Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 256

jects. This happens by stopping settling whenever the
target output (object 2) gets above an activity of .6 (if
this doesn't happen, settling stops after 200 cycles).
The spatial processing pathway has a sequence of two
layers of spatial representations, differing in the level of
spatial resolution. As in the object pathway, each unit
in the spatial pathway represents 3 adjacent spatial lo-
cations, but unlike the object pathway, these units are
not sensitive to particular features. Two units per lo-
cation provide distributed representations in both layers
of the spatial pathway. This redundancy will be useful
for demonstrating the effects of partial damage to this
pathway.

Spat2 Output Obj2

Spat1

Obj1

V1

Object 2 (Target)

Object 1 (Cue)

Input

Figure 8.25: The simple spatial attention model.

Locate the attn_ctrl overall control panel.
This control panel contains a number of important pa-
rameters, most of which are wt_scale values that de-
termine the relative strength of various pathways within
the network (and all other pathways have a default
strength of 1). As discussed in chapter 2, the connec-
tion strengths can be uniformly scaled by a normalized
multiplicative factor, called wt_scale in the simula-
tor. We set these weight scale parameters to determine
the relative influence of one pathway on the other, and
the balance of bottom-up (stimulus driven) versus top-
down (attentional) processing.
As emphasized in figure 8.23, the spatial pathway in-
fluences the object pathway relatively strongly (as de-
termined by the spat_obj parameter with a value of
2), whereas the object pathway influences the spatial
pathway with a strength of .5. Also, the spatial sys-
tem will be responsive to bottom-up inputs from V1
because the v1_spat parameter, with a value of 2,
makes the V1 to Spat1 connections relatively strong.
This strength allows for effective shifting of attention
(as also emphasized in figure 8.23). Finally, the other
two parameters in the control panel show that we are
using relatively slow settling, and adding some noise
into the processing to simulate subject performance.

8.5.2 Exploring the Simple Attentional Model

Open the project attn_simple.proj.gz in
chapter_8 to begin.
Let's step through the network structure and con-
nectivity, which was completely pre-specified (i.e., the
network was not trained, and no learning takes place,
because it was easier to hand-construct this simple ar-
chitecture). As you can see, the network basically
resembles figure 8.23, with mutually interconnected
Spatial and Object pathways feeding off of a V1-
like layer that contains a spatially mapped feature array
(figure 8.25). In this simple case, we're assuming that
each “object” is represented by a single distinct feature
in this array, and also that space is organized along a sin-
gle dimension. Thus, the first row of units represents the
first object's feature (which serves as the cue stimulus)
in each of 7 locations, and the second row represents
the second object's feature (which serves as the target)
in these same 7 locations.

, !

Now select r.wt in the network window and click on
the object and spatial units to see how they function via
their connectivity patterns.
The object processing pathway has a sequence of
3 increasingly spatially invariant layers of representa-
tions, with each unit collapsing over 3 adjacent spatial
locations of the object-defining feature in the layer be-
low. Note that the highest, fully spatially invariant level
of the object pathway plays the role of the output layer,
and is used for measuring the reaction time to detect ob-

, !

Perceiving Multiple Objects

Although much of the detailed behavioral data we will
explore with the model concerns the Posner spatial cue-
ing task, we think the more basic functional motiva-
tion for visual attention is to facilitate object recog-