Computational models of recurrent cortical networks

Our proposal to the McDonnell Foundation is directed towards developing computational approaches to simulate neural networks. The models we will create will account for the highly dynamic nature of the functional properties of neurons in the cerebral cortex. An evolving idea of cortical function is that individual neurons and cortical areas are adaptive processors, changing their function according to context and top-down influences of attention, expectation and perceptual task. This is accomplished by altering the effective connectivity, or input selection, within cortical circuits. In this view information is represented by a complex set of interactions across cortical areas, rather than being localized to a specific area. The model will implement the dynamic switching of cortical networks by incorporating biologically realistic components of spiking neurons and including known components of cortical circuitry. Although the model is intended to address the nature of cortical information processing and brain function, the principles incorporated, including lateral interactions, recurrent processes and state switching, are likely to be common to other complex biological systems, ranging from signal transduction pathways to ecosystems.

An emerging theme of neural systems research is that, even at early stages of information processing, neurons and the cortical areas within which they reside are influenced by context, learning and higher order cognitive influences. Rather than performing stereotyped operations on a sensory stimulus, neurons are adaptive processors, executing different operations according to the immediate behavioral demands. The classical idea about sensory processing is that early stages perform a stereotyped operation, dealing with simple attributes at early stages and gradually increasing the complexity of the stimuli for which they are selective as one proceeds along the cortical sensory pathway. This is essentially a feedforward hierarchical model, even though it had long been known that feedback connections from higher to lower levels of the hierarchy existed. In addition, our discovery of long range horizontal connections in the visual cortex suggested that neurons integrated information over larger parts of the visual field than suggested by mapping their receptive fields with simple, short line segment, stimuli. In fact, we find that even in primary visual cortex neurons are capable of encoding information about complex visual stimuli extending over large parts of the visual field. Furthermore, the stimuli for which they are most responsive depend on the perceptual task that the animal is performing, a phenomenon that we refer to as a top-down influence. As a result neurons are not merely selective for the physical characteristics of the stimulus (its shape, for example) but by higher order cognitive influences. These influences depend on learned internal representations of stimulus configurations, so that the responses of cortical neurons reflect a comparison between an expected (or remembered) stimulus and information coming from the sensory periphery. One can think of this as a form of hypothesis testing, whereby object recognition requires countercurrent streams of processing, with bottom up information containing the representation of the physical stimulus and top-down processes conveying expectation and perceptual task.

Our modeling effort will attempt to represent the mechanisms, at the level of cortical circuitry, by which this process occurs. There are several components to cortical circuitry that play a critical role in the integrative processes we will model. The connections operate in a push-pull mode, with a balance between direct monosynaptic excitatory connections and disynaptic inhibitory connections, which together achieve a combination of an additive excitatory effect and a divisive inhibition. Within each cortical area there are long-range horizontal connections running between neurons that represent disparate locations in the cortical functional architecture. In the visual cortex, these connections mediate linkage between line segments that run along contours, and they therefore play a role in contour integration. Each neuron receives input from thousands of other widely distributed neurons in the cortical network. We propose that these connections are not constitutively active, but instead that subsets of connections change their effectiveness according to task demands, enabling each neuron to select subsets of connections to perform different operations. Finally, while a number of different areas of the cerebral cortex represent a hierarchy of information processing, the connections between layers of this hierarchy are not just feedforward, but each stage communicates with antecedent stages via feedback connections. Our model incorporates the idea that the role of feedback connections is to mediate the gating of intrinsic horizontal connections. Thus the function of any neuron is generated by a heterosynaptic interaction between feedback and intrinsic connections.