as observed during WM experiments (30, 31). Because τF ≫ τD, the decay of the utilization factor is balanced by the increase produced by the PSs, so that u remains at sufficiently high levels for subsequent PSs to emerge. Persistent PSs can be terminated by reducing background input, thus restoring the network to the transient regime. A different model of persistent PSs is based on the increase in asynchronous transmitter release (32). In the simulation presented in Fig. 2B, neurons coding for a given memory exhibit highly coherent firing during the PSs. In more realistic networks, PSs could be broader and consist of random subpopulations of neurons, resulting in less pronounced synchrony (fig. S2). If nonspecific background input increases further, the network exhibits bistability between a spontaneous state and an asynchronous state of enhanced firing rate (Fig. 2C). Information about the memory is maintained in both synaptic and spiking form. This regime exists for sufficiently strong recurrent connections, which could possibly result from extensive learning.

The use of residual calcium at synaptic terminals as a memory “buffer” requires low emission rates (Fig. 2, histograms). Moreover, the buffer content is not substantially affected by the neural activity in the rest of the cortex. When we presented the network with a noisy input that targeted a random subset of excitatory neurons for a brief duration, the increased firing of neurons receiving the input suppressed the memory-related spiking activity (Fig. 3, teal shading). The information, however, remained in the increased utilization factor in the memory population; hence, the spiking activity resumed after the termination of the suppressing input. The same feature enables the network to keep multiple items simultaneously with interleaved PSs (33). We illustrated this possibility in two different conditions: (i) when the network has a single activity state and PSs result from a sequence of reactivating signals (Fig. 3A) or (ii) when the network exhibits persistent PSs (Fig. 3B). When a new item is presented, the previous memory is temporarily suppressed (dark shading), after which the network maintains both memories by subsequent reactivations of the two populations. When the network is in the regime of externally generated PSs, two-item WM results in the same oscillation frequency in the global activity as that of the one-item memory (30), whereas in the regime of persistent PSs, two-item memory results in higher global frequency. In more realistic implementations of the model, an increase in frequency may be less pronounced also in the persistent regime (fig. S2).

Fig. 3.

Robustness to noise and two-item memory. The first item is loaded into memory at t = 0 (dark shading). The second item is loaded into memory at t = 2.7 s. Teal shading indicates a random nonspecific input to 15% of the excitatory neurons. (A) Periodic sequence of nonspecific external inputs is used to refresh the memory (gray shading). (B) Persistent PSs. Dots, rasters of 10% of the first (0 to 79) and second (80 to 159) populations' neurons; red and blue curves, same as in Fig. 2.

Consequently, we propose that WM can be maintained by short-term synaptic facilitation. Accumulation of residual calcium in the pre-synaptic terminals could carry the information about the recalled memory in a working form, reducing the need for metabolically costly action potentials. The memories are transformed into spiking activity, either as a result of global reactivating input to the network or by virtue of the intrinsic network dynamics. Not all encountered stimuli enter WM, and we thus expect the basal modality of the network to be the transient one. The decision to allow items into WM is mediated by attention, which we suggest is represented by the global excitatory input, either in tonic or oscillating mode. The performance of human observers on memory tasks is positively correlated with the level of neural activity during the presentation of the items (34).

The model predicts that residual calcium at the presynaptic terminals should be tonically enhanced during WM, even when there is no noticeable increase in the firing rate. This prediction is in contrast to the model of (7) where WM is mediated by propagating calcium wavefronts along dendritic processes. Suppressing the spiking activity for a period of several hundred milliseconds (35) should still allow for the memory to reactivate after the suppressing input is withdrawn. We also expect that groups of neurons could exhibit brief epochs of coherent firing. The model provides a possible target for a pharmacological interference with WM. In particular, manipulations that modify the facilitation/depression balance in the memory-related cortical areas (36) are predicted to have a strong effect on the stability and duration of memory.

Supporting Online Material

www.sciencemag.org/cgi/content/full/319/5869/1543/DC1

SOM Text

Figs. S1 and S2

Table S1

References

References and Notes

View Abstract

Navigate This Article