CNS Vision Lab

Spiking Bipole Cells

This model of cortical Layer 2/3 pyramidal cells shows how grouping can be done at the level of small cell populations. Model cells consist of two compartments governed by Hodgkin-Huxley equations and are connected via slow synapses. The network topology is based on the presence of recurrent horizontal connections with slow propagation speeds.

The bipole cell model was introduced as a way to account for our ability to extract illusory contours from the visual scene. Consider for example the Kanizsa square in Figure 1. Under appropriate viewing conditions, this stimulus yields the percept of a white square surface although with no well defined luminance edges. The percept seems to result from the grouping of signals coming from adjacent pacman shapes.

The Boundary Contour System and Feature Contour System model (BCS/FCS) postulates that this grouping is carried out by groups of oriented cells in visual cortex that respond to collinear inputs outside of their classical receptive field (Grossberg and Mingolla, 1985). These cells are termed bipole cells since they combine cooperating collinear inputs from two sides. This is further illustrated in Figure 1, where a bipole cell (blue ellipse) is seen to receive signals from neighboring complex cells (indicated in red). The activated bipole cell can then signal the presence of a contour at its location.

Figure 1 The percept of a square surface emerges from the grouping of the pacman figures. A bipole cell (blue ellipse) performs the grouping of signals coming from colinear complex cells placed outside of its classical receptive field.

Previous models incorporating bipole cells (e.g. Grossberg and Mingolla (1985), Grossberg and Raizada (1999)) did not explicitely attempt to map model cells to individual cortical cells, which would require the use of spiking dynamics. Rather, the grouping process was implemented with shunting equations, thereby loosing the potential to test the model with respect to realistic physiological constraints. The goal of this project is to show that the grouping mechanisms can be implemented with more plausible bipole cells.

In order to do this, a small network of cells is implemented using Hodgkin-Huxley spiking dynamics. Each cell is composed of two compartments, including one soma and one passive dendrite. Spiking bipole cells are interconnected via anisotropic excitatory long-range horizontal collaterals with slow conduction velocity. Such connections have been found in many species including macaque monkey (Angelucci et al., 2002) and cat (Gilbert et al., 1989). It is the presence of these collaterals that enable bipole cells to be sensitive to inputs separated far apart. Given the extra degree of complexity due to biophysical details, we follow the simplified methodology employed in Grossberg and Somers (1991) and Yazdanbakhsh and Grossberg (2004) and simulate the network with 1D patterns.

Figure 2 illustrates the kind of behavior expected from the network. On the left, the network simply reproduces a constant luminance input. On the right, the network produces a near identical response although a large gap is present in the input. Although not illustrated here, it is also necessary that this response be sensitive to the magnitude of contrast of the input stimuli and to size of the inducers with respect to the gap length. Physiological evidence exists that confirms each of these computational properties in a variety of species (Crook et al., 2002, Kapadia et al, 1995, 2000; Peterhans et al., 1989; Von der Heydt et al., 1989). Simulations aim at reproducing these data.

The bipole property Left) Network representation for a luminance defined boundary contour. Right) Network representation for an illusory contour. In this case, there is inward completion but no outward propagation. This is the bipole property.

A variety of physiological results (Peterhans et al., 1989; Von der Heydt et al, 1989) and theoretical considerations (Grossberg and Mingolla, 1985) suggest that bipole-like cells are present not only in primary visual cortex but also in prestriate cortex (area V2). The final phase of this project is devoted to combining V1 and V2 bipole cells in a single network that exhibits the same desired computational properties while remaining true to the physiological evidence.

This work is supported in part by the National Science Foundation (NSF SBE-0354378) and the Office of Naval Research (ONR N00014-01-1-0624).

This research is supported in part by

References