data on short-term memory in topologically different parts of the PFC during delay tasks having the what-then-where design. The model contained different populations of neurons (as found experimentally) in attractor networks that responded in the delay period to the stimulus object, the stimulus position, and to combinations of both object and position information. These neuronal populations were arranged hierarchically and global inhibition mediated through inhibitory interneurons was employed to implement competition. The relative activity of the different attractor populations required to perform what-then-where and where-then-what short-term memory tasks was provided by an external attentional signal that biases the different neuron populations within the framework of the biased competition model of attention (Desimone & Duncan 1995; Rolls & Deco 2002).
It was shown that their model could account for the neurophysiological activity seen in both the ventrolateral and dorsolateral PFC during the delay periods of working memory tasks utilizing the what-then-where design, obtained by Rao et al. (1997). Furthermore, the Deco et al. model generated simulated fMRI patterns that matched experimental findings during a what-then-where short-term memory task for both PFC sectors as shown by the fMRI findings of Postle & D'Esposito (1999). However, this could not be done if it was assumed that the difference between ventrolateral and dorsolat-eral PFC followed the organization-by-stimulus-domain hypothesis, with the dorsal PFC being specialized for spatial processing and ventral PFC being specialized for object processing. Rather, Deco et al. (2004) had to assume that the differences between these two prefrontal regions arose from having a larger amount of inhibition in the dorsolateral portion of the PFC than in the ventrolateral part.
The Deco et al (2004) study is important for several reasons. First, it is the first large-scale neural model applied to fMRI data that employed spiking neurons. Second, it demonstrates how this type of modeling can shift the terms of a scientific debate. There is great deal of controversy concerning the different functional roles of dorsal and ventral PFC. Heretofore, the controversy was cast in cognitive terms (i.e., spatial vs. object processing on the one hand, ^-
Fig. 5. Performance of the auditory model of Husain et al. (2004) for the auditory continuity illusion. Above-threshold activity of 5 or more neurons in the response module (FR) indicates a match between the two stimuli of a DMS trial. The top graph shows that with a short duration gap in the second stimulus, the model (like actual subjects) indicates a match, thus grouping the parts of the tonal contour into a perceptual whole. As the gap widens, a non-match results. If noise is inserted in the gap, and is of weak intensity (green), the tonal contour is not considered as continuing through the gap (bottom); if the noise is more intense (red), then perceptual grouping occurs, but only if the band of noise is in the part of frequency space occupied by the tonal contour. See Husain et al. (2005) for details. Modified from Horwitz & Glabus (2005)
manipulation vs. maintenance on the other). The results of the Deco et al. study demonstrate that one can rephrase the debate in neural terms (i.e., different neuronal populations compared to different levels of inhibition), and in these terms, a variety of neuroscientific results can be brought forward to resolve the issue.
The next study we present comes from Chadderon and Sporns (in press) and it also focuses on prefrontal cortex. A large-scale computational model of prefrontal cortex and associated brain regions was constructed. It was designed to investigate the mechanisms by which working memory and task state interact to select adaptive behaviors from a behavioral repertoire. The model consisted of multiple brain regions containing neuronal populations with realistic physiological and anatomical properties: extrastriate visual cortical regions, inferotemporal cortex, prefrontal cortex, striatum, and midbrain dopamine neurons. Like the visual model of Tagamets and Horwitz (1998) discussed earlier, the Chadderon-Sporns model used Wilson-Cowan leaky integrator neurons (Wilson & Cowan 1972).
In the Chadderon-Sporns model the onset of a delayed match-to-sample or delayed non-match-to-sample task triggers tonic dopamine release in pre-frontal cortex, which causes a switch into a persistent, stimulus-insensitive dynamic state that promotes the maintenance of stimulus representations within prefrontal networks. Other modeled prefrontal and striatal units select cognitive acceptance or rejection behaviors according to which task is active and whether prefrontal working memory representations match the current stimulus. Working memory task performance and memory fields of prefrontal delay units were degraded by extreme elevation or depletion of tonic dopamine levels. Analyses of cellular and synaptic simulated activity indicated that hyponormal dopamine levels resulted in increased prefrontal activation, whereas hyper-normal dopamine levels led to decreased activation.
Chadderon and Sporns also used their simulated results to derive synthetic fMRI signals, in a similar manner to that discussed earlier (Horwitz & Tagamets 1999). They found that under normal dopamine conditions, there was a significant increase in PFC fMRI activity in a DMS working memory task, as compared to an "idle" control condition. If a relatively fast hemodynamic delay function was used to derive fMRI signals, the increase was confined to the delay periods of the task, and absent during the cue/distractor/target periods. Decreasing tonic dopamine levels led to higher baseline activation of PFC, but the activity differences between idle and working memory conditions were not significant. Conversely, if tonic dopamine levels were elevated, baseline activation of PFC was reduced, and activity during a working memory task was decreased with respect to corresponding control trials, thus showing a reversal of the increase found under normal dopamine levels.
The Chadderon and Sporns model (Chadderdon & Sporns in press) represents an important step forward over the previous models we discussed in that it explicitly incorporates a modulatory neurotransmitter (dopamine) so that more complex behavior can be addressed. In particular, they implemented two separate domains of prefrontal working memory. Task identity (i.e., delayed match-to-sample, delayed nonmatch-to-sample and their corresponding control tasks) was maintained by a segregated set of recurrently excitatory and mutually inhibitory cell populations. Stimulus feature memory was maintained by tonic dopamine level. It seems likely that in the near future, the level of model complexity will increase even further.
The final example we present in this subsection involves a model that simulates EEG/MEG dynamics. As mentioned in the introductory section, EEG/MEG activity has engendered a number of computational neural modeling efforts (e.g., David & Friston 2003; Jirsa & Haken 1997; May et al. 1999; Nunez 1981; Robinson et al. 2005). Here, we shall discuss a recent study by David et al. (2005) that describes a neurally plausible forward model designed to reproduce responses seen empirically, thus allowing mechanistic enquiries into the generation of evoked and induced EEG/MEG responses to be made. As these authors point out, most neural models of EEG/MEG were constructed to generate alpha rhythms (e.g., Jansen & Rit 1995; Stam et al. 1999), but recent work has shown that models that produce the entire spectrum of EEG/MEG oscillations can be created (e.g., David & Friston 2003; Robinson et al. 2001). The model developed by David and colleagues (David et al. 2005) focuses on simulating event-related activity.
One important aspect of the David et al. model is that the basic functional unit, based on the work of Jansen and Rit (1995), uses three neuronal subpopulations to represent a cortical area; one subpopulation corresponds to pyramidal neurons, a second represents excitatory interneurons and the third corresponds to inhibitory interneurons. A second important feature is that their neural mass model consists of hierarchically arranged areas using three kinds of inter-area connections (forward, backward and lateral). The excitatory interneurons can be considered to be spiny stellate cells found primarily in layer 4 of the cortex; they receive forward connections. The excitatory pyramidal neurons are the output cells of a cortical column, and are found in the agranular layers, as are the inhibitory interneurons. The MEG/EEG signal is taken to be a linear mixture of the averaged depolarization of the pyramidal neurons.
Using this model, David et al. (2005) investigated how responses, at each level of a cortical hierarchy, depended on the strength of connections. They did this in the context of deterministic responses and then with stochastic spontaneous activity. One important result of their simulations was that with the presence of spontaneous activity, evoked responses could arise from two distinct mechanisms: (1) for low levels of (stimulus related and ongoing) activity, the systems response conforms to a quasi-linear superposition of separable responses to the fixed and stochastic inputs, which is consistent with the traditional assumptions that motivate trial averaging to suppress spontaneous activity and reveal the event-related response; (2) when activity is sufficiently high, there are nonlinear interactions between the fixed and stochastic inputs, which results in a phase resetting, which in turn leads to a different explanation for the appearance of an evoked response.
Was this article helpful?