Mean-Field Theory of Irregularly Spiking Neuronal Populations and Working Memory in Recurrent Cortical Networks - Computational Neuroscience: A Comprehensive Approach - page 462

15.3.4 Persistent neural activity in an object working memory model

As we have just demonstrated, the self-consistency equations whose solutions pro-
vide the firing rates of the different neural populations in the steady-states of the
recurrent network can, in some cases, have multiple solutions for the same set of
parameters and external inputs to the network. When this is the case, transient exter-
nal inputs can switch the state of the network among its possible stationary solutions.
Conceptually, the network can now function as a short-term or working memory sys-
tem, as its state of activity is no longer uniquely specified by the 'static' variables of
the system (cellular or network parameters, unspecific external inputs, etc) but also
carries information about the recent history of transient inputs to the network. More
generally, a recurrent network can display multi-stability, whereby a resting state of
spontaneous activity coexists with multiple attractor states (stable neural firing pat-
terns), each of which encodes a different sensory stimulus. Therefore, the identity
of a transient input is encoded and stored in the level of spiking activity of a distinct
neural assembly in the recurrent circuit. Such stimulus-selective persistent activity
has been documented during electrophysiological experiments on behaving monkeys
during working memory tasks [35, 41, 42, 43, 44, 45, 46, 47, 48, 55, 56, 73, 79, 81,
85, 86, 89, 100, 124].
We now describe in some detail an object working memory model that has been
analyzed both at the mean-field level and with numerical simulations. For the model
to comply with the basic phenomenology of the data from object working memory
experiments, the simple bistable network presented in the previous section has to be
considerably enlarged. First, local networks in association cortices are likely to be
endowed with much more than two attractors. The experiments in the temporal lobe
with a large number of stimuli (up to 100) [81, 85, 86, 100] suggest the following
picture:

• In the absence of external stimulation, networks in the temporal lobe are in a

spontaneous activity state, in which all neurons fire at low levels of several Hz;

• Upon presentation of a particular familiar stimulus, a small sub-population of

neurons in localized areas of the temporal lobe exhibit persistent activity; this
fraction of neurons can be estimated to be around 1% or a few % [81]. Thus,
the representation of familiar stimuli in these areas is sparse;

• Representations of different stimuli have very small overlaps, since neurons

typically respond to only one or a few images in the set of shown images [81].

A model for object working memory based on these observations has been built in
several stages [9, 11, 12, 21]. The model of [11] is a network of randomly connected
excitatory and inhibitory neurons much as the one discussed in Section 15.3.1.3.
In addition, the excitatory population is divided in sub-populations, where a given
sub-population is assumed to have a strong visual response to a particular stimulus.
A schematic representation of the architecture of the model is shown in Fig-
ure 15.6A. The network consists of two large pools of interacting pyramidal cells
and interneurons. Both populations are fully connected with themselves and with