Workshop 2 Summary by Barry Horwitz and John Rinzel:

Workshop 2: System Level Modeling

Because the levels of investigation varied so widely - from relatively small ensembles of neurons to essentially a large portion of the human brain - multiple modeling approaches were presented, and multiple kinds of neural data were considered. What became clear as the Workshop progressed was (1) that there were a variety of different neural modeling styles - bottom-up approaches, top-down approaches; approaches that stayed within one level of description; approaches that integrated data from several levels of investigation, and (2) that different approaches are required because there are a variety of questions that are being dealt with, and each approach is limited in what it can address.

The first session - levels of investigation - included presentations of the types of data available to computational modelers, and the kinds of questions that these data generate that modeling can help address. Included were experimental data obtained at the single unit level of investigation (Bair), at the mesoscopic (ensemble) level of investigation (Nicolelis), and at the whole-brain level (Horwitz and George). Electrical (spiking) activity is what is generally measured at the neuronal level of investigation, and Wyeth Bair presented an overview of this approach. His talk, which focused on data obtained from visual cortex, touched on several issues that were to be brought up several times by other speakers: (1) the role of context in altering the signals measured; (2) the fact that one type of signal (e.g., single unit activity) may appear to be inconsistent with data obtained by other types of signals (e.g., optical imaging data); and (3) the crucial role that feedback connections play. Miguel Nicolelis gave an overview of his work employing multiunit recordings to obtain population activity from multiple brain regions simultaneously. By combining such data with a computational model, one seeks to use the measured activity in one or more parts of the brain to predict activity in another part of the brain. A stunning example showed that one could predict and manipulate motor output by identifying preparatory neural activity in motor, premotor and parietal cortex. Both Barry Horwitz and John George discussed the role computational modeling can play in understanding data obtained primarily from human subjects by fMRI (Horwitz) and MEG (George). It was emphasized that fMRI and MEG data are particularly complex, and that computational modeling will be essential for understanding how such data are to be interpreted in terms of neural activity. This is particularly the case for fMRI data, which reflects changes in local hemodynamic activity and are only indirectly related to neural activity. It was emphasized that functional brain imaging data have not, until recently, been the focus of computational neuromodeling.

Several of the subsequent talks reinforced the points expressed in the first three presentations. Detlef Heck demonstrated (using intra- and extracellular data from the frontal cortex of the rat) how integration of synaptic inputs is affected by the background network activity. Gustavo Deco, Martin Stetter and Malle Tagamets presented fairly detailed network models (spatially distributed, cell and local-circuit based) that simultaneously generated both simulated single unit activity and simulated fMRI/PET data corresponding to specific cognitive tasks.

The second session of the workshop was devoted to sensorimotor processing. Steve Lisberger discussed the neural system mediating smooth pursuit eye movement. He introduced a type of control model for smooth pursuit (in which different modules perform such tasks as sensory-motor transformations and gain control), and then used single unit recordings from monkeys to identify neuroanatomical circuits that correspond to the modules in the model (e.g., frontal motor cortex associated with gain control - part of a parieto-frontal circuit). The remaining talks of this session focused primarily on the role of basal ganglia and cerebellar circuits in motor control. Jim Houk presented two models: one involved a cortical-basal ganglia circuit encoding the serial order of sensory events; the second concerned how a cortical-cerebellar circuit can be used to predictively regulated movement commands. Jose (Pepe) Contreras-Vidal, Dan Bullock and David Hansel all presented models that hypothesized that frontal-basal ganglia circuits function primarily to select internal models that had become established by prior learning. David Terman also focused on the basal ganglia, demonstrating with a computational model of spiking neurons the important role of the indirect pathway of the cortical-basal ganglia loops in Parkinson's disease and in how deep brain stimulation may act in helping alleviate some its symptoms by changing dynamic activity patterns.

There were two sessions on cognitive function. A group of presentations by Jonathan Cohen, Marius Usher, Phil Holmes, and Todd Braver were related to the issue of cognitive control. Like the Lisberger presentation mentioned above, Cohen, Usher and Braver started with a set of phenomena [e.g., tasks where the stronger (or more likely) of two responses must sometimes be inhibited], and a model (a connectionist model in this case) that is able to account for the behavioral performance of human subjects on such tasks. Lesion and fMRI data were then introduced that enabled different parts of the connectionist model to be associated with different parts of the brain (e.g., the anterior cingulate was proposed to play a critical role in conflict mediation). Holmes presented a model of the response of locus coeruleus neurons to target detection, which is a central component of the cognitive control hypothesis.

A variety of additional approaches to neural modeling of cognitive function were presented by other speakers. Steve Grossberg reviewed a large body of research centered on his Adaptive Resonance Theory of cognitive function that emphasizes the importance of matching bottom-up sensory data against top-down expectations. In a similar vein, Randy McIntosh discussed the way context affects how sensory stimuli are handled by the brain. In particular, he showed how this could be detected by applying techniques such as structural equation modeling to PET/fMRI data. Dana Ballard presented a model that enables synchronous spike codes on both feed forward and feedback connections between the Lateral Geniculate Nucleus (LGN) and visual cortex to form oriented receptive fields given natural images as input. This model incorporates both spiking synchrony and a rate code. In a effort to account for the fact that many human psychophysical results have been explained using Bayesian models, Rajesh Rao presented a neural model of cortex that can perform Bayesian inferences. The model was applied to a visual motion detection task. David Horn's presentation addressed the question of how is it possible to learn what are the perceptual features of importance. He used a PDP model to suggest how internal representations of new perceptual features are created through repeated exposure.

A number of presentations focused on working memory. Neural models that enable a short-term memory of a presented stimulus to be maintained were included in the presentations by Gustavo Deco and Malle Tagamets. Tagamets reviewed a large-scale neural model of the ventral visual processing pathway that uses, at each stage, an ensemble of Wilson-Cowan units, each representing a basic local computational element. The model mimics performance of a delayed match-to-sample task for simple shapes, and is able to reproduce the electrical activities of monkey neurons in multiple brain regions (including prefrontal cortex) and as well, the PET/fMRI activities observed in humans performing this type of task. The Deco model, which focuses primarily on prefrontal cortex, uses a network model of spiking neurons (developed by Nicholas Brunel and Xiao-Jing Wang). This model can account for both prefrontal electrical activity and prefrontal fMRI results in a delayed matching task incorporating both object and spatial components. XJ Wang showed how one could account for performance measures (e.g., reaction times) by incorporating a simple decision rule in such a modeling framework. Brunel elaborated on the mean field approach (embedded in the Wilson-Cowan formulation) by showing how it could be extended so that the synaptic input could be described by both a mean and a variance. This will permit one to have a network of irregularly firing neurons modeled using a mean field theory type of formalism.

Finally, there were a series of presentations based on the theme of Modeling Strategies for Multi-Scale Integration. Included here was a talk by Martin Stetter, who focused on the use of mean field theory at multiple scales in the mammalian primary visual cortex area V1. Among the topics he discussed was the importance of contextual information in modulating the response properties of visual orientation selective cells. Carson Chow presented models (both spiking and rate-based) to account for a number of experimental observations of visual binocular rivalry. John Rinzel also discussed both spiking and rate models but in a non-cognitive setting with models that attempt to account for the slow episodic population rhythms (with a time scale on the order of minutes) that are seen in chick embryonic spinal cord. This structure is of great interest, since all the synaptic currents, including those associated with GABA, are excitatory, yet oscillations in activity still occur. Synaptic depression plays a critical role in generating these oscillations.

Among the key conclusions from the multiple presentations at this workshop were the following: (1) much of importance in the way neurons affect behavior is based on their interaction with other neural populations, and to gain insights into these interactions require computational approaches; (2) increasingly, top-down effects (which come under a variety of names - context, recurrent inputs, attention, feedback) are being shown to play a central role in neural processing; (3) integration of multiple levels of data will become important for understanding neural systems, and such integration will be based on computational modeling; and (4) a variety of computational frameworks will be needed that range from, and bridge between, biophysically-based network models to high-level descriptions that employ cognitive/psychological-based state variables, and statistically-based formulations for discrimination and decision-making.