Laminart PDF

Vision Research 43 (2003) 801–829
www.elsevier.com/locate/visres
A laminar cortical model of stereopsis and three-dimensional

surface perception q
Stephen Grossberg *, Piers D.L. Howe
Department of Cognitive and Neural Systems and Center for Adaptive Systems, Boston University, 677 Beacon Street, Boston, MA 02215, USA
Received 6 March 2002; received in revised form 11 September 2002
Abstract
A laminar cortical model of stereopsis and later stages of 3D surface perception is developed and simulated. The model describes
how initial stages of monocular and binocular oriented filtering interact with later stages of 3D boundary formation and surface
filling-in in the lateral geniculate nucleus and cortical areas V1, V2, and V4. In particular, it details how interactions between layers
4, 3B, and 2/3A in V1 and V2 contribute to stereopsis, and clarifies how binocular and monocular information combine to form 3D
boundary and surface representations. Along the way, the model modifies and significantly extends the disparity energy model.
Neural explanations are given for psychophysical data concerning: contrast variations of dichoptic masking and the correspondence
problem, the effect of interocular contrast differences on stereoacuity, PanumÕs limiting case, the Venetian blind illusion, stereopsis
with polarity-reversed stereograms, da Vinci stereopsis, and various lightness illusions. By relating physiology to psychophysics, the
model provides new functional insights and predictions about laminar cortical architecture.
Ó 2003 Elsevier Science Ltd. All rights reserved.
Keywords: Cortical model; Depth perception; Stereopsis; Surface perception; Cortical layers; Lightness perception; Monocular–binocular
interactions
1. Introduction plain the 3D surface percepts that form an integral part

of our visual consciousness.
This article describes a model of how the lateral The present model, shown in Fig. 1, goes beyond
geniculate nucleus (LGN) and cortical areas V1, V2 and these previous analyses in several ways. First, it provides
V4 utilize both monocular and binocular visual in- a refined model of stereopsis in V1 which clarifies the
formation to produce three-dimensional (3D) surface role of cells in cortical layers 4, 3B, and 2/3A. In par-
percepts. Despite some explanatory successes, many ticular, the model revises how the disparity energy
previous cortical models, for example the disparity en- model achieves stereopsis, in a manner that is more
ergy model of Ohzawa, DeAngelis, and Freeman (1990), consistent with recent data. Second, the model shows
considered only stereopsis, which is an early stage of how monocular and binocular information are com-
depth perception that occurs in cortical area V1. Stere- bined and selected in V2 to form 3D boundary repre-
opsis is important, but on its own is insufficient to ex- sentations. Third, the model shows how these 3D
boundaries give rise to visible 3D surface percepts in V4.
Taken together, these model processes are used to ex-
q plain and simulate a much larger set of neurophysio-
Authorship in alphabetical order. SG was supported in part by
the Air Force Office of Scientific Research (AFOSR F49620-01-1-0397)
logical, anatomical, and psychophysical data about
and the Office of Naval Research (ONR N00014-01-1-0624). PDLH stereopsis and 3D surface perception than has previ-
was supported in part by the Air Force Office of Scientific Research ously been possible.
(AFOSR F49620-01-1-0397), the Defense Advanced Research Projects The modelÕs explanatory range is larger still since it is
Agency and the Office of Naval Research (ONR N00014-95-1-0409), consistent with, and generalizes, a recent laminar model
and the Office of Naval Research (ONR N00014-95-1-0657).
*
Corresponding author. Tel.: +1-617-353-7858; fax: +1-617-353-
of V1 and V2, called the LAMINART model, that
7755. does not incorporate binocular interactions (Gross-
E-mail address: [email protected] (S. Grossberg). berg, 1999a, 1999b; Grossberg, Mingolla, & Ross, 1997;
0042-6989/03/$ - see front matter Ó 2003 Elsevier Science Ltd. All rights reserved.
doi:10.1016/S0042-6989(03)00011-7
802 S. Grossberg, P.D.L. Howe / Vision Research 43 (2003) 801–829
Fig. 2. (a) The same-sign hypothesis: only edges that have the same
contrast polarity can be stereoscopically fused to produce a percept of
depth. (b) As it is traversed, the boundary of the ellipse changes its
contrast polarity relative to the background, thereby illustrating the
need for object boundaries to be represented in a contrast-invariant
manner. See text for details.
Fig. 1. Model circuit diagram.
depth perception will occur only if the two edges belong

Grossberg & Raizada, 2000; Grossberg & Williamson, to the same object. This is commonly referred to as
2001; Raizada & Grossberg, 2001). The LAMINART the correspondence problem (Howard & Rogers, 1995;
model explained data about perceptual development, Julesz, 1971). An early step in solving the correspon-
learning, grouping, and attention. Because the pre- dence problem is to binocularly fuse only edges with
sent model consistently generalizes the LAMINART approximately the same magnitude of contrast (McKee,
model to 3D vision, it is called the 3D LAMINART Bravo, Taylor, & Legge, 1994). This constraint naturally
model. Section 4 outlines how this synthesis leads to arises when the brain fuses edges that derive from the
a model that predicts how cellular and network mecha- same objects in the world. The model satisfies this
nisms of 3D vision are linked to mechanisms of devel- constraint through interactions between excitatory and
opment, learning, grouping, and attention. inhibitory cells in layer 3B of V1 that endow the bin-
The model achieves these goals by embodying five ocular cells there with an obligate property (Poggio,
basic psychophysical constraints in its neural circuitry: 1991), whereby they respond preferentially to left and
(1) Reconciles contrast-specific binocular fusion with right eye inputs of approximately equal size.
contrast-invariant boundary perception. It is well known (3) Solves the correspondence problem. Even if all
that only edges in the left and right retinal images that binocular matches are of the same-sign and similar
have the same contrast polarity (i.e., their luminance contrast magnitude, there can still exist many false
gradients have the same sign) can be binocularly fused binocular matches between edges that did not derive
to form a percept of depth (Howard & Rogers, 1995). from the same objects. This problem has often been
Otherwise expressed, binocular fusion obeys the same- approached by imposing a unique-matching rule, which
sign hypothesis (see Fig. 2). However, fused boundaries states that any given feature in one retinal image is
must also be able to form around objects whose contrast matched at most with one feature in the other retinal
polarity with respect to the background can reverse image (Grimson, 1981; Marr & Poggio, 1976; for a re-
along their perimeters (Grossberg, 1994). In other view see Howard & Rogers, 1995, pp. 42–43). However,
words, binocular boundaries need to be represented in a this rule fails in situations like PanumÕs limiting case
contrast-invariant way. How can the brain reconcile (Gillam, Blackburn, & Cook, 1995; McKee, Bravo,
contrast-specific fusion with the need to form contrast- Smallman, & Legge, 1995; Panum, 1858) where a bar
invariant object boundaries? The model proposes that presented to one eye is simultaneously matched to two
both constraints are realized by interactions between separate bars presented to the other eye. The present
cells in layers 4, 3B, and 2/3A of cortical area V1 in- model does not enforce unique matches. Rather, the
terblobs (see Fig. 1). model encourages them by using a disparity filter that is
(2) Implements the contrast constraint on binocular proposed to occur in the pale stripes of cortical area V2,
fusion. The brain needs to determine which of the many possibly in layer 3B (see Fig. 3). This disparity filter uses
potential same-sign edges in the two retinal images two types of inhibitory interactions: line-of-sight inhi-
should be binocularly fused, since veridical stereoscopic bition and inhibition across depth but within cyclopean
S. Grossberg, P.D.L. Howe / Vision Research 43 (2003) 801–829 803
Once monocular information is included, the prob-

lem immediately arises about how to combine mon-
ocular and binocular boundaries. This is a problem
because monocular boundaries do not have a definite
depth associated with them. How, then, can we decide to
which depth they should be assigned? A proposed ap-
proach to this monocular–binocular interface problem
was suggested in Grossberg (1994, 1997) in order to
explain data about 3D figure-ground perception. Here
the same hypothesis is shown to play a crucial role in
explaining many other data about 3D surface percep-
tion. Namely, the model assumes that the outputs of the
monocular boundary cells are added to all depth planes
in the pale stripes of cortical area V2 along their re-
spective lines-of-sight, possibly in layer 4 (see Figs. 1 and
4a). The disparity filter, which helps to solve the corre-
Fig. 3. The V2 disparity filter. The V1 binocular boundaries network spondence problem, also solves the monocular–binocu-
matches an edge in one retinal image with every other edge in the other lar interface problem by automatically eliminating most
retinal image whose relative disparity is not too great, that has the
same contrast polarity and whose magnitude of contrast is not too
of the monocular boundaries that are not at the correct
different. In response to this image, the V1 boundary network creates depths.
four matches, with the two not in the fixation plane being false matches (5) Forms 3D surface percepts. So far we have con-
between edges that do not correspond to the same object. As described sidered only how the brain constructs a 3D boundary
in the text, these false matches are suppressed by the disparity filter in representation of an object. There is considerable evi-
V2, wherein each neuron is inhibited by every other neuron that shares
either of its monocular inputs (i.e., shares a monocular line-of-sight
dence that boundary representations on their own do
represented by the solid lines) or is directly in front of or behind it (i.e., not give rise to visible percepts, which rather are a
is connected to it by a dashed line). Note in particular that the solid property of surface representations (Grossberg, 1994).
lines that represent the monocular lines-of-sight also represent the al- In the present model, surface representations derive
lelotropic shifts: an edge in the left retinal image is shifted to the right from a filling-in process whereby lightness and color
for matches increasingly further away whereas an edge in the right
retinal image is shifted in the opposite direction.
mark the depths at which the surfaces occur. Filling-in is
needed to recover lightness and color estimates in re-
gions where they have been suppressed by the process of
position, to encourage unique matches, as described in discounting the illuminant (Grossberg & Todorovic,
more detail below. 1988). Boundaries control the depths at which particular
(4) Combines monocular and binocular information to lightnesses and colors can fill-in, a process that we call
form depth percepts. Although PanumÕs limiting case
seems to be a laboratory curiosity at first, there are
many naturally occurring situations where there is only
one edge in one eye and two possible edges with which
to match it in the other eye. For example, due to the
lateral displacement of the eyes, an objectÕs edge that is
seen by one eye may be occluded in the other eye, as
occurs during da Vinci stereopsis (Nakayama & Shim-
ojo, 1990). Despite this lack of binocular information,
the monocularly viewed region has a definite depth
conferred to it by the binocularly viewed parts of the
scene. The brain can thus utilize monocular informa-
tion to build up seamless 3D percepts of the world. In
fact, in experiments involving PanumÕs limiting case,
varying the relative contrast of the bars alters the per-
ception of depth in a manner that reveals clear monoc- Fig. 4. (a) Open and connected boundaries. (b) Filling-in of surface
ular–binocular interactions (Smallman & McKee, 1995). lightness is contained or not depending on the connectedness of the
Dichoptic masking, where an object presented to one boundary. Note that the monocular boundaries (i.e., two horizontal
boundaries and the right vertical boundary) have been added to all
eye is obscured (i.e., masked) by one presented to the
depth planes whereas the binocular boundary (i.e., the left vertical
other eye, illustrates a third way in which monocular boundary) is present only in the near depth plane, thereby creating a
and binocular information may interact (McKee et al., connected boundary, and thus containment of filling-in, only in the
1994). near depth plane.
3D surface capture. The present article considers only The 3D LAMINART model overcomes this limi-
the filling-in of achromatic lightness. tation using identified cells in laminar circuits, and
How does the brain ensure that lightness fills-in at resimulates all the data previously simulated by
only the correct depths? Grossberg (1994) proposed McLoughlin and Grossberg (1998), in particular the
properties of this boundary–surface interaction that data on contrast variations of the correspondence
helped to explain many data about 3D figure-ground problem and dichoptic masking. In addition, the new
perception. Here, one of these properties proved essen- model can simulate still more psychophysical data than
tial to explain 3D surface percepts that arise in stere- its non-laminar predecessors, including: the Venetian
opsis research. Namely, visible surfaces arise in cortical blind illusion, four different examples of da Vinci ste-
area V4 only if they are enclosed by connected bound- reopsis (Gillam, Blackburn, & Nakayama, 1999;
aries (see Fig. 4). In particular, a rectangular connected Nakayama & Shimojo, 1990), stereopsis with opposite-
boundary may be composed of one vertical binocular contrast stimuli, the effect of interocular contrast dif-
boundary, one vertical monocularly viewed boundary, ferences on stereoacuities and various lightness illusions.
and two horizontal boundaries that code no disparity In so doing, it demonstrates more of the roles that
information. This connected boundary can support a boundary and surface representations play in depth
visible surface percept at the depth corresponding to the perception. The 3D LAMINART model also makes
binocular boundary if all other constraints are satisfied. neurophysiological predictions, including that there
Such a boundary can contain the filling-in process. exist: (1) In V1 cells that obey the ratio constraint on
However, if the vertical binocular boundary is missing, binocular fusion. The model proposes that some bin-
as it would be at a different depth plane, then the total ocular simple cells in layer 3B obey an obligate property
boundary is not connected, and a visible percept will whereby they can be activated only if they receive ap-
not be evident at that depth because filling-in can dis- proximately equal inputs from both left and right eye
sipate out of the boundary gap. This example illus- monocular simple cells in layer 4. The constraints that
trates how the monocular–binocular interface problem determine cell firing depend upon the ratios of left and
(item (4) above), and thus the correspondence problem right monocular cell activity. This property explains the
(item (3) above), influence visible percepts of 3D sur- ratio constraint on stereoscopic fusion that is illustrated
faces. in Fig. 10 below. The obligate property is predicted to be
The present model refines aspects of the FACADE caused by a balance between excitatory inputs from
model of 3D vision and figure-ground perception layer 4 monocular simple cells and inhibitory inputs
(Grossberg, 1994, 1997). The FACADE model included from layer 3B inhibitory interneurons. The interneurons
a (non-laminar) model of stereopsis and 3D planar are themselves activated by layer 4 monocular simple
surface perception (Grossberg & McLoughlin, 1997; cells and mutually inhibit each other, in addition to in-
McLoughlin & Grossberg, 1998) that modified and hibiting the binocular simple cells. (2) In V2 cells that
generalized the disparity energy model of stereopsis solve the correspondence problem using a disparity fil-
(Ohzawa et al., 1990). This generalization incorporated ter. (3) In V4 a filling-in mechanism that completes
rectification prior to binocular combination, absent visible 3D surface representations within connected
from the original disparity energy model, which has boundaries. These results were briefly reported in Howe
recently received independent experimental support and Grossberg (2001).
(Cumming, 2002; Read, Cumming, & Parker, 2002). It
also proposed that positional shifts between left and
right eye cortical inputs code disparities, rather than
phase shifts, which has also received experimental sup- 2. Model description
port (Tsao & Livingstone, in press). The FACADE
model also incorporated a disparity filter to help solve The model consists of four component networks
the correspondence problem (Howard & Rogers, 1995) which process: V1 binocular boundaries, V1 monocular
as well as mechanisms for filling-in 3D surface percepts boundaries, V2 boundaries, and V4 surfaces. For a
from 3D boundary representations. In particular, the mathematical description, the reader is referred to Ap-
FACADE model explained the fact that stereoscopic pendix A. A description of the neurophysiological and
fusion is generally impossible when the left and right eye anatomical evidence that supports all the model pro-
stimuli differ too much in contrast (Smallman & McKee, cessing stages is found in Section 4.1. In order to reduce
1995). However, in the form developed by Grossberg the computational load, the model currently considers
and McLoughlin, the FACADE model could not ex- only horizontal and vertical contours and five depth
plain why stereoscopic fusion is always possible in the planes. Even so, the model includes approximately
special case where each eye sees only a single bar, re- 185,000–333,000 cells depending on the simulation. Al-
gardless of the contrast difference of the two bars though model cells and cells in vivo will be clearly dis-
(McKee et al., 1994; Smallman & McKee, 1995). tinguished in the text, model cells will be referred to by
physiological labels because their properties so closely These proposed interactions between layers 4, 3B and
match those found in vivo. 2/3A are consistent with neurophysiological data, as
detailed in Section 4.1, and instantiate key operations of
the disparity energy model (Ohzawa et al., 1990), which
2.1. V1 binocular boundaries itself is strongly supported by physiological evidence; for
a review (see Ohzawa (1998)). As discussed in Section 1,
The network that processes the V1 binocular the need for preprocessing before the site of binocular
boundaries is located in the V1 interblob region and combination, such as that carried out by layer 4 of our
includes the binocular cells in layers 3B and 2/3A. It model, has recently been demonstrated by Cumming
carries out stereoscopic fusion of vertical contours, but (2002) and Read et al. (2002), who showed that pre-
not of horizontal contours, which it assumes cannot be processing was required to explain subtleties in physio-
stereoscopically fused. This network implements the logical data not captured by the original disparity
same-sign hypothesis (see Section 1, item (1)). As shown energy model.
in Fig. 1, inputs to the left and right eyes activate
monocular simple cells in layer 4 of the V1 interblob 2.2. V1 monocular boundaries
regions. Left and right eye monocular simple cells con-
jointly activate binocular simple cells in layer 3B whose The network that processes the V1 monocular
depth sensitivity is determined by the relative retinal boundaries comprises the monocular cells in layers 4, 3B
disparity of the layer 4 monocular cells that project to and 2/3A of the V1 interblob region. It is similar to the
them. The model implements the same-sign hypothesis binocular boundaries network, but represents both
by assuming that only layer 4 simple cells with the same horizontal and vertical boundaries whereas the binocu-
contrast polarity project to a single layer 3B simple cell. lar boundaries network represents only vertical bound-
These layer 3B simple cells are therefore selective for aries. Binocular boundary cells preferentially represent a
binocular disparity and a prescribed contrast polarity. particular depth plane, but this is not true of monocular
Binocularly fused vertical contours that occupy corre- boundary cells. How, then, do monocular and binocular
sponding points on the two retinas are seen as a single boundaries interact? A proposed solution of this mon-
boundary in the fixation plane, whereas vertical con- ocular–binocular boundary interface problem assumes
tours that are displaced relative to each other are seen as that the outputs of the monocular boundary cells are
a single boundary either in front of or behind the fixa- added to all depth planes in cortical area V2 along their
tion plane, depending on their displacement, as detailed respective lines-of-sight (see Section 1, item (4) and Fig.
in Appendix A (Eq. (A.10)). 3). Appendix A, Eqs. (A.12) and (A.13), describe this
There are also inhibitory cells in layer 3B. As is de- process quantitatively.
scribed in Appendix B, these cells ensure that the bin- As noted in Section 1 (item (4)), the V2 disparity filter
ocular simple cells act like the ‘‘obligate cells’’ of Poggio helps to solve the monocular–binocular interface prob-
(1991): The activity of such a binocular simple cell is lem, as well as the correspondence problem, by elimi-
suppressed by these inhibitory interneurons if the mag- nating most of the monocular representations that are
nitudes of the left and right eye inputs differ too much not at the correct depth. This previously unexpected
(see Section 1, item (2)). In particular, these obligate property of the disparity filter is crucial to understand-
cells respond to binocular, but not to monocular, stim- ing the monocular–binocular interactions described in
ulation. These obligate cells help to solve the corre- this paper. It can best be understood by studying the
spondence problem by ensuring that only similar stimuli model simulations in Section 3.
in the left and right eye retinal images are stereoscopi-
cally fused. 2.3. V2 boundaries
The next processing stage implements contrast-
invariant boundary detection (see Section 1, item (1)). The disparity filter network that processes V2
Layer 3B simple cells that are sensitive to the same po- boundaries is located in the V2 pale stripes (see Section
sition and disparity, but opposite contrast polarities, 1, item (3)). The V1 binocular boundaries network at-
pool their signals at layer 2/3A complex cells. These tempts to match every vertical edge in one retinal image
complex cells therefore respond to both contrast polar- with every other nearby vertical edge in the other retinal
ities and so can generate three-dimensional object image that has the same contrast polarity and approxi-
boundaries even if the objectÕs contrast polarity, with mately the same magnitude of contrast. Fig. 3 shows the
respect to the background, reverses as the boundary is resultant matches if each eye sees two bars. V1 makes
transversed. In summary, the two layers 3B and 2/3A, four matches. Only the two in the fixation plane are
acting together, can realize the same-sign hypothesis and correct matches. The other two are false matches be-
also begin to compute object boundaries in front of tween retinal images that do not correspond to the same
textured backgrounds. object. Such false matches are known to occur in V1 but
less readily in V2 (Bakin, Nakayama, & Gilbert, 2000; and that interact through on-center, off-surround cir-
Cumming & Parker, 2000). As they typically do not give cularly symmetric receptive fields. The present model
a veridical depth perception, these false matches must be utilizes such model neurons, which are analogous to
suppressed. those found in the LGN, as summarized in Section 4.1
Fig. 3 illustrates how the disparity filter works. To and defined in Appendix A (Eqs. (A.1)–(A.3)). These
encourage unique matching, the model assumes that model neurons are excited by spots of light applied to
each neuron inhibits all other neurons that share either the center of their receptive fields but are inhibited by
of its monocular inputs; that is, shares one of its mon- those applied outside this central region. The excitatory
ocular lines-of-sight. This is represented by the solid and inhibitory components of the receptive fields are
lines between neurons in Fig. 3. This rule on its own balanced so that cell responses are attenuated to spa-
could ensure that only two of the four initial matches in tially uniform or slowly varying stimulation. The cells
Fig. 3 survive, but it could not guarantee that it is the therefore respond preferentially to luminance borders.
false matches that are suppressed. A second form of At a later processing stage, these border signals propa-
inhibition ensures this. This inhibition acts across depth gate throughout those surface regions that are com-
and within cyclopean position. It is represented by the pletely enclosed by boundaries to complete the lightness
dashed line between each neuron with every other neu- representation. Propagation occurs via a filling-in pro-
ron that is directly in front of or behind it. These two cess that is akin to a diffusion process, as defined in
types of inhibition work together to ensure that the two Appendix A (Eqs. (A.17)–(A.23)). Propagating signals
matches in the fixation plane typically win, thereby can dissipate across space unless the region is sur-
solving the correspondence problem. It should be rounded by a connected boundary (see Fig. 4). As in
stressed that the disparity filter operates only on verti- Grossberg (1994), the present model proposes that the
cally oriented cells, as the model assumes that horizontal final stage of filling-in occurs in V4, where visible surface
boundaries cannot be fused and therefore cannot give percepts are predicted to occur. Section 3 summarizes
rise to false matches. It will be shown in Section 3.1.4 how such a filling-in process, when confined by the 3D
how this filter is also able to explain how, in some sit- boundaries of the present model, can explain da Vinci
uations, double matching can occur, as in PanumÕs stereopsis, as well as many aspects of lightness percep-
limiting case, an example of stereopsis that many pre- tion, thereby linking the modelÕs explanations of surface
vious models (e.g., Grimson, 1981; Marr & Poggio, depth and lightness.
1976) could not explain.
2.4. Surfaces 3. Model simulations
Boundaries help give rise to 3D surface percepts in This section summarizes simulations that predict how
the manner summarized in Section 1 (item (5)). Al- monocular and binocular information interact in the
though our main goal is to explain percepts of surface visual cortex. We will consider, in turn, contrast varia-
depth, percepts of surface lightness are also simulated to tions of dichoptic masking, stereoacuity, PanumÕs lim-
show that our development of cortical depth perception iting case, contrast variations of the correspondence
mechanisms are consistent with simulations in related problem, the Venetian blind illusion, stereopsis with
modeling studies of surface brightness and lightness opposite-contrast stimuli, da Vinci stereopsis, and the
(e.g., Grossberg & Kelly, 1999; Kelly & Grossberg, Craik–OÕBrian–Cornsweet lightness illusion. The main
2000). Such a unified set of simulations supports the key aim of these simulations is to illustrate how the modelÕs
FACADE prediction that the same process fills-in four component networks interact with each other to
surface lightness, color, and depth (Grossberg, 1994). explain the percepts reported by human subjects. These
Previous simulations of lightness often focused on explanations constitute testable predictions for linking
computing the relative lightnesses of surface regions (but psychophysical percepts to their cortical mechanisms.
see Grossberg, Mingolla, & Williamson, 1995). Once Like the model diagram shown in Fig. 1, the simulation
relative lightness is estimated, then absolute lightness figures should be read from the bottom up, with the
can be computed in many cases by assuming that the bottom two rows representing the input and the V1
lightest surface of the group is white and calculating the boundary representations, the next two rows represent-
absolute lightnesses of all other surfaces relative to that ing the V2 boundary representations and the top row
one (Wallach, 1976). representing the V4 surface representations. Further-
Grossberg and Todorovic (1988) computed the rela- more for each of the top four rows, depth increases from
tive lightness of two surfaces by first discounting the left to right, with the middle plot representing the fixa-
effects of a spatially non-uniform illumination (see Sec- tion plane, the two leftmost plots representing the two
tion 1, item (5)). Discounting the illuminant can be near depth planes and the two right plots representing
achieved by neurons that obey cell membrane equations the two far depth planes.
3.1. Dichoptic masking ond plot of this row. The vertical boundaries in this
disparity plane are consequently stronger than those in
3.1.1. The basic paradigm the other four depth planes, which they then suppress
In the basic paradigm considered by McKee et al. via the line-of-sight inhibition of the V2 disparity filter
(1994), the contrast threshold for the detection of a low (cf., Fig. 3) to give the final V2 boundary representations
contrast bar presented to one eye was found to increase shown in the fourth row. Notice, in particular, that all
radically when a high contrast bar was presented to the horizontal boundaries have survived since the disparity
other eye. Furthermore it was not necessary for the two filter only inhibits vertical boundaries. In contrast to the
bars to be at retinal correspondence. The model expla- horizontal boundaries, only the vertical boundaries in
nation of this percept is as follows. The high contrast the near disparity plane, represented by the second plot
bar is presented to the left eye and the low contrast bar of this row, have survived. As explained in Section 2.4,
to the right, as shown by the middle two plots in the lightness signals, originating at the location of the
bottom row of Fig. 5. The outer two plots of the bottom boundaries, propagate throughout this disparity plane.
row show the simulated monocular boundary repre- Because the near disparity plane contains a connected
sentations. Since their contrasts differ greatly, these two boundary that completely encloses a bar-shaped region,
bars cannot be stereoscopically fused in V1 due to the these boundaries can contain the filling-in of the light-
inhibitory circuit in layer 3B, as explained in Section 2.1. ness signals to cause the bar-shaped surface percept
This accounts for the absence of V1 binocular bound- shown in one plot of the top row. The other filling-in
aries representations in the second row. As the monoc- signals dissipate and do not give rise to a conscious
ular boundaries do not yet have a depth associated with surface percept (see Fig. 4). Because the bars in the left
them, they are added to all depth planes in V2 along and right eye inputs are perceived to occupy the same
their respective monocular lines-of-sight, as shown in position in 3D space, the high contrast bar masks the
the third row of this figure. In this row, each of the five low contrast bar. In summary, this simulation shows
plots represent a different depth, with those on the left how the left and right inputs can be fused to form a
representing depth planes nearer than the fixation plane single percept in V4 even though their contrasts are so
and those on the right the converse. As we move across different that they cannot be fused by the binocular cells
this row the allelotropic shifts (cf., Fig. 3) cause the left in V1.
monocular boundaries to be added to locations further The fact that the V2 disparity filter can fuse bars
to the right in successive depth planes, while the right whose contrasts are too different to be fused in V1 has
monocular boundaries are added to locations further to ramifications for stereoacuity. In particular Schor and
the left. The left and right monocular boundaries coin- Heckmann (1989) noted that increasing the contrast of
cide in the near disparity plane represented by the sec- the image equally in both eyes increases stereoacuity,
but increasing the contrast of the image in just one eye
decreases stereoacuity. The model explanation is simply
that in the first case fusion could occur in V1 but in the
second only in V2. Since V1 cells in general have smaller
receptive fields than V2 cells that correspond to the same
region of visual space, the model is therefore able to
explain why stereoacuity is greater in the first case than
in the second.
3.1.2. Release from dichoptic masking

McKee et al. (1994) continued their study of dich-
optic masking by demonstrating that, for the particular
case where the two bars of Fig. 5 were in retinal corre-
spondence, the addition of a second high contrast bar to
the right input releases the low contrast bar from
masking. The model explanation is summarized in Fig.
6a. As before, the stimuli are shown in the middle two
Fig. 5. Model simulation showing that retinal correspondence is not plots of the first row and the monocular boundary
needed for dichoptic masking (McKee et al., 1994). The first row representations in the outer two plots. Now the high
represents the inputs and the V1 monocular boundaries, the second contrast bar of the left input is able to binocularly fuse
row the V1 binocular boundaries, the third and fourth rows the V2
with the high contrast bar of the right input to form the
boundaries and the fifth row the V4 surface percepts. In the top four
rows depth increases from left to right for successive plots, with the vertical V1 binocular ‘‘far’’ boundary representations
middle plot representing the fixation plane. All other simulation plots shown in the fourth plot of the second row. As before,
use the same format. See text for details. the monocular boundaries are added to all disparity
position as the low contrast bar of the right input. The

low contrast bar is therefore no longer masked. This
simulation explains the observation of McKee et al.
(1994) that the addition of a high contrast bar to the
right input causes the low contrast bar of the right input
to be released from dichoptic masking.
In Fig. 6b, instead of placing a high contrast bar next
to the low contrast of the right input bar, as in Fig. 6a, a
low contrast bar is placed next to the high contrast bar
of the left input. The resultant simulation is very similar
to that shown in Fig. 6a. In particular, the two low
contrast bars binocularly fuse and the resultant allelo-
tropic shifts mean that once again the high contrast bar
of the left input is not perceived to cover the low con-
trast bar of the right input resulting in a release from
dichoptic masking.
Fig. 6a and 6b together show that the release from
dichoptic masking can be achieved by adding either a
high or low contrast bar to the original stimulus of Fig.
5. Since McKee et al. (1994) considered only the stim-
ulus configuration of Fig. 6a, Fig. 6b represents a novel
prediction.
3.1.3. Return to dichoptic masking

McKee et al. (1994) also observed that the release
from dichoptic masking observed with the stimulus
configuration of Fig. 6a did not occur when the addi-
tional bar in the right input had a low contrast. Fig. 7a
Fig. 6. (a) Simulation of the release from dichoptic masking reported describes the model simulation of this property. The key
by McKee et al. (1995). (b) Simulation of another way a release from difference between Figs. 7a and 6a is that the right bar of
dichoptic masking may be achieved. See text for details. the right input no longer fuses with the bar of the left
input because, as explained in Section 2.1, their con-
trasts differ too greatly. This accounts for the lack of V1
planes in V2 along their respective monocular line-of- boundary representations in the fourth plot of the sec-
sights (cf., Fig. 3) as shown in the third row. In addition, ond row. The situation is now very similar to the basic
the V1 binocular boundaries are also added to V2, co- dichoptic masking paradigm, depicted in Fig. 5, thereby
inciding with the right bar representation of the fourth explaining the return to dichoptic masking.
plot of this row. The vertical boundaries of this bar The monocular boundaries are added to all disparity
representation are therefore stronger then the vertical planes in V2, shown in the third row of this figure, in
boundaries in the other depth planes, in particular those exactly the same manner as in the previous section.
of the fixation plane, which they consequently suppress Because the left bar of the right input is in retinal cor-
via the line-of-sight inhibition of the V2 disparity filter. respondence with the single bar of the left input, their
This then is the reason why the left bar of the right input boundary representations overlap in the zero disparity
is not perceived to lie in the fixation plane even though it plane, thereby forming the leftmost bar representation
is seen only monocularly. The vertical boundaries of the in the middle plot of this row. The vertical boundaries of
left bar representation of the fourth plot are not sup- this bar representation are consequently stronger and so
pressed because they do not share any lines-of-sight with suppress all the other vertical boundaries that share
the vertical boundaries of the right bar of this plot. The either of their lines-of-sight. They do not, however, sup-
final V2 boundary representations are shown in the press the vertical boundaries corresponding to the right
fourth row. bar representation of this plot because these boundaries
Only those vertical boundaries in the fourth plot have do not share any of their lines-of-sight. The final V2
survived. Consequently, only the fourth plot contains boundary representations are shown in the fourth row.
regions that are completely enclosed by boundaries, and Only the vertical boundaries in the fixation plane,
so give rise to surface percepts in V4, as is shown in the represented by the middle plot, have survived. These two
fourth plot of the top row. The high contrast bar of the sets of boundaries completely enclose two bar-shaped
left input is no longer perceived to occupy the same 3D regions. As before, these boundaries confine the light-
inhibition, but does not enforce it. One advantage of this

is that the model can simulate PanumÕs limiting case,
where a bar in one eye is simultaneously fused with two
bars in the other eye (Gillam et al., 1995; McKee et al.,
1995; Panum, 1858; but see Frisby (2001) and Wang,
Wu, Ni, & Wang (2001) for variations where double
matching does not seem to occur). Fig. 7b shows the
model simulation where a bar in one eye masks equally
two bars presented to the other eye as reported by
McKee et al. (1995).
The left eye sees a single bar and the right eye sees
two bars as shown by the middle two plots of the first
row. The resultant monocular boundaries are shown by
the outer two plots of this row. Area V1 fuses the bar of
the left input with both bars of the right input, to form
binocular boundaries in both a near and a far disparity
plane, represented by the second and fourth plots of the
second row. The monocular boundaries are added to all
disparity planes in V2 along their respective lines-of-
sight, as shown by the third row. The left monocular
boundaries form the left bar representation in the first
two plots, the middle bar representation in the third
plot, and the right bar representation in the fourth and
fifth plots of the third row. Similarly, the right monoc-
ular boundaries form the two right bar representations
in the first two plot, the outer two bar representations in
the third plot, and the leftmost two bar representations
Fig. 7. (a) Simulation of the return to dichoptic masking reported by in the fourth and fifth plots of this row. The V1 binoc-
McKee et al. (1995). (b) Simulation of dichoptic masking in PanumÕs ular boundaries are also added to V2, coinciding with
limiting case reported by McKee et al. (1995). See text for details.
the vertical boundaries of the left bar representation in
the second plot and the right bar representation in the
ness signals, which originate at the locations of fourth plot. Those boundaries in V2 that receive bin-
the boundaries, to generate the two bar-shaped surfaces ocular input are stronger than and consequently sup-
shown in the middle plot of the top row. The high press, via the recurrent inhibition of the V2 disparity
contrast bar of the left input coincides with, and there- filter, those V2 vertical boundaries that receive only
fore again masks, the left bar of the right input. In monocular input and that share one of their lines-of-
summary, this simulation explains the observation sight. The surviving V2 boundary representations are
of McKee et al. (1994) that dichoptic masking returns shown in the fourth row. Those regions in V2 that are
when the contrast of the right bar of the right input is enclosed by a connected boundary give rise to surface
reduced. percepts in V4. The model correctly predicts that the bar
A key difference between this simulation and that of the left input is matched with both bars of the right
shown in Fig. 6a is that in this simulation the monoc- input, and so masks them both equally (McKee et al.,
ularly viewed bars are perceived to lie in the fixation 1995).
plane whereas in the previous simulation the line-of-
sight inhibition of the V2 disparity filter prevented the 3.2. Contrast variations of the correspondence problem
monocularly viewed bar of that simulation from being
perceived to lie in the fixation plane. Taken together The previous simulations have demonstrated the
these two simulations show how the line-of-sight inhi- crucial role the disparity filter plays in explaining prop-
bition of the V2 disparity filter can interact with the erties of dichoptic masking. However, as was discussed
spatial layout of the stimuli to determine the perceived in Section 1, the disparity filter also helps to solve the
depth arrangement. correspondence problem by eliminating matches be-
tween edges that belong to different objects. The unique-
3.1.4. Dichoptic masking in Panum’s limiting case matching rule that various other models have imposed
As described in the next section, the present model cannot hold in general, since PanumÕs limiting case
solves the correspondence problem by using a disparity shows that, in certain circumstances, a feature in one
filter that encourages unique matching, via line-of-sight eye can be matched to two features in the other eye. This
is why the disparity filter of the present model encour- the second plot of this row, which is caused by the in-
ages unique matching but does not enforce it. The model appropriate fusion of the right bar of the left input with
will now be shown to simulate all the data from the the left bar of the right input. As usual, the monocular
Smallman and McKee (1995) extensive study of the boundaries are added to all depth planes in the V2 dis-
correspondence problem even though it does not enforce parity filter along their respective monocular lines-
the uniqueness constraint. In so doing, it clarifies the of-sight, as shown in the third row of this figure. In
crucial role that monocular–binocular interactions play addition, the binocular bar representations are also
in these percepts. added to V2, coinciding with the middle bar representa-
tion in the second plot and both bar representations of
the fourth plot. Those vertical boundaries that receive
3.2.1. Control experiment
binocular input, being stronger, quickly inhibit via the
Smallman and McKee (1995) initiated their study by
V2 disparity filter all other vertical boundaries that
performing a control experiment in which each eye was
share their lines-of-sight and only receive monocular
presented with two bars, all four bars having the same
input. The two sets of vertical boundaries in the fourth
high contrast. Subjects reported seeing two identical
plot, both of which receive binocular input, cooperate
bars, both in the far disparity plane. Fig. 8a shows the
via the disparity filter to inhibit the vertical boundaries
corresponding model simulation.
of the middle bar representation of the second plot,
Since the left input is displaced leftwards relative to
which also receive binocular input. This happens be-
the right input, the vertical edges of the two bars fuse in
cause the middle bar boundaries receive binocular in-
the far disparity plane in V1, as is shown by the fourth
puts that share monocular inputs with their inhibitors.
plot of the second row. In addition to this, there is a
The final V2 boundary representations are shown in the
false match in the near disparity plane of V1, shown in
fourth row. The model correctly predicts that subjects
see both bars in the far disparity plane. In summary, this
simulation shows how the line-of-sight inhibition of the
V2 disparity filter ensures that the false match that is
present in V1 (second plot of the second row) is elimi-
nated. The V2 disparity filter is therefore the reason why
the model can solve the correspondence problem.
Fig. 8b shows a more complicated version of the
correspondence problem. Once again the false matches
are shown in the second plot of the second row and the
correct matches in the fourth plot. Since there are more
correct matches than false matches, the latter are again
suppressed by the former via the line-of-sight inhibition
of the V2 disparity filter. This simulation shows that the
model can be applied to more general versions of the
correspondence problem than that shown in Fig. 8a. In
Section 3.3 the model is applied to a particularly com-
plex version of the correspondence problem known as
the Venetian blind illusion. These simulations of the
correspondence problem, the Venetian blind illusion
(Figs. 11 and 12) and da Vinci stereopsis (Figs. 14 and
15), among others, clarify how the model will generalize
to natural images by showing how it deals with a variety
of potentially confusing matches within the fusion
range.
3.2.2. Contrast variations

After performing their control experiment, Smallman
and McKee (1995) then proceeded to study contrast
variations. They first considered the case where the left
bar of the left input had a much lower contrast than
Fig. 8. (a) Simulation of the control experiment Smallman and McKee
(1995) used for subsequent studies of the correspondence problem. (b)
the other three bars. They found that observers per-
Simulation of a more complicated version of the correspondence ceived this bar to lie in the zero disparity plane while
problem. See text for details. also perceiving two high contrast bars, the left lying in a
vertical boundaries of the leftmost bar representation of

the middle plot because these boundaries do not share
any of their lines-of-sight. The final V2 boundary rep-
resentations are shown in the plots of the fourth row.
After the surviving connected boundaries fill-in, three
surface representations form where humans see them.
Smallman and McKee (1995) then studied the inverse
situation where the left bar of the left input had a much
higher contrast than the other three bars, which all had
the same contrast, as depicted by the middle two plots of
the first row of Fig. 9b. They found that this situation
produced very similar results to the last situation with
the left bar of the left input being perceived to lie in the
zero disparity plane and the two bars of the right input
being perceived to lie in the near and far disparity planes
as before.
According to the model, the left bar of the left input
once again cannot fuse with either of the bars of the
right input, this time because its contrast is too high.
The situation is therefore virtually identical to that de-
picted by Fig. 9a, thereby explaining the similar percept
reported by the subjects.
3.2.3. The ratio rule

The only difference between the control experiment
and these last two experiments was that, in the latter, the
left bar of the left input had a significantly different
Fig. 9. (a) Simulation of a contrast variant of the correspondence contrast from the other three bars. A key question is
problem studied by Smallman and McKee (1995). (b) Simulation of how great this contrast difference must be to cause the
another contrast variant of the correspondence problem studied by percept to change from that obtained in the control
Smallman and McKee (1995). See text for details. experiment to that obtained in the last two experiments.
Smallman and McKee (1995) determined that this crit-
ical difference was best described in terms of a ratio
near disparity plane and the right lying in a far dis-
constraint on the magnitudes of the contrast of the in-
parity plane, as depicted by plots in the top row of
puts to the two eyes, where the exact value of the critical
Fig. 9a. The model simulation explains this percept as
ratio varied between subjects. Fig. 10 shows the maxi-
follows.
mum contrast difference between the two eyes that still
The model asserts that the left bar of the left input
allows the model to perform stereoscopic fusion, plotted
cannot be matched with either of the bars of the right
on logarithmic axes. ‘‘’’ designates a data point ob-
input because its contrast differs too greatly from theirs.
tained when the odd bar had a lower contrast than the
Instead, the right bar of the left input matches with both
other three bars (i.e., the situation depicted by Fig. 9a)
bars of the right input, forming near and far disparity
and ‘‘+’’ the converse situation (Fig. 9b). The line of best
vertical boundary representations in V1, as shown by
fit for the two sets of data combined is a straight line of
the second and fourth plots of the second row. As de-
slope equal to 1, indicating that the modelÕs behavior is
tailed in Section 3.2.1, the monocular boundary repre-
consistent with the ratio rule reported by Smallman and
sentations are added to all depth planes in the V2
McKee (1995).
disparity filter along their respective lines-of-sight, as is
shown by the plots in the third row of this figure. The V1
binocular boundary representations are also added to 3.2.4. Exception to the ratio rule
the V2 disparity filter, coinciding with the middle bar Smallman and McKee (1995) also discovered one
representation of the second plot and the right bar notable exception to the ratio rule: When each eye sees
representation of the fourth plot. These two sets of only a single bar, these bars match regardless of their
boundaries, being stronger, then suppress, via the re- contrast difference. This situation was simulated as an
current line-of-sight inhibition of the V2 disparity fil- example of dichoptic masking without retinal corre-
ter, those vertical boundary representations that share spondence, and the reader is referred back to Section
their lines-of-sight. They cannot, however, suppress the 3.1.1 (Fig. 5) for an explanation.
Fig. 11. Simulation of the Venetian blind effect (Howard & Rogers,
1995).
Fig. 10. Simulation of the ratio constraint on stereoscopic fusion. On

logarithmic axes the contrast of the higher contrast bar is plotted simply. Instead, we will divide the stimulus into two
against the minimum contrast of the low contrast bar that can still be components and consider these separately.
fused with it. Crosses (+) denote the situation where one bar has a
higher contrast than the other three bars (e.g., Fig. 9a) and circles ()
First, we note that every second bar of the left input is
the converse (e.g., Fig. 9b). The line of best fit for the total data set has in retinal correspondence with every third bar of the
a slope of 1. The model thus obeys the ratio constraint on stereoscopic right input. We extract these bars to form the stimulus
fusion. shown in the middle two plots of the first row of Fig.
12a. Since the bars in the left and right inputs of this
figure are in retinal correspondence, the model correctly
predicts that they will appear in the zero disparity plane,
3.3. The Venetian blind effect
as shown by the middle plot of the fifth row.
We now consider the remaining bars, which are
We continue our discussion of the correspondence
shown in Fig. 12b. The right eye sees exactly twice the
problem by considering the Venetian blind effect. A
number of bars as the left eye. This is therefore an ex-
Venetian blind stereogram is shown in Fig. 7.21 of
ample of PanumÕs limiting case that was considered
Howard and Rogers (1995) and consists of two gratings,
Section 3.1.4. As before, the model predicts that each
a low frequency one that is presented to the left eye, and
bar of the left input is fused with two bars of the right
a high frequency one presented to the right. When fused,
input to generated the percept shown in the top row of
the frequency of the gratings are such that every second
this figure.
bar of the left grating is in retinal correspondence with
Adding together the percepts shown in top rows of
every third bar of the right grating. The stimulus is
Fig. 12a and 12b, we achieve the percept shown in the
reproduced in the middle two plots of the first row of
top row of Fig. 11, thereby explaining the Venetian
Fig. 11.
blind effect. The insight that the model provides is that
According to Howard and Rogers, this stereogram
the Venetian blind effect is just a complex version of the
produces a percept of short ramps, each containing
correspondence problem and PanumÕs limiting case,
three bars, sloping up from left to right interspaced with
when it is properly understood by combining early ste-
steep returns. The total percept is that of a Venetian
reo matching, later selection by a disparity filter, and
blind. The model is able to correctly predict this percept,
surface filling-in of those regions that are completely
as shown by the top row of Fig. 11. Numbering from left
enclosed by boundaries.
to right, this row shows that the first bar of the percept is
in the zero disparity plane, the second in the near dis-
parity plane, then there is a step return to the third bar 3.4. Stereopsis with opposite-contrast stimuli
which is located in the far disparity plane after which the
pattern repeats. Polarity-reversed stereograms are those stereograms
Although the model is able to correctly simulate the where corresponding elements in the two stereo half im-
percept, this simulation is too complicated to explain ages have opposite luminances. In other words, for every
Fig. 12. (a) Simulation of one component of the Venetian blind effect. Fig. 13. Simulation of stereopsis with a polarity-reversed stereogram.
(b) Simulation of the other component. See text for details.
white element in one stereo half image there is a corre-

sponding black element in the other stereo half image. contains regions completely enclosed by boundaries,
The model simulation of the depth percept induced by a which is why surfaces are perceived only in this disparity
polarity-reversed stereogram is shown in Fig. 13a. plane, as shown by the fifth row.
Here the left eye sees a black bar and the right eye a This simulation suggests that the stereoscopic depth
white bar both on the same gray background. Since they perception induced by polarity-reversed stereograms is
have the same contrast polarity, the left edge of the white mediated by the fusion of those edges in the two stereo
bar fuses with the right edge of the black bar to form a half images that have the same contrast polarity. In
vertical boundary in a far disparity plane of V1, shown particular, the model predicts that the degree of the in-
by the fourth plot of the second row. Even though they duced depth should be completely determined by the
have the same contrast polarity, the right edge of the disparity of these fused edges. An alternative prediction
white bar cannot fuse with the left edge of the black bar would be that, since most of the display is perceived only
because they are too disparate. As always, the monocular monocularly, and since monocular objects tend to be
boundaries are added to V2 along their respective lines- perceived to lie in the fixation plane (Krol & van de
of-sight, shown by the plots in the third row. The vertical Grind, 1983), the depth perceived should be biased to-
binocular V1 boundary is also added to V2 and coincides wards the fixation plane. Howe and Watanabe (in press)
with the middle vertical boundary in the far disparity ran a series of psychophysical experiments to investigate
plane represented by the fourth plot. This boundary, which of these two predictions were true, and found that
being stronger, then suppress, via the line-of-sight inhi- the prediction of the model was a better description of
bition of the V2 disparity filter, the middle two vertical the degree of depth experienced by subjects.
boundaries in all other disparity planes of this row, re- Suppose, however, that the subjectÕs vergence were to
sulting in the final V2 boundary representations shown in change so that the stimulus became that shown in Fig.
plots of the fourth row. Only the fourth plot of this row 13b. Now, because they have the same contrast polarity,
the right edge of the white bar fuses with the left edge of situations are often caused by each eye viewing the
the black bar to form a boundary representation in the world from a slightly different position, leading to par-
far disparity plane of V1; see the fourth plot of the tial occlusions where part of a scene is visible to only one
second row. Also, the left edge of the white bar fuses eye. The model clarifies how the percept of depth caused
with the right edge of the black bar to form a boundary by such stimuli can be explained in terms of monocular–
representation in the near disparity plane of V1; see the binocular interactions.
second plot of the second row. Unlike Fig. 13a, there are
now two boundary representations in V1. The monoc-
3.5.1. Stimuli of Nakayama and Shimojo (1990)
ular boundaries are added to V2 along their respective
In this set of experiments, a thick bar was presented
lines-of-sight. The binocular boundaries are also added
to both eyes and a thin bar only to the right eye, as
to V2, overlapping with the middle vertical boundaries
shown in the first row of plots of Fig. 14a. Subjects re-
in the second and fourth plots of the third row. These
ported perceiving the thin bar behind the thick bar, at a
two boundaries, being stronger, suppress all other ver-
depth that was consistent with the right edge of the thin
tical boundaries via the recurrent inhibition of the V2
bar of the right input being fused with the right edge of
disparity filter. However, because they are equally
thick bar of the left input.
strong, they cannot suppress each other. The final
The model explanation is as follows. The vertical
boundary representations are shown in fourth row. No
boundaries of the thick bar are registered binocularly in
regions are completely enclosed by boundaries and so
the near disparity plane in V1, as shown by the second
the model predicts that there will be no stable depth
plot of the second row, and the right edge of the thin bar
percepts.
is matched with the right edge of the thick bar to be
This prediction is correct in as far as it goes, in that
registered binocularly in the far disparity plane in V1, as
subject do not achieve any stable surface percepts, but
shown by the fourth plot. The left edge of the thin bar is
in practice unstable surface percepts may form if sub-
registered only monocularly because it cannot be mat-
jects experience binocular rivalry. Describing binocular
ched with either of the edges of the left input. As usual,
rivalry is beyond the scope of our simulations. However,
the monocular boundaries are added to all depth planes
it has been qualitatively modeled in Grossberg (1987) in
a manner that is consistent with the present model
simulations.
The key point here is that whether or not an anti-
correlated stereogram induces a stable depth percept
depends on the vergence of the subject. One vergence
position enables the visual system to match the left and
right inputs only in a single way. Other vergence posi-
tions lead to two binocular boundaries in V1, and
consequently no stable depth percepts in V4, as dem-
onstrated by Fig. 13b. Subjects may also be able to use
attention to choose between the two possible ways of
matching the left and right inputs. Section 4.4 shows
how the model may be extended to incorporate atten-
tional effects.
Regardless of whether subjects use vergence or at-
tention to make sure their visual system can only fuse
the left and right inputs in one way, as more elements
are included in the left and right inputs, the harder it is
to ensure unambiguous fusion. The model suggests that
this is the reason why complex anticorrelated stereo-
grams (i.e., those anticorrelated stereograms that con-
tain many separate elements) induce little or no depth
perception whereas simple anticorrelated stereograms
do (Howard & Rogers, 1995; Julesz, 1971).
3.5. Da Vinci stereopsis

Fig. 14. (a) Simulation of the depth percept invoked by the conven-
Da Vinci stereopsis describes those situations where a tional da Vinci stereopsis stimuli of Nakayama and Shimojo (1990).
monocular object has a definite depth conferred to it by (b) Simulation of the depth percept invoked by the polarity-reversed
its relationship to a binocularly viewed object. Such da Vinci stereopsis stimuli. See text for details.
in the V2 disparity filter along their respective monoc-

ular lines-of-sight, as shown by the plots in the third row
of this figure. The vertical binocular boundaries are also
added to the disparity filter, overlapping with the ver-
tical boundaries of the thick bar representation in the
second plot and with the rightmost vertical boundary in
the fourth plot. These vertical boundaries, being stron-
ger, eliminate all other vertical boundaries that share
their lines-of-sight via the disparity filterÕs line-of-sight
inhibition. However, they do not eliminate the vertical
boundaries originating from the left edge of the thin bar
because these do not share any of their lines-of-sight.
The final V2 boundary representations are shown in the
fourth row. As usual, V4 fills in surfaces in those regions
that are completely enclosed by a connected boundary.
This produces a percept of a thick bar in a near disparity
plane, represented by the second plot of the top row,
and a thin bar in a far disparity plane, represented by
the fourth plot. The very small squares seen in the top
row are artifacts of the implementation of the diffusion
process with a relatively small number of pixels and
have no physiological significance. In particular, they
disappear when the simulations are carried out at a
sufficiently high resolution, at a high computational
cost. The model therefore correctly predicts that the thin
bar will appear behind the thick bar at a depth that is
consistent with the right edge of the thin bar being ste-
reoscopically fused with the right edge of the thick bar, Fig. 15. (a) Simulation of depth percept invoked by the da Vinci ste-
as has been reported experimentally (Nakayama & reopsis stimuli of Gillam et al. (1999). (b) Simulation of a variant of the
Shimojo, 1990). original Gillam et al. stimuli.
Nakayama and Shimojo (1990) also showed that the
percept remained the same even when the monocular
and binocular bars had opposite luminance polarities: bars, the left in the near disparity plane and other in the
specifically, the binocular bar being white and the far disparity plane. Gillam et al. suggested that, because
monocular bar being black. It might be thought that the right eye input contains a gap not present in the left
the model could not simulate this observation, since the eye input, this display demonstrates that stereopsis can
model can only binocularly fuse edges that have the be induced by monocular gaps. It should be stressed
same contrast polarity. The modelÕs successful simula- that, although they are superficially very similar, this
tion is shown in Fig. 14b. display is quite different to that of Nakayama and
The only difference between this simulation and the Shimojo (1990) (Fig. 14) which demonstrated an entirely
previous simulation is that in the previous simulation it different point: that depth perception could be deter-
was the right edge of the thin bar of the right input that mined by the separation of a monocular bar from a
fused with the right edge of the thick bar of the left input binocular bar.
whereas in this simulation it is the left edge of the thin The model explanation is as follows. The model
bar that fuses with the right edge of the thick bar of the suggests that the left edge of the thick bar fuses with the
right input since these two edges now have the same left edge of the left thin bar to appear in a near disparity
contrast polarity. This simulation then proceeds in the plane in V1, represented by the second plot of the sec-
same manner as the previous one, thereby explaining the ond row, while the right edge of the thick bar fuses with
similar percept. Contrary to the claims of Nakayama the right edge of the right thin bar to appear in a far
and Shimojo (1990), this simulation shows that the re- disparity plane in V1, represented by the fourth plot of
ported percept can be generated without independent this row, since in both cases these edges have the same
knowledge of occlusion relationships (see Section 4.2). contrast polarity. The two other vertical edges of the
thin bars of the right input are registered only mon-
3.5.2. Stimulus of Gillam et al. (1999) ocularly because they cannot be matched to either of the
In Fig. 15a the right eye sees two thin bars and the left edges of the left input. As usual, the V1 monocular
eye a single thick bar. Subjects report seeing two thin boundary representations are added to all depth planes
in the V2 disparity filter along their respective lines-of- Todorovic (1988) to explain several lightness illusions. It
sight. This is why two thin bar representations and one is therefore claimed that the present model can explain
thick bar representation are seen in all disparity planes the same large set of lightness illusions. (See Grossberg
of the third row, with the slight complication that in all and Kelly (1999), Grossberg and Pessoa (1998), Kelly
cases the thick bar representation overlaps with at least and Grossberg (2000) and Pessoa, Mingolla, and Neu-
one of the two thin bar representations. The V1 binoc- mann (1995) for other articles that explain additional
ular boundary representations are also added to the V2 lightness and brightness data using this filling-in mech-
disparity filter, overlapping with the leftmost vertical anism.)
boundary in the second plot and the rightmost vertical The Todorovic–OÕBrian–Cornsweet effect (COCE) is
boundary in the fourth plot. These vertical boundaries, simulated to illustrate this claim. In Grossberg and
being stronger, inhibit, via the recurrent line-of-sight Todorovic (1988), the COCE was simulated using only a
inhibition of the disparity filter, all the other vertical monocular input. The simulation herein uses inputs to
boundaries that share any of their lines-of-sight. This both eyes and shows that the binocular model can also
means that they do not inhibit those vertical boundary simulate this percept. The stimuli are shown in the
representations originating from the two monocularly middle two plots of the bottom row of Fig. 16. Both eyes
viewed edges of the right input because these vertical see the same stimulus, which consists of two abutting
boundaries do not share any of their lines-of-sight. regions of the same uniform lightness separated by a
The final V2 boundary representations are shown in the lightness cusp. Subjects report perceiving both regions
fourth row. V4 fills-in surfaces in those regions that are as having uniform lightness, with the left region ap-
completely enclosed by boundaries, resulting in the pearing darker than the right.
percept of a thin near bar and a thin far bar, as reported The model explains the COCE as follows. The input
by human subjects (Gillam et al., 1999). is binocularly fused to form three vertical binocular
In the previous display, at least one edge of each re- boundaries in the nearest disparity plane of V1, repre-
gion could be binocularly fused. In contrast, in Fig. 15b sented by the leftmost plot of the second row. As always,
the middle bar of the right eye stimulus is perceived both the V1 binocular and monocular boundaries are
entirely monocularly. added to the V2 disparity filter, with the monocular
The model simulation is as follows. The left eye sees a boundaries being added to all depth planes along their
single bar while the right eye sees three separate bars. The respective lines-of-sight, as shown by the plots in the
left edge of the bar of the left input again fuses with the third row. The vertical boundaries in the nearest dis-
left edge of the leftmost bar of the right input to form a parity plane are stronger because they receive both
binocular boundary in the second plot of the second row. monocular and binocular inputs. They therefore inhibit
Similarly, the right edge of the bar of the left input again the vertical boundaries in the other disparity planes via
fuses with the right edge of the rightmost bar of the right the recurrent line-of-sight inhibition of the disparity
input to form a binocular boundary in the fourth plot of filter. The final V2 boundaries are shown by the plots in
the second row. Again the monocular boundaries are the fourth row. The boundaries in the nearest disparity
added to V2 along their respective lines-of-sight, as plane confine the V4 diffusion of the lightness signals
shown by the third row. The binocular V1 boundaries
are also added to V2. The binocular boundary in the
second plot of the second row overlaps with the first
vertical boundary in the second plot of the third row.
Similarly, the binocular boundary of the fourth plot of
the second row overlaps with the last vertical boundary
in the fourth plot of the third row. The surviving V2
boundaries are shown in the fourth row.
Only those boundaries that completely enclose a re-
gion can contain the lightness signals that originate at
the location of the boundaries, and so only these regions
give rise to surface percepts in V4. The model therefore
correctly predicts that three surfaces will be seen, each at
a different depth as reported experimentally (Gillam
et al., 1999).
3.6. Lightness illusions
The filling-in mechanism utilized by the model V4 Fig. 16. Simulation of the Craik–OÕBrian–Cornsweet lightness illu-
simulations is equivalent to that used by Grossberg and sion. See text for more details.
that originate at the edges of the regions. Those lightness of its cells are simple (Hubel & Wiesel, 1968; Schiller,
signals originating from the left side of the cusp are Finlay, & Volman, 1976).
darker than those originating from the right side. This As discussed in Section 2.1, the model assumes that
lightness difference is propagated, by the V4 filling-in polarity-specific binocular matching occurs in layer 3B.
mechanism, throughout the respective regions, causing This is consistent with observations that a significant
the left region to appear uniformly darker than the right, proportion of layer 3B comprises simple cells (Dow,
as shown by the leftmost plot of the top row. 1974), that layer 3 contains a significant number of
binocular cells (Hubel & Wiesel, 1968; Poggio, 1972),
and that projections to it can be independent of ocular
4. Discussion dominance (Katz, Gilbert, & Wiesel, 1989).
The model suggests that binocular layer 2/3A cells
4.1. Supporting physiological and anatomical data pool responses from layer 3B cells of both contrast po-
larities so that they can represent the boundaries of
This section shows that all the relevant physiological objects whose contrast polarity, with respect to the
and anatomical data of which we are aware support the background, changes as the boundary is transversed. In
model. The model does not, however, consider cortical keeping with this suggestion, it is known that layer 3B
areas V3, V3A and MT, even though there is evidence projects throughout layer 2/3A (Callaway, 1998), and
that these areas play a role in depth perception (e.g., that layers 2 and 3 each contain significant numbers of
Backus, Fleet, Parker, & Heeger, 2001). These areas binocular and complex cells (Poggio, 1972).
were not needed to simulate the modelÕs targeted data. The model further suggests that there is a group of
The function of area V3A appears to be particularly cells in layer 2/3A and 3B that respond only to binoc-
controversial, with studies suggesting that it is variously ular, and not to monocular, stimulation. Such ‘‘obligate
concerned with relative disparity (Backus et al., 2001), cells’’ are known to exist in macaque V1 (Poggio & Fi-
saccades (Nakamura & Colby, 2000a, 2000b) and pre- scher, 1977; Smith, Chino, Ni, & Cheng, 1997), with
hensile hand movements (Nakamura et al., 2001). As a about 40% of tuned excitatory neurons being obligatory
further complication, there is some evidence that the (Poggio & Talbot, 1981), including almost all ‘‘tuned
function of macaque V3A differs from that performed zero’’ neurons (Poggio, 1991). Obligate cells do not
by human V3A (Tootell et al., 1997). appear to be as prevalent in cat (Anzai, Bearse, Free-
When the model diagram in Fig. 1 is compared to the man, & Cai, 1995).
list of data below, it can be seen that the model makes The model predicts that all these interactions occur in
predictions concerning brain physiology and anatomy the V1 interblob regions, which is in keeping with ob-
beyond what is known. One prediction is that there is an servations that V1 interblobs are highly selective for
inhibitory circuit in V1 which causes the binocular cells orientation but relatively unselective for color (Merigan
in layers 3B and 2/3A not to respond if the inputs to the & Maunsell, 1993).
left and right eyes differ too greatly in contrast. Another
is that there is a disparity filter in V2 that employs line-
4.1.2. V1 monocular boundaries
of-sight inhibition. A third prediction is that there is a
The model suggests that the V1 monocular bound-
surface filling-in mechanism that leads to visible per-
aries are formed by a process that is a simplification of
cepts and is located in V4 (among other places; see
that which forms the V1 binocular boundaries. Conse-
Grossberg, 1994). This section should be read in con-
quently, much of the above data applies equally to the
junction with Fig. 1, which interprets each model stage
monocular boundaries network. Additional support
anatomically.
for this network comes from observations that layer 3
(Hubel & Wiesel, 1968; Poggio, 1972) and layer 2
4.1.1. V1 binocular boundaries
(Poggio, 1972) of V1 each comprise a large proportion
Consistent with the model, the LGN contains circu-
of monocular cells.
larly symmetric on-center, off-surround receptive fields
(Kandel, Schwartz, & Jessell, 2000, pp. 529). LGN le-
sion studies have shown that the parvocellular, but not 4.1.3. V2 boundaries
the magnocellular, pathway is critical for fine stereopsis The model assumes that the V2 boundaries are lo-
(Schiller, Logothetis, & Charles, 1990a, 1990b). Just as cated in the V2 pale stripes. This is consistent with ob-
V1 layer 4 is the major recipient of this parvocellular servations that the V2 pale stripes receive the major
input in vivo (Callaway, 1998), it is also the input layer projection from the V1 interblob regions, while receiving
of model V1. Also, in accord with the model, layer 4 is no significant projection from the V1 blob regions, and
known to output to layer 3B, but not to layer 2/3A, of are highly orientationally selective (Roe & TsÕo, 1997),
V1 (Callaway, 1998), a large proportion of it is mon- while also containing a complete map of visual space
ocular (Hubel & Wiesel, 1968; Poggio, 1972), and many (Roe & TsÕo, 1995).
The model is further consistent with data that V2 is that, at least for certain stimuli, binocular neurons that
mainly binocular (Hubel & Livingstone, 1987; Roe & are tuned to different spatial frequencies will respond to
TsÕo, 1997), is mainly disparity-sensitive (Poggio & different false matches. Consequently they argue that
Fischer, 1977; von der Heydt, Zhou, & Friedman, false matches can be eliminated simply by pooling the
2000), contains many complex cells (Hubel & Living- responses of several binocular neurons, each tuned to a
stone, 1987), receives input into layer 4 (Rockland & different spatial frequency. Although they demonstrated
Virga, 1990) and outputs to V4 (Xiao, Zych, & Fell- the proficiency of their model when it was presented
eman, 1999), which itself is highly selective for disparity with white noise stimuli, it is not clear how their model
(Merigan & Maunsell, 1993). In addition, the V2 pale could be extended to other stimuli, in particular those
stripes are disparity-selective (Peterhans, 1997). situations were contrast affects the perceived solution of
According to the model, an important function of V2 the correspondence problem (Section 3.2) or where
is to suppress false matches by utilizing a disparity filter. monocular information contributes to depth perception
This is consistent with observations that cells readily (Section 3.5).
exhibit false matches in V1 (Cumming & Parker, 2000), Another way to solve the correspondence problem is
but not in V2 (Bakin et al., 2000). to utilize a disparity filter that implements the unique-
matching rule, which states that any given feature in one
4.1.4. Surfaces retinal image is matched at most with one feature in the
Surfaces are built up through interactions between other retinal image (Grimson, 1981; Marr & Poggio,
the V1 blobs, the V2 thin stripes, and V4, consistent with 1976; for a review see Howard & Rogers, 1995, pp. 42–
the fact all these regions are linked by major projections 43). As discussed in Section 1, this rule fails in PanumÕs
(Livingstone & Hubel, 1984; Xiao et al., 1999), that the limiting case (Gillam et al., 1995; McKee et al., 1995;
V2 thin stripes are the least orientationally selective area Panum, 1858).
of V2 (Peterhans, 1997) and contain a complete map of This failure caused Grossberg and McLoughlin
visual space (Roe & TsÕo, 1995). (1997) and McLoughlin and Grossberg (1998) to design
a disparity filter that encouraged unique matching
4.2. Comparison with other theories and models without enforcing it. Their model forms the foundation
for our own and can simulate much of the same data,
One of the most popular explanations of monocular– including most of the dichoptic masking and the corre-
binocular interactions is the ecological optics hypothesis spondence problem data. Their model also makes an
of Nakayama and Shimojo (1990). This hypothesis incorrect psychophysical prediction: that if each eye sees
suggests that visual systems attempt to interpret un- a single bar, then the ratio constraint on stereoscopic
paired image points in terms of occlusion. For example, fusion (Smallman & McKee, 1995) ensures that fusion
in Fig. 14, both eyes see a thick bar but only the right will occur only if the magnitudes of the contrasts of the
eye a thin bar. According to the ecological optics hy- two bars do not differ too greatly. This is inconsistent
pothesis, the visual system interprets these stimuli by with experimental findings which indicate that the ratio
assuming that the thin bar is located behind the thick constraint does not apply to this special case (McKee
bar at the exact distance that would cause the thick bar et al., 1994; Smallman & McKee, 1995).
to hide it from the left, but not from the right, eye. The present model refines the Grossberg and
While this hypothesis is consistent with the percepts McLoughlin model to correct this short-coming. In
evoked by the stimuli in Figs. 14 and 15, it cannot ex- particular, for the purposes of the disparity filter, the
plain the percept evoked by the stimuli of Fig. 13, be- Grossberg and McLoughlin model assigned all unfused
cause this stimulus cannot be explained in terms of boundaries to the fixation plane, whereas the present
occlusion. If we wish to understand the response of the model adds unfused boundaries to all fixation planes
visual system to all possible stimuli, not just the ones and then lets the V2 disparity filter eliminate boundary
that can be interpreted in terms of occlusion, then it is representations as necessary. As explained in Section
necessary to offer a mechanistic account that can deal 3.1.1, this procedure allows in the special case where
with a broader data set in a unified way, as the present each eye sees only a single bar the two bars to be bin-
model does. ocularly fused regardless of their contrast difference. The
One of the most successful mechanistic models of present model has simulated all the data considered by
stereopsis is the disparity energy model (Ohzawa et al., McLoughlin and Grossberg (1998), specifically the data
1990). However, this model does not solve the corre- on contrast variations of dichoptic masking and the
spondence problem in that it may match vertical con- correspondence problem, and has also simulated addi-
tours in the two retinal images that correspond to tional data including the Venetian blind illusion, four
different objects. Fleet, Wagner, and Heeger (1996) have different examples of da Vinci stereopsis (Gillam et al.,
proposed how the disparity energy model could be ex- 1999; Nakayama & Shimojo, 1990), stereopsis with op-
tended to avoid this problem. In their paper they note posite-contrast stimuli, the effect of interocular contrast
differences on stereoacuity and the Craik–OÕBrian– both the correspondence problem and the monocular–
Cornsweet lightness illusion. Furthermore, unlike its binocular interface problem by utilizing its disparity fil-
predecessor, it has been mapped onto known cortical ter. The correspondence problem arises because V1
cells and laminar circuits within cortical areas V1, V2 sometimes incorrectly fuses contours that belong to
and V4. different objects. The monocular–binocular interface
The model presented in this article and the Grossberg problem is caused because the V1 monocular bound-
and McLoughlin model both instantiated key aspects of aries, not having a definite depth association, are initially
FACADE theory. Another model, which is also a sim- added to all depth planes. Finally, the surface network
plified version of FACADE theory, was used to explain includes cells in V4, the V2 thin stripes and the V1 blobs.
a series of experiments on the McCollough effect, an It is necessary because it is surface percepts, not
orientation-sensitive, long-lasting, chromatic after-effect boundary percepts, that subjects report in the experi-
(Grossberg, Hwang, & Mingolla, 2002). Unlike the mental studies considered by this paper and also because,
present model, the binocular cells in the McCollough as illustrated by all of the simulations, not all boundaries
effect model did not exhibit the ‘‘obligate’’ property in give rise to a percept of depth.
that they responded to monocular inputs, albeit less If anything, the model in Fig. 1 is too simple to ex-
strongly than to binocular inputs. Although, the plain all data about depth perception. Fortunately, the
McCollough effect model did not make use of obligate analysis in this article has opened a clear path to gen-
cells, inserting such cells into the model would not dis- eralize the model, as illustrated below.
rupt its simulations. These obligate cells would merely
be unnecessary. Similarly, the addition of non-obligate
binocular cells into V1 of the present model, while un- 4.4. Generalizing to natural images, 3D boundary com-
necessary, would not reduce its explanatory power. In pletion and 3D attention
particular, such an addition would merely increase the
number of false matches that occur in V1. These false One of the long-term goals of this modeling work is
matches would be eliminated by the V2 disparity filter, to extend the present model so that it can be applied to
in the manner outlined in Section 2.3, and so would not natural images. The simulations already done show that
contribute to the final percept. Taken together, these the model can resolve a wide range of potentially con-
two models help explain the differing roles of obligate fusing false matches. There remain, however, two im-
and non-obligate binocular cells in the broader context pediments that the model first needs to overcome.
of FACADE theory and help to functionally explain First, the present model can represent only 3D planes
why both obligate and non-obligate cells have been that are flat and perpendicular to the observer. To ana-
found experimentally to exist (Poggio, 1991). lyze natural images, the model needs to be extended to
represent slanted and curved surfaces in 3D. A parallel
4.3. Model robustness and complexity line of research has begun to demonstrate how it can be
consistently generalized to explain such data (Grossberg
The model is robust in the sense that the absolute & Swaminathan, 2003; Swaminathan & Grossberg,
values of the model parameters can be varied over large 2001).
ranges without disrupting its explanations of data; only Second, the present model shows how boundaries can
their values relative to each other are important. Fur- be formed using bottom-up inputs from the outside
thermore, there is considerable scope when choosing world. It does not, however, indicate how horizontal
individual parameter values, since no single parameter interactions can be used to complete these boundaries
proves to be critical in any simulation. where pixels are missing either due to internal brain
The model is minimally complex in the sense that each imperfections, such as the blind spot in the retina, or due
of its four interacting networks, V1 binocular bound- to incomplete contours in external inputs, whether due
aries, V1 monocular boundaries, V2 boundaries, and V4 to noise, occluding surfaces, spatially discrete texture
surfaces, are essential. The V1 binocular boundaries elements, illusory contour stimuli, or even missing pixels
network is needed to explain stereopsis and the contrast in impressionist paintings. Nor does it clarify how these
ratio constraint observed in stereoscopic fusion (Small- circuits can develop, be modified by learning, or mod-
man and McKee, 1995). The V1 monocular boundaries ulated by top-down attention. This omission can be
network plays a role in explaining da Vinci stereopsis overcome as follows.
(Gillam et al., 1999; Nakayama & Shimojo, 1990), A parallel line of modeling has developed quantita-
dichoptic masking (McKee et al., 1994), contrast varia- tive explanations and simulations of how processes of
tions of the correspondence problem (Smallman and perceptual development, learning, grouping, and atten-
McKee, 1995) as well as some examples of stereopsis tion may be achieved by laminar cortical circuits
with opposite-contrast stimuli (e.g., Howe & Watanabe, (Grossberg, 1999a, 1999b; Grossberg et al., 1997; Gross-
in press). The V2 boundaries network is needed to solve berg & Raizada, 2000; Grossberg & Williamson, 2001;
Raizada & Grossberg, 2001). This LAMINART model

did not, however, investigate how these boundaries may
be completed in three-dimensions. We now show that
the LAMINART model of boundary completion is
consistent with the model of three-dimensional bound-
ary formation that is summarized in Fig. 1. In particu-
lar, we demonstrate below how the LAMINART model
be can generalized to cope with the positional dis-
placement, or allelotropia, that is characteristic of bin-
ocular fusion (cf., Fig. 3).
Fig. 17 summarizes some of the key LAMINART
interactions that govern perceptual grouping, and
attention, without regard to its three-dimensional re-
presentation. Fig. 18 shows how the LAMINART
perceptual grouping and attention circuit naturally
generalize to a 3D LAMINART model that is consistent
with the system interactions in Fig. 1. This extended
model clarifies how 3D boundaries can be completed
and how attention can be selectively paid to objects in
3D. The following new features in Fig. 18 show how
these properties obtain. First, layer 4 no longer directly
activates layer 2/3, as in Fig. 17. Instead, layer 4 simple
cells first activate layer 3B simple cells, which in turn
activate layer 2/3A complex cells, as shown in Fig. 1.
The layer 2/3A cells can then interact via horizontal Fig. 17. The LAMINART model: (a) The LGN directly activates V1
layers 4 and 6. Layer 6, in turn, sends a pattern of on-center, off-
interactions, like those summarized in Fig. 17c and e, to
surround inputs to layer 4. These layer 6 inputs can strongly inhibit
complete boundaries. Second, binocular cells in layer layer 4 through the off-surround, but the excitatory and inhibitory
2/3A can represent different disparities, and thus differ- inputs in the on-center are approximately balanced so that layer 6 can
ent relative depths from an observer. Interactions be- modulate the excitability of layer 4 cells, but not fully drive them to fire
tween layer 2/3A cells that represent the same relative vigorously. The direct connections from LGN to layer 4 carry out this
driving function. (b) This layer 6-to-4 circuit can be used by top-down
depth from the observer can be used to complete signals from V2 layer 6 to attentionally modulate the excitability of V1
boundaries between object contours that lie at the same layer 4 cells. (c) Boundary completion can occur when layer 4 cells
depth. activate layer 2/3 cells, which communicate with their layer 2/3
Due to the binocular fusion that occurs in layer 3B, neighbors via long-range horizontal excitatory connections and
the binocular boundaries that are formed in layers 3B shorter-range inhibitory interneurons. The balance between these ex-
citatory and inhibitory interactions allows boundaries to form in-
and 2/3A can be positionally displaced, or shifted, rel-
wardly between properly oriented image contrasts, as in the case of
ative to their monocular input signals to layers 6 and 4. many illusory contours, but not outwardly from individual contrasts.
Fig. 17c suggests that these layer 2/3 boundaries feed The strongest boundary groupings in layer 2/3 can support themselves
signals back to layer 6 in order to select the winning best through the positive feedback loop between layers 2/3-to-6-to-4-
groupings that are formed in layer 2/3. How can the to-2/3, even as their strong inhibitory signals in the layer 6-to-4 off-
surround can inhibit weaker groupings. (d) A top-down on-center,
positionally displaced binocular boundaries in layer
off-surround from V1 layer 6 to the LGN acts like the top-down sig-
2/3A of Fig. 17 contact the correct monocularly acti- nals from V2 layer 6 to V1 layer 4. (e) The LAMINART system ar-
vated cells in layers 6 and 4, so as to complete the chitecture. Note that the horizontal interactions within V2 layer 2/3
feedback loop 2/3A-to-6-to-4-to-3B-to-2/3A that can can have a broader spatial extent than those in V1 layer 2/3. The
select the winning three-dimensional groupings? In longer-range V2 interactions carry out the type of perceptual group-
ings that are familiar in illusory contours, texture grouping, comple-
particular, how can the feedback signal from a layer 2/ tion of occluded objects, and bridging the blind spot. (Reprinted with
3A cell that is positionally displaced with respect to its permission from Grossberg & Raizada (2000).)
monocular inputs activate horizontal signals that can
activate the correct layer 6 monocular sources?
We propose that horizontal connections that are and Williamson (2001) have simulated how layer 2/3
known to occur in layer 5 (Callaway & Wiser, 1996) connections and layer 6-to-4 connections may originate
accomplish this. Feedback signals from layer 2/3A during development. The selective layer 5-to-6 contacts
propagate vertically to layer 5, whose cells activate hor- are proposed to form according to similar laws. In
izontal axons in this layer that contact the appropriate summary, inward horizontal layer 4-to-3B and 2/3A-
layer 6 cells. These layer 5-to-6 contacts are assumed to to-2/3A connections are proposed to form binocular
be selectively formed during development. Grossberg cells and their groupings, while outward layer 5-to-6
A.1. LGN
The LGN cells obey membrane, or shunting, equa-

tions that receive input from the retina and are assumed
to have circularly symmetric on-center, off-surround
receptive fields. When these fields are approximately
balanced, the network discounts the illuminant and
contrast-normalizes its cell responses (Grossberg &
Todorovic, 1988). The LGN cell membrane potentials,
L=R
Xij , obey the following differential equation:
dXij
L=R
L=R L=R L=R L=R
X
L=R
¼ eXij þ ða Xij ÞIij Xij gpqij Ipq ;
dt p6¼i;q6¼j
ðA:1Þ
where L=R designates that the cell belongs to the left or

right monocular pathway, indices i and j denote the
position of the input on the retina, e is a constant (105 )
Fig. 18. A 3D LAMINART model, including 3D boundary comple- that represents the rate of decay of the cell membrane
tion and attention, as well as the binocular and monocular interactions
potential, a is a constant (9.9) that represents the max-
summarized in Fig. 1. See text for details. L=R
imum membrane potential, Iij is the unnormalized lu-
minance of the left or right retinal image, and gpqij is
a Gaussian kernel that represents the inhibitory off-
connections are proposed to close the feedback loops surround:
that help to select the correct three-dimensional group- !
ings. ðp iÞ2 þ ðq jÞ2
gpqij ¼ exp ; ðA:2Þ
Once the generalization to 3D boundaries is 2r2
made, top-down attentional modulation of these bound-
aries follows directly by using the same circuits as in where r represents the size of the kernel (1.5). The
Fig. 17. steady-state cell membrane potentials of these cells are
given by:
L=R
Appendix A. Model equations L=R aIij
Xij ¼ P L=R
: ðA:3Þ
eþ p;q gpqij Iij
To make the equations easier to read, capital letters
denote variables and lower case letters denote constants. The steady-state equation (A.3) was used in the simu-
Appendix A is most easily read in conjunction with Fig. lations. Below all equations that were solved at steady-
1, which depicts the model. Each eyeÕs stimulus was state are given in their steady-state form.
presented on a grid 55 units high and 70 units wide
except Figs. 11 and 12 which were presented on a 55 by
A.2. V1 layer 4 simple cells
126 grid and Fig. 14b which was presented on a 55 by 85
grid. In all simulations, white had a luminance value, in
All cells in V1 layer 4 are modeled as monocular
arbitrary units, of 2. In Figs. 5–9, 11, 12 and 15, the light
simple cells that are sensitive to either dark-light or
gray bars (if any) had a luminance of 0.85 and the dark
light-dark contrast polarity, but not both, depending on
gray bars 0.68. In Figs. 13 and 14 medium gray was
their receptive field structure. At steady-state the mem-
represented by a luminance of 0.75, and black by 0.3. In H =V ;L=R;þ
brane potentials, Sij , of simple cells that respond
Fig. 16, the simulation of the COCE, there was a lu-
to dark-light contrast polarity are given by:
minance cusp which ranged from 0.4 to 0.9. Simulations
X h iþ
were performed using the Matlabâ software package. H =V ;L=R;þ H =V L=R
Sij ¼ kpq Xiþp;jþq ; ðA:4Þ
Analytical equilibrium solutions of the differential p;q
equations were used in all cases except for the V2 dis-
parity filter equation (A.15) and the V4 diffusion equa- where H =V designates that the cell responds to hori-
tion (A.18), which could not be solved analytically, and zontal or vertical boundaries, + indicates that the sim-
were instead solved using EulerÕs method, and then ple cell responds to dark-light contrast polarity,
solved again using a different step size to verify the ac- ½x
þ ¼ maxðx; 0Þ, and kpqH =V
is a Gabor function repre-
curacy of the original solution. senting the simple cell receptive field kernel:
Table 1
The allelotropic shift (s) is the amount that the left and right monocular contours must be displaced to form a single fused binocular contour. It
depends on the disparity. It is zero for matches in the fixation plane because these matches are between contours at retinal correspondence
Disparity (d)
V. Near disparity Near disparity Zero disparity Far disparity V. Far disparity
Allelotropic shift (s) )8 )4 0 +4 +8
Fig. 3 illustrates the allelotropic shift and shows that a left monocular contour needs to be shifted more to the right for matches that are further from
the observer, whereas a right monocular contour needs to be shifted in the opposite direction.
" !# of the inhibition, d is the disparity to which the model

H =V 2Pr 1 p 2 q2
kpq ¼ / sin exp þ ; ðA:5Þ neuron is tuned and s is the allelotropic shift that de-
s 2 r2p r2q pends on the disparity and is defined in Table 1.
where /, s, rp , rq are constants (4.4, 3, 0.6, 0.6) repre-
senting the amplitude and dimensions of this kernel; A.5. Layer 3B binocular cells
r ¼ p for cells that respond to vertical boundaries; and
r ¼ q for those that respond to horizontal boundaries. The layer 3B binocular cells, all of which are verti-
The cell membrane potentials of the simple cells with cally oriented and receive excitatory input from layer 4
light-dark contrast polarity are the inverse of the pre- and inhibitory input from the layer 3B inhibitory cells
vious cell membrane potentials: that correspond to the same position and disparity, have
V ;B;þ=
X h iþ membrane potentials Bijd . At equilibrium:
H =V ;L=R;
Sij
H =V ;L=R;þ
¼ Sij ¼ H =V
kpq
L=R
Xiþp;jþq : ðA:6Þ h iþ h iþ
V ;B;þ= 1 V ;L;þ= V ;R;þ=
p;q Bijd ¼ SðiþsÞj þ SðisÞj
c1
h iþ h iþ
V ;L;þ= V ;L;=þ
A.3. Layer 3B monocular simple cells a Qijd þ Qijd
H =V ;L=R;þ=
h iþ h iþ
At steady-state the membrane potentials, Bij , þ Qijd
V ;R;þ=
þ Qijd
V ;R;=þ
; ðA:10Þ
of the layer 3B monocular cells are given by:
H =V ;L=R;þ=
Bij
H =V ;L=R;þ= þ
¼ 2½Sij
; ðA:7Þ where c1 and a and are constants (0.29, 6) representing
the rate of decay of the membrane potential and the
where the multiplicative factor of 2 compensates for the strength of the inhibition. Appendix B proves that the
fact that the monocular simple cells receive inputs from exact values of a and c1 are not critical. Under mild
only one eye whereas the binocular simple cells, dis- constraints on these parameters, the binocular cells act
cussed in the next section, receive input from both eyes. like the ‘‘obligate cells’’ of Poggio (1991), responding
only when their left and right inputs are approximately
A.4. Layer 3B inhibitory cells equal in magnitude. Eq. (A.10) was solved at equilib-
rium, using the theorem described in Appendix B to
The layer 3B inhibitory cells, all responding only to speed up the simulations. Fig. 19 shows that the calcu-
vertical boundaries, receive excitatory input from layer 4 lated and simulated values are essentially identical.
and inhibitory input from all other inhibitory interneu-
rons that correspond to the same position and disparity.
Their cell membrane potentials, Qijd
V ;L=R;þ=
, are deter- A.6. Layer 2/3A monocular and binocular complex cells
mined at equilibrium by the following equations:
h V1 layer 2/3A consists of both monocular and bin-
V ;L;þ= 1 h V ;L;þ= iþ V ;R;þ=
iþ
ocular complex cells. These complex cells pool the cell
Qijd ¼ SðiþsÞj b Qijd
c2 membrane potentials of monocular/binocular layer 3B
h iþ h iþ
V ;R;=þ V ;L;=þ simple cells of like orientation and both contrast polar-
þ Qijd þ Qijd ðA:8Þ
ities at each position. At steady-state their membrane
H =V ;L=R=B
potentials, Cijd , are given by:
and
h h iþ h iþ
1 h V ;R;þ= iþ iþ H =V ;L=R=B H=V ;L=R=B;þ H =V ;L=R=B;
V ;R;þ= V ;L;þ= Cijd ¼ Bijd þ Bijd : ðA:11Þ
Qijd ¼ SðisÞj b Qijd
c2
h iþ h iþ
V ;L;=þ V ;R;=þ
þ Qijd þ Qijd ; ðA:9Þ A.7. V2 layer 4
where c2 and b are constants (4.5, 4) representing the In V2, virtually all cells are binocularly driven
decay rate of the membrane potential and the strength (Hubel & Livingstone, 1987), consistent with the model
monocular inputs (represented by the solid lines) or

corresponds to the same 2D position but a different
depth (represented by the dashed line). The cell mem-
brane potentials of the vertically oriented layer 3B cells
are given by:
V
dNijd h iþ
V
¼ Nijd V
þ Jijd dV
dt
X h iþ
V
g mdd 0 Nðiþs 0 sÞjd 0
d 0 6¼d
h iþ h iþ
V V
þ mdd 0 Nðiþss 0 Þjd 0 þ l Nijd 0 ; ðA:15Þ
where d and d 0 represent disparities; s and s0 are the

corresponding allelotropic shifts, defined by Table 1; dV
is a constant (0.15) that represents the threshold of a
vertically oriented layer 4 cell, g is a constant (0.38) that
scales the total inhibition that each cell receives, l is a
Fig. 19. Output of a V1 layer 3B binocular simple cell using the pa- constant (0.1) that represents the inhibition from
rameter values of Appendix A (Eq. (A.10)). (a) Simulated and (b)
boundaries directly in front or behind, and mdd 0 repre-
calculated.
sents the inhibition from all other neurons that share an
input, as detailed in Table 2.
The disparity filter is robust in that its behavior is
hypothesis that the left and right monocular inputs are stable across a range of parameter values. The key fea-
combined in V2. The model assumes that this is done in tures of the disparity filter, as illustrated in Table 2, is
layer 4. Since the monocular inputs do not yet have a that it is symmetrical about the fixation plane (i.e., the
depth associated with them, they are added to all depth near and far disparity planes equally inhibit and are
planes along their respective lines-of-sight (cf., Section equally inhibited by the zero disparity plane) and that it
H
2.3). At steady-state their membrane potentials, Jijd , are favors the zero disparity plane in that this plane inhibits
given by: the near and far disparity planes more than they inhibit
h iþ h iþ it. Eq. (A.15) was solved using EulerÕs method and then
H H ;L H ;R
Jijd ¼ CðiþsÞj h þ CðisÞj h ; ðA:12Þ resolved with a different step size to check the accuracy
of the solution.
where h is a constant (1.42) representing the threshold of
the V1 layer 2/3A cells and s is the allelotropic shift
defined in Table 1. Similarly, at steady-state the cell A.9. V2 layer 2/3A complex cells
membrane potentials of the vertically oriented layer 4
complex cells are given by: The V2 layer 2/3A cells receive input from V2 layer
h iþ h iþ h iþ 3B. In the present model, they merely scale this input,
V V ;B V ;L V ;R
Jijd ¼ Cijd h þ b CðiþsÞj h þ CðisÞj h ; but see Grossberg (1999a, 1999b) for a discussion of
their function in perceptual grouping. Analogous to
ðA:13Þ layer 4, their steady-state cell membrane potentials,
H =V
where b is a constant (0.21) representing the strength of Tijd , are given by:
the monocular connections.
Table 2
A.8. V2 layer 3B
The inhibition coefficients mdd 0
V. Near Near Zero Far V. Far
Analogous to layer 4, at steady-state the cell mem-
H
brane potentials, Nijd , of the horizontally oriented layer V. Near – 3 5 3 2
Near 0.4 – 2.8 1.5 0.4
3B cells are given by: Zero 0.2 1.3 – 1.3 0.2
h iþ Far 0.4 1.5 2.8 – 0.4
H H
Nijd ¼ Jijd : ðA:14Þ V. Far 2 3 5 3 –
Each neuron is inhibited by every other neuron that shares either of its
V2 layer 3B contains the disparity filter (cf., Fig. 3) in inputs by an amount that depends on the disparities of the inhibited
which each vertically oriented cell is inhibited by every and inhibiting neurons (cf., Fig. 3). See text for further discussion of
other vertically oriented cell that shares either of its parameter choices.
h iþ f
H=V H =V
Tijd ¼ 50 Nijd ; ðA:16Þ Ppqijd ¼ H V H V
;
1 þ hðTði0:5Þðjþ0:5Þd þ Tði0:5Þðjþ0:5Þd þ Tðiþ0:5Þðjþ0:5Þd þ Tðiþ0:5Þðjþ0:5Þd Þ
where the multiplicative factor of 50 allows the V2 plots if p ¼ i and q ¼ j þ 1:

in Figs. 5–9 and 11–16 to be plotted to the same scale as ðA:23Þ
the V1 plots. As shown by these equations, the boundary lattice is
offset by [0.5, 0.5] relative to the lightness lattice, cor-
responding to the idea that these two processing streams
A.10. V4
are spatially displaced with respect to one another in the
cortical map. Constants f and h (1000, 10000) represent
V4 receives lightness signals from the LGN via the V1
the diffusion rate and the strength of the gating inhibi-
blobs and the V2 thin stripes, and boundary signals via
tion. Spurious lattice edge effects were avoided by using
the V1 interblob regions and the V2 pale stripes. It
the wrap-round technique according to which the last
combines the monocular lightness signals from the two
element of a row/column is adjacent to the first element
eyes that correspond to the same 3D location. Its
of the same row/column.
lightness input, Zijd , is given by:
h iþ h iþ
L R
Zijd ¼ XðiþsÞj þ XðisÞj ; ðA:17Þ
Appendix B. Proof of the obligate theorem
where i, j are positional indices, d represents the dis-
The obligate property is proved below. The proof
parity to which the model neuron is tuned and s the
shows that binocular simple cells in layer 3B can be
allelotropic shift defined in Table 1. The on-center off-
activated only if they receive approximately equal inputs
surround receptive field structure of the LGN ensures
from both left and right eye monocular simple cells in
that these lightness signals are only present at the bor-
layer 4. The constraints that determine cell firing de-
ders of surfaces, where the lightness signals represent the
pend upon the ratios of left and right monocular cell
ratio of the luminances on either side of the border. In
activity. This property explains the ratio constraint
V4, these lightness signals fill-in by a process whose
on stereoscopic fusion that is illustrated in Fig. 10.
lateral spread is gated, and thereby contained, by
The proof also shows that the activities converge expo-
boundary signals. Following Grossberg and Todorovic
nentially to unique equilibrium activities in response
(1988), the V4 cell membrane potentials, Wijd , are mod-
to any combination of constant monocular inputs.
eled by a diffusion equation that is solved at steady-
The obligate property is caused by a balance between
state:
P excitatory inputs from layer 4 monocular simple cells
Zijd þ p;q2nij Wpqd Ppqijd and inhibitory inputs from layer 3B inhibitory inter-
Wijd ¼ P : ðA:18Þ neurons. The interneurons are themselves activated by
1 þ p;q2nij Ppqijd
layer 4 monocular simple cells and mutually inhibit each
Eq. (A.18) was solved iteratively until equilibrium was other, in addition to inhibiting the binocular simple
achieved. Diffusion of potential occurs between the cells.
nearest-neighbor locations nij of ði; jÞ: The obligate property is proved below for layer 3B
nij ¼ fði; j 1Þ; ði 1; jÞ; ði þ 1; jÞ; ði; j þ 1Þg: ðA:19Þ binocular cells with dark-light contrast polarity (BVijd;B;þ ).
By symmetry, the same proof holds for cells with light-
The gating coefficients, Ppqijd , in (A.18) represent the V ;B;
dark contrast polarity (Bijd ).
H =V
inhibition of the diffusion by the boundary signals Tpqd .
They are defined by: Obligate theorem. Consider the system:
f h iþ h iþ
Ppqijd ¼ H
1 þ hðTði0:5Þðjþ0:5Þd V
þ Tði0:5Þðjþ0:5Þd H
þ Tði0:5Þðj0:5Þd V
þ Tði0:5Þðj0:5Þd Þ
; dBVijd;B;þ
¼ c1 BVijd;B;þ þ SðiþsÞj
V ;L;þ V ;R;þ
þ SðisÞj
if p ¼ i 1 and q ¼ j; dt
h iþ h iþ
ðA:20Þ
a QVijd;L;þ þ QVijd;L;
f
Ppqijd ¼ H V H V
; h iþ h iþ
1 þ hðTðiþ0:5Þðjþ0:5Þd þ Tðiþ0:5Þðjþ0:5Þd þ Tðiþ0:5Þðj0:5Þd þ Tðiþ0:5Þðj0:5Þd Þ
þ QVijd;R;þ þ QVijd;R; ; ðB:1Þ
if p ¼ i þ 1 and q ¼ j;
ðA:21Þ
h iþ h iþ
f dQVijd;L;þ
Ppqijd ¼ ; ¼ c2 QVijd;L;þ þ SðiþsÞj
V ;L;þ
b QVijd;R;þ
H
1 þ hðTði0:5Þðj0:5Þd V
þ Tði0:5Þðj0:5Þd H
þ Tðiþ0:5Þðj0:5Þd V
þ Tðiþ0:5Þðj0:5Þd Þ dt
h iþ h iþ
if p ¼ i and q ¼ j 1; þ QVijd;R; þ QVijd;L; ; ðB:2Þ
ðA:22Þ
h iþ h iþ Case 1
dQVijd;R;þ
¼ c2 QVijd;R;þ þ SðisÞj
V ;R;þ
b QVijd;L;þ V ;L;þ
0 < SðiþsÞj V ;R;þ
; SðisÞj ; V ;L;
SðiþsÞj V ;R;
; SðisÞj < 0;
dt
h iþ h iþ V ;L;þ V ;R;þ
b SðiþsÞj b SðisÞj
þ QVijd;L; þ QVijd;R; ; ðB:3Þ 6 V ;R;þ ; and 6 V ;L;þ : ðB:11Þ
c2 SðisÞj c2 SðiþsÞj
h iþ h iþ
dQVijd;L; Under these conditions, (B.4) and (B.5) imply that, for
¼ c2 Qijd þ SðiþsÞj b QVijd;R;
V ;L; V ;L;
sufficiently large t,
dt
h iþ h iþ
þ Qijd V ;R;þ
þ Qijd V ;L;þ
; ðB:4Þ QVijd;L; ; QVijd;R; 6 0: ðB:12Þ
V ;L;þ V ;R;þ
By (B.12), and recalling for this case 0 < SðiþsÞj ; SðisÞj ,
and (B.2) and (B.3) can be approximated at large times:
h iþ h iþ
dQVijd;R; dQVijd;L;þ h iþ
¼ c2 QVijd;R; þ SðisÞj
V ;R;
b QVijd;L; ¼ c2 QVijd;L;þ þ SðiþsÞj
V ;L;þ
b QVijd;R;þ ðB:13Þ
dt dt
h iþ h iþ
þ QVijd;L;þ þ QVijd;R;þ ; ðB:5Þ and
dQVijd;R;þ h iþ
V ;L;þ
where SðiþsÞj V ;R;þ
and SðisÞj are monocular simple cell activities ¼ c2 QVijd;R;þ þ SðisÞj
V ;R;þ
b QVijd;L;þ : ðB:14Þ
dt
that are defined by (A.4) and (A.6), 0 < c1 and
Eqs. (B.13) and (B.14) are used to draw the phase-plane
0 < b < c2 < a < c2 þ b: ðB:6Þ plot shown in Fig. 20a. Eq. (B.11) implies:
Under these conditions, the system converges expo- V ;R;þ
SðisÞj V ;L;þ
SðiþsÞj
nentially to the unique equilibria specified by (1)–(4) 0< 6 ðB:15Þ
provided that the inputs are constant. c2 b
and
V ;L;þ V ;R;þ V ;L; V ;R;
(1) if 0 < SðiþsÞj ; SðisÞj ; SðiþsÞj ; SðisÞj < 0;
V ;L;þ V ;R;þ
V ;L;þ V ;R;þ
b=c2 6 SðiþsÞj =SðisÞj ; and b=c2 6 SðisÞj V ;R;þ V ;L;þ
=SðiþsÞj , SðiþsÞj SðisÞj
0< 6 : ðB:16Þ
c2 b
then at equilibrium

1 a From these equations and where the nullclines intersect
BVijd;B;þ ¼ 1 V ;L;þ
SðiþsÞj V ;R;þ
þ SðisÞj ; ðB:7Þ
c1 c2 þ b the axes in Fig. 20a, it follows that the nullclines must
V ;L;þ
(2) if 0 < SðiþsÞj V ;R;þ
; SðisÞj V ;L;
; SðiþsÞj V ;R;
; SðisÞj < 0; and cross each other at a point where
V ;R;þ V ;L;þ
SðisÞj =SðiþsÞj < b=c2 , 0 6 QVijd;L;þ ; QVijd;R;þ : ðB:17Þ
then at equilibrium This allows us to remove the rectification in (B.13) and
(B.14) which in turn allows us to perform local analysis
1 a
BVijd;B;þ ¼ V ;R;þ
SðisÞj þ 1 V ;L;þ
SðiþsÞj ; ðB:8Þ on the linear system
c1 c2
V ;L;þ V ;R;þ V ;L; V ;R; V ;L;þ

(3) if 0 < SðiþsÞj ; SðisÞj ; SðiþsÞj ; SðisÞj < 0; and SðiþsÞj = c2 b
J¼ : ðB:18Þ
V ;R;þ
SðisÞj < b=c2 , b c2
then at equilibrium The eigenvalues are c2 b. By (B.6), b < c2 , and so

both eigenvalues are negative. Thus the crossing of the
V ;B;þ 1 V ;L;þ a V ;R;þ
Bijd ¼ S þ 1 S ; ðB:9Þ nullclines represents a unique equilibrium point to which
c1 ðiþsÞj c2 ðisÞj the system exponentially converges. This equilibrium
(4) for all other values of SðiþsÞj , SðisÞj , SðiþsÞj , SðisÞj , point can be found by adding (B.13) and (B.14) and
recalling that for this case QVijd;L;þ and QVijd;R;þ are non-
at equilibrium BVijd;B;þ 6 0: ðB:10Þ
negative:
Fig. 18a and b shows the simulated and calculated outputs d V ;L;þ
Qijd þ QVijd;R;þ ¼ ðc2 þ bÞ QVijd;L;þ þ QVijd;R;þ
for the above system for the parameter values summarized dt
in Appendix A. V ;L;þ
þ SðiþsÞj V ;R;þ
þ SðisÞj : ðB:19Þ
Proof. First note by (A.6) that, out of the four possible The equilibrium point is:
inputs SðiþsÞj , SðisÞj , SðiþsÞj , and SðisÞj , at most only two
can be positive. This greatly simplifies the subsequent 1 V ;L;þ
QVijd;L;þ þ QVijd;R;þ ¼ V ;R;þ
SðiþsÞj þ SðisÞj : ðB:20Þ
analysis. c2 þ b
and
QVijd;L;þ > 0: ðB:25Þ
Eqs. (B.24) and (B.25) imply that (B.13) and (B.14) can
be rewritten as:
dQVijd;L;þ
¼ c2 QVijd;L;þ þ SðiþsÞj
V ;L;þ
ðB:26Þ
dt
and
dQVijd;R;þ
¼ c2 QVijd;R;þ þ SðisÞj
V ;R;þ
bQVijd;L;þ : ðB:27Þ
dt
Linear analysis of

c2 0
J¼ ðB:28Þ
b c2
yields only a single eigenvalue, c2 . Because this eigen-

value is negative, the intersection of nullclines represents
an equilibrium point to which the system converges.
Solving (B.26) at equilibrium yields:
V ;L;þ
SðiþsÞj
QVijd;L;þ ¼ : ðB:29Þ
c2
Fig. 20. Phase-plots used in Appendix B: (a) Case 1; (b) Case 2. In By (B.12), (B.24) and (B.29), and recalling that for this
both cases the dQVijd;R;þ =dt ¼ 0 nullcline crosses the ordinate and the V ;L;þ V ;R;þ
V ;R;þ V ;R;þ V ;L;þ
case 0 < SðiþsÞj ; SðisÞj , it follows that at large times, (B.1)
abscissa at SðisÞj =c2 and SðisÞj =b, while the dQijd =dt ¼ 0 nullcline
is approximated by:
V ;L;þ V ;L;þ
crosses them at SðiþsÞj =b and SðiþsÞj =c2 .

dBVijd;B;þ a
¼ c1 BVijd;B;þ þ 1 S V ;L;þ þ SðisÞj
V ;R;þ
; ðB:30Þ
dt c2 ðiþsÞj
By (B.12) and (B.20), and recalling for this case 0 <
V ;L;þ V ;R;þ
SðiþsÞj ; SðisÞj , we see that at large times (B.1) is approx- which has the equilibrium solution:
imated by:

1 a
dBVijd;B;þ a BVijd;B;þ ¼ 1 V ;L;þ
SðiþsÞj V ;R;þ
þ SðisÞj : ðB:31Þ
¼ c1 BVijd;B;þ þ 1 V ;L;þ
SðiþsÞj V ;R;þ
þ SðisÞj : c1 c2
dt c2 þ b
ðB:21Þ
Case 3
Eq. (B.21) then exponentially converges to: V ;L;þ V ;R;þ V ;L; V ;R;
0 < SðiþsÞj ; SðisÞj ; 0 < SðiþsÞj ; SðisÞj ;
1 a
SðiþsÞj V ;R;þ
þ SðisÞj : ðB:22Þ V ;L;þ
SðiþsÞj
c1 c2 þ b b
and V ;R;þ < : ðB:32Þ
SðisÞj c2
Case 2
V ;R;þ
By symmetry with Case 2:
SðisÞj b
0< V ;L;þ
SðiþsÞj V ;R;þ
; SðisÞj ; V ;L;
SðiþsÞj V ;R;
; SðisÞj < 0; and < :
V ;L;þ
SðiþsÞj c2 1 a
BVijd;B;þ ¼ V ;L;þ
SðiþsÞj þ 1 V ;R;þ
SðisÞj : ðB:33Þ
c1 c2
ðB:23Þ
First note that (B.12)–(B.14) apply to this case, allowing
Case 4
us to draw the phase-plane plot shown in Fig. 20b. From
where the nullclines intersect the axes and from (B.23), SðiþsÞj ; SðisÞj <0 and SðiþsÞj ; SðisÞj < 0: ðB:34Þ
we see that the nullclines must cross each other at a
point where From (B.1) we see that at equilibrium:
QVijd;R;þ < 0 ðB:24Þ BVijd;B;þ 6 0: ðB:35Þ

Case 5 and
V ;L;þ
V ;L;þ
0 < SðiþsÞj V ;R;
; SðisÞj ; V ;L;
SðiþsÞj V ;R;þ
; SðisÞj < 0; SðiþsÞj
QVijd;L;þ ¼ : ðB:47Þ
V ;L;þ V ;R; c2
b SðiþsÞj b SðisÞj
6 V ;R; ; and 6 V ;L;þ : ðB:36Þ Using these equations, and recalling that for this case
c2 SðisÞj c2 SðiþsÞj V ;L;þ V ;R;þ
0 < SðiþsÞj and SðisÞj < 0, we see that at large times (B.1)
By analogy with Case 1, in particular (B.12) and (B.20), is approximated by:
at equilibrium:
dBVijd;B;þ a V ;L;þ
V ;L;þ
SðiþsÞj : ðB:48Þ
QVijd;L; ; QVijd;R;þ 6 0 ðB:37Þ dt c2
and This converges to:

1 V ;L;þ V ;B;þ
Bijd ¼
1
1
a
S V ;L;þ : ðB:49Þ
QVijd;L;þ þ QVijd;R; ¼ V ;R;
SðiþsÞj þ SðisÞj : ðB:38Þ c1 c2 ðiþsÞj
c2 þ b
From (B.6) we see c2 < a. Thus (B.49) implies
Using (B.37) and (B.38), and recalling that for this case
V ;L;þ
and SðisÞj < 0, we see that at large times (B.1) BVijd;B;þ < 0: ðB:50Þ
is approximated by:
dBVijd;B;þ a V ;L;þ Case 7
V ;L;þ
V ;R;
SðiþsÞj þ SðisÞj :
dt c2 þ b V ;L;þ
SðiþsÞj b
V ;L;þ V ;R; V ;L; V ;R;þ
ðB:39Þ 0 < SðiþsÞj ; SðisÞj ; SðiþsÞj ; SðisÞj < 0; and V ;R; < :
SðisÞj c2
Its equilibrium solution is: ðB:51Þ

1 a a By analogy with Case 2, in particular (B.12), (B.24) and
SðiþsÞj V ;R;
SðisÞj :
c1 c2 þ b c2 þ b (B.29), at equilibrium:
ðB:40Þ
QVijd;L; ; QVijd;R;þ 6 0; ðB:52Þ
V ;L;þ V ;R;
Recalling that for this case ðb=c2 ÞSðiþsÞj
we 6 SðisÞj ,
convert (B.40) into an inequality: QVijd;L;þ < 0; ðB:53Þ

1 a a b V ;L;þ and
BVijd;B;þ 6 1 V ;L;þ
SðiþsÞj SðiþsÞj :
c1 c2 þ b c2 þ b c2 V ;R;
SðisÞj
ðB:41Þ QVijd;R; ¼ : ðB:54Þ
c2
From (B.6) we see that c2 < a. This allows (B.41) to be Using these equations, and recalling that for this case
rewritten V ;L;þ
and SðisÞj < 0, we see that at large times (B.1)
is approximated by:
V ;B;þ 1 c2 V ;L;þ c2 b V ;L;þ
Bijd < 1 S S :
c1 c2 þ b ðiþsÞj c2 þ b c2 ðiþsÞj dBVijd;B;þ a V ;R;
ðB:42Þ ¼ c1 BVijd;B;þ þ SðiþsÞj
V ;L;þ
SðisÞj : ðB:55Þ
dt c2
After factorization and cancellation we find: This converges to:

BVijd;B;þ < 0: ðB:43Þ V ;B;þ 1 V ;L;þ a V ;R;
Bijd ¼ S S : ðB:56Þ
c1 ðiþsÞj c2 ðiþsÞj
V ;L;þ V ;R;
Case 6 Recalling that for this case ðc2 =bÞSðiþsÞj < SðisÞj , we
V ;R;
convert (B.56) into an inequality:
SðisÞj b
0 < SðiþsÞj ; SðisÞj ; SðiþsÞj ; SðisÞj < 0; and < : 1 a V ;L;þ
V ;L;þ
SðiþsÞj c2 BVijd;B;þ < 1 S : ðB:57Þ
c1 b ðiþsÞj
ðB:44Þ From (B.6) we see that b < a. Thus (B.57) implies:
By analogy with Cases 1 and 2, in particular (B.12), BVijd;B;þ < 0: ðB:58Þ
(B.24) and (B.29), at equilibrium:
QVijd;L; ; QVijd;R;þ 6 0; ðB:45Þ Case 8

QVijd;R; < 0; ðB:46Þ SðiþsÞj ; SðisÞj < 0 and 0 < SðiþsÞj ; SðisÞj : ðB:59Þ
By analogy with Cases 5–7: Grossberg, S. (1999a). How does the cerebral cortex work? Learning,
attention and grouping by the laminar circuits of visual cortex.
BVijd;B;þ < 0: ðB:60Þ Spatial Vision, 12, 163–186.
Grossberg, S. (1999b). The link between brain learning, attention, and
Case 9 consciousness. Consciousness and Cognition, 8, 1–44.
Grossberg, S., Hwang, S., & Mingolla, E. (2002). Thalamocortical
V ;L;þ
SðiþsÞj V ;R;þ
; SðisÞj V ;L;
; SðiþsÞj V ;R;
; SðisÞj ¼ 0: ðB:61Þ dynamics of the McCollough effect: Boundary-surface alignment
through perceptual learning. Vision Research, 39, 3796–3816.
By inspection, (B.2)–(B.5) imply that at equilibrium, Grossberg, S., & Kelly, F. (1999). Neural dynamics of binocular
brightness perception. Vision Research, 39, 3796–3816.
QVijd;L;þ ; QVijd;R;þ ; QVijd;L; ; QVijd;R; ¼ 0: ðB:62Þ Grossberg, S., & McLoughlin, N. (1997). Cortical dynamics of three-
dimensional surface perception: Binocular and half-occluded scenic
images. Neural Networks, 10, 1583–1605.
By (B.62) and (B.1) implies that at equilibrium,
Grossberg, S., Mingolla, E., & Ross, W. D. (1997). Visual brain and
visual perception: How does the cortex do perceptual grouping?
BVijd;B;þ ¼ 0: ðB:63Þ
Trends in Neuroscience, 20, 106–111.
Grossberg, S., Mingolla, E., & Williamson, J. (1995). Synthetic
As we have now considered all possible cases, the the- aperture radar processing by a multiple scale neural system for
orem is proved. boundary and surface representation. Neural Networks, 8, 1005–
1028.
Grossberg, S., & Pessoa, L. (1998). Texture segregation, surface
References representation, and figure-ground separation. Vision Research, 38,
2657–2684.
Anzai, A., Bearse, M. A., Freeman, R. D., & Cai, D. (1995). Contrast Grossberg, S., & Raizada, R. D. (2000). Contrast-sensitive perceptual
coding by cells in the catÕs striate cortex: Monocular vs. binocular grouping and object-based attention in the laminar circuits of
detection. Visual Neuroscience, 12, 77–93. primary visual cortex. Vision Research, 40, 1413–1432.
Backus, B. T., Fleet, D. J., Parker, A. J., & Heeger, D. J. (2001). Grossberg, S., & Swaminathan, G. A laminar cortical model for
Human cortical activity correlates with stereoscopic depth percep- 3D perception of slanted and curved surfaces and of 2D
tion. Journal of Neurophysiology, 86, 2054–2068. images: Development, attention and bistability. Submitted for
Bakin, J. S., Nakayama, K., & Gilbert, C. D. (2000). Visual responses publication.
in monkey area V1 and V2 to three-dimensional surface configu- Grossberg, S., & Todorovic, D. (1988). Neural dynamics of 1-D and
rations. The Journal of Neuroscience, 20, 8188–8198. 2-D brightness perception: A unified model of classical and recent
Callaway, E. M. (1998). Local circuits in primary visual cortex of the phenomena. Perception and Psychophysics, 43, 241–277.
macaque monkey. Annual Review of Neuroscience, 21, 47–74. Grossberg, S., & Williamson, J. R. (2001). A neural model of how
Callaway, E. M., & Wiser, A. K. (1996). Contributions of individual horizontal and interlaminar connections of visual cortex develop
layer 2–5 spiny neurons to local circuits in macaque primary visual into adult circuits that carry out perceptual groupings and learning.
cortex. Visual Neuroscience, 13, 907–922. Cerebral Cortex, 11, 37–58.
Cumming, B. G. (2002). Receptive field structure and disparity tuning Howard, I. P., & Rogers, B. J. (1995). Binocular vision and stereopsis.
in primate V1. Vision ScienceS Society (abstracts), 288. New York: Oxford University Press.
Cumming, B. G., & Parker, A. J. (2000). Local disparity not perceived Howe, P. D. L., & Grossberg, S. (2001). Laminar cortical circuits for
depth is signaled by binocular neurons in cortical area V1 of the stereopsis and surface depth perception. Society for Neuroscience
macaque. The Journal of Neuroscience, 20, 4758–4767. Abstracts, 164.17.
Dow, B. M. (1974). Function classes of cells and their laminar Howe, P. D. L., & Watanabe, T. (in press). A new test of the disparity
distribution in monkey visual cortex. Journal of Neurophysiology, energy model of stereopsis. Vision Research.
37, 927–946. Hubel, D. H., & Livingstone, M. S. (1987). Segregation of form, color,
Fleet, D. J., Wagner, H., & Heeger, D. J. (1996). Neural encoding of and stereopsis in primate area 18. The Journal of Neuroscience, 7,
binocular disparity: Energy models, position shifts and phase 3378–3415.
shifts. Vision Research, 36, 1839–1857. Hubel, D. H., & Wiesel, T. N. (1968). Receptive fields and functional
Frisby, J. P. (2001). Limited understanding of PanumÕs limiting case. architecture of monkey striate cortex. Journal of Physiology, 195,
Perception, 30, 1151–1152. 215–243.
Gillam, B., Blackburn, S., & Cook, M. (1995). PanumÕs limiting case: Julesz, B. (1971). Foundations of cyclopean perception. Chicago: The
Double fusion, convergence error, or Ôda Vinci stereopsisÕ. Percep- University of Chicago Press.
tion, 24, 333–346. Kandel, E. R., Schwartz, J. H., & Jessell, T. M. (2000). Principles of
Gillam, B., Blackburn, S., & Nakayama, K. (1999). Stereopsis based neural science (4th ed.). Chicago: University of Chicago Press.
on monocular gaps: Metrical encoding of depth and slant without Katz, L. C., Gilbert, C. D., & Wiesel, T. N. (1989). Local circuits and
matching contours. Vision Research, 39, 493–502. ocular dominance columns in monkey striate cortex. The Journal of
Grimson, W. E. (1981). A computer implementation of a theory of Neuroscience, 9, 1389–1399.
human stereo vision. Philosophical Transactions of the Royal Kelly, F. J., & Grossberg, S. (2000). Neural dynamics of 3-D surface
Society (B), 292, 217–253. perception: Figure-ground separation and lightness perception.
Grossberg, S. (1987). Cortical dynamics of three-dimensional form, Perception and Psychophysics, 62, 1596–1619.
color, and brightness perception: II. Binocular theory. Perception Krol, J. D., & van de Grind, W. A. (1983). Depth from dichoptic edges
and Psychophysics, 41, 117–158. depends on vergence tuning. Perception, 12, 425–438.
Grossberg, S. (1994). 3D vision and figure-ground separation by visual Livingstone, M. S., & Hubel, S. H. (1984). Anatomy and physiology of
cortex. Perception and Psychophysics, 55, 48–120. a color system in the primate visual cortex. The Journal of
Grossberg, S. (1997). Cortical dynamics of three-dimensional figure- Neuroscience, 4, 309–356.
ground perception of two-dimensional figures. Psychological Re- Marr, D., & Poggio, T. (1976). Cooperative computation of stereo
view, 104, 618–658. disparity. Science, 194, 283–287.
McKee, S. P., Bravo, M. J., Smallman, H. S., & Legge, G. E. (1995). grouping, attention, and orientation contrast. Visual Cognition, 8,
The Ôuniqueness constraintÕ and binocular masking. Perception, 24, 431–466.
49–65. Read, J. C. A., Cumming, B. C., & Parker, A. J. (2002). Simple cells
McKee, S. P., Bravo, M. J., Taylor, D. G., & Legge, G. E. (1994). can show non-linear binocular combination. Vision ScienceS
Stereo matching precedes dichoptic masking. Vision Research, 34, Society (abstract), 287.
1047–1060. Rockland, K. S., & Virga, A. (1990). Organization of individual
McLoughlin, N., & Grossberg, S. (1998). Cortical computation of cortical axons projecting from area V1 (area 17) to V2 (area 18) in
stereo disparity. Vision Research, 38, 91–99. the macaque monkey. Visual Neuroscience, 4, 11–28.
Merigan, W. H., & Maunsell, J. H. R. (1993). How parallel are the Roe, A. W., & TsÕo, D. Y. (1995). Visual topography in primate V2:
primate visual pathways. Annual Review of Neuroscience, 16, 369– Multiple representation across functional stripes. The Journal of
402. Neuroscience, 15, 3689–3715.
Nakamura, K., & Colby, C. L. (2000a). Visual, saccade-related, and Roe, A. W., & TsÕo, D. Y. (1997). The functional architecture of area
cognitive activation of single neurons in monkey extrastriate area V2 in the macaque monkey. Cerebral Cortex, 12, 295–333.
V3A. Journal of Neurophysiology, 84, 677–692. Schiller, P. H., Finlay, B. L., & Volman, S. F. (1976). Quantitative
Nakamura, K., & Colby, C. L. (2000b). Updating of the visual studies of single-cell properties in monkey striate cortex. I.
representation in monkey striate and extrastriate cortex during Spatiotemporal organization of receptive fields. Journal of Neuro-
saccades. Proceedings of the National Academy of Sciences, 99, physiology, 39, 1288–1319.
4026–4031. Schiller, P. H., Logothetis, N. K., & Charles, E. R. (1990a). Role of the
Nakamura, H., Kuroda, T., Wakita, M., Kusunoki, M., Kato, color-opponent and broad-band channels in vision. Visual Neuro-
A., Mikami, A., Sakata, H., & Itoh, K. (2001). From three- science, 5, 321–346.
dimensional space vision to prehensile hand movements: The Schiller, P. H., Logothetis, N. K., & Charles, E. R. (1990b). Functions
lateral intraparietal area links the area V3A and the anterior of the colour-opponent and broad-band channels of the visual
intraparietal area in macaque. The Journal of Neuroscience, 21, system. Nature, 343, 68–70.
8174–8187. Schor, C., & Heckmann, T. (1989). Interocular differences in contrast
Nakayama, K., & Shimojo, S. (1990). da Vinci stereopsis: Depth and and spatial frequency: Effects on stereopsis and fusion. Vision
subjective occluding contours from unpaired image points. Vision Research, 29, 837–847.
Research, 30, 1811–1825. Smallman, H. S., & McKee, S. P. (1995). A contrast ratio constraint on
Ohzawa, I. (1998). Mechanisms of stereoscopic vision: The disparity stereo matching. Proceedings of the Royal Society of London B, 260,
energy model. Current Opinion in Neurobiology, 8, 509–515. 265–271.
Ohzawa, I., DeAngelis, G. C., & Freeman, R. D. (1990). Stereoscopic Smith, E. L., Chino, Y., Ni, J., & Cheng, H. (1997). Binocular
depth discrimination in the visual cortex: Neurons ideally suited as combination of contrast signals by striate cortical neurons in the
disparity detectors. Science, 249, 1037–1041. monkey. Journal of Neurophysiology, 78, 366–382.
Panum, P. L. (1858). Physiologische Untersuchungen ueber das Sehen Swaminathan, G., & Grossberg, S. (2001). Laminar cortical circuits for
mit zwei Augen. Kiel: Schwerssche Buchhandlung (translated by C. the perception of slanted and curved 3D surfaces. Society for
Hubscher, Hanover, NH: Dartmouth Eye Institute, 1940). Neuroscience Abstracts, 619.49.
Pessoa, L., Mingolla, E., & Neumann, H. (1995). A contrast- and Tootell, R. B. H., Mendola, J. D., Hadjikhani, N. K., Ledden,
luminance-driven multiscale network model of brightness percep- P. J., Liu, A. K., Reppas, J. B., Sereno, M. I., & Dale,
tion. Vision Research, 35, 2201–2223. A. M. (1997). Functional analysis of V3A and related areas in
Peterhans, E. (1997). Functional organization of area V2 in the awake human visual cortex. The Journal of Neuroscience, 17, 7060–
monkey. Cerebral Cortex, 12, 335–357. 7078.
Poggio, G. F. (1972). Spatial properties of neurons in striate cortex of Tsao, D. Y., & Livingstone, M. S. (in press). Spatiotemporal maps of
unanesthetized macaque monkey. Investigative Ophthalmology, 11, disparity-selective simple cells in macaque V1. Neuron.
369–377. von der Heydt, R., Zhou, H., & Friedman, H. S. (2000). Represen-
Poggio, G. F. (1991). Physiological basis of stereoscopic vision. In tation of stereoscopic edges in monkey visual cortex. Vision
Vision and visual dysfunction. Binocular vision (pp. 224–238). Research, 40, 1955–1967.
Boston, MA: CRC. Wallach, H. (1976). On perception. New York: Quadrangle/The New
Poggio, G. F., & Fischer, B. (1977). Binocular interaction and depth York Times Book Co., p. 8.
sensitivity in striate and prestriate cortex of behaving rhesus Wang, Z., Wu, X., Ni, R., & Wang, Y. (2001). Double fusion does not
monkey. Journal of Neurophysiology, 40, 1392–1405. occur in PanumÕs limiting case: Evidence from orientation dispar-
Poggio, G. F., & Talbot, W. H. (1981). Mechanisms of static and ity. Perception, 30, 1143–1149.
dynamic stereopsis in foveal cortex of the rhesus monkey. Journal Xiao, Y., Zych, A., & Felleman, D. J. (1999). Segregation and
of Physiology, 315, 469–492. convergence of functionally defined V2 thin stripe and interstripe
Raizada, R., & Grossberg, S. (2001). Context-sensitive bindings by the compartment projections to area V4 of macaques. Cerebral Cortex,
laminar circuits of V1 and V2: A unified model of perceptual 9, 792–804.

Laminart PDF

Uploaded by

Copyright:

Available Formats

Laminart PDF

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Laminart PDF

Uploaded by

Copyright:

Available Formats

Vision Research 43 (2003) 801–829

A laminar cortical model of stereopsis and three-dimensional

1. Introduction plain the 3D surface percepts that form an integral part

Fig. 1. Model circuit diagram.

depth perception will occur only if the two edges belong

Once monocular information is included, the prob-

2.4. Surfaces 3. Model simulations

3.1.2. Release from dichoptic masking

position as the low contrast bar of the right input. The

3.1.3. Return to dichoptic masking

inhibition, but does not enforce it. One advantage of this

3.2.2. Contrast variations

vertical boundaries of the leftmost bar representation of

3.2.3. The ratio rule

Fig. 10. Simulation of the ratio constraint on stereoscopic fusion. On

white element in one stereo half image there is a corre-

3.5. Da Vinci stereopsis

in the V2 disparity ﬁlter along their respective monoc-

3.6. Lightness illusions

Raizada & Grossberg, 2001). This LAMINART model

The LGN cells obey membrane, or shunting, equa-

where L=R designates that the cell belongs to the left or

  " !# of the inhibition, d is the disparity to which the model

monocular inputs (represented by the solid lines) or

where d and d 0 represent disparities; s and s0 are the

where the multiplicative factor of 50 allows the V2 plots if p ¼ i and q ¼ j þ 1:

then at equilibrium The eigenvalues are c2  b. By (B.6), b < c2 , and so

yields only a single eigenvalue, c2 . Because this eigen-

QVijd;R;þ < 0 ðB:24Þ BVijd;B;þ 6 0: ðB:35Þ

and This converges to:

QVijd;L; ; QVijd;R;þ 6 0; ðB:45Þ Case 8

You might also like

" !# of the inhibition, d is the disparity to which the model

then at equilibrium The eigenvalues are c2 b. By (B.6), b < c2 , and so