Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 235

.projects/Remove/Project_0 . Or, if you wish to
stop now, quit by selecting Object/Quit .

Question 8.3 Explain the sense in which these probe
stimuli represent edges, making reference to the rela-
tionship between the Input_pos and Input_neg
patterns.

8.3.3 Summary and Discussion

This model showed how Hebbian learning can develop
representations that capture important statistical corre-
lations present in natural images. These correlations
reflect the reliable presence of edges in these images,
with the edges varying in size, position, orientation, and
polarity. Because the resulting representations capture
many of the important properties of actual V1 receptive
fields, this model can provide a computational explana-
tion of why these properties arise in the brain. This is
an important advantage of computational models — ex-
perimentalists can record and document the properties
of visual representations, and a computational model
can show how these properties result from the interac-
tion between general learning and architectural princi-
ples and the structure of the natural environment.
Another advantage of understanding the principled
basis for these visual representations is that we can po-
tentially understand the commonalities between these
representations and those in other sensory modalities.
For example, the phenomenon of topography is widely
observed throughout all primary sensory representa-
tions (e.g., somatosensory representations are arranged
like a miniature human or homunculus in the brain).
We can expect that other sensory modalities will tend
to represent the strong correlational features present in
the input.
One interesting indication that general principles are
at work in shaping perceptual representations comes
from an experiment by Sur, Garraghty, and Roe (1988).
They rerouted visual projections from the thalamus into
the auditory cortex of ferrets and found that neurons in
the auditory cortex developed several response proper-
ties typical of visual neurons. Although this and other
similar experiments are far from conclusive, they do
provide suggestive evidence that the different areas of
cortex use the same basic kinds of learning principles.
Finally, we should note that the early visual system
has been the subject of a large number of computa-
tional models, exploiting a range of different principles
(for recent reviews, see Swindale, 1996; Erwin et al.,

Now, let's present these patterns to the network.

Select act in the network window, and then
do StepProbe in the control panel. Continue to
StepProbe through the next 3 events, and note the
relationship in each case to the weight-based receptive
fields shown in the grid log.
You should observe that the units that coded for the
orientation and directionality of the probe were acti-
vated.
If you are interested, you can draw new patterns into
the probe events, and present them by the same proce-
dure just described. In particular, it is interesting to see
how the network responds to multiple edges present in
a single input event.
Finally, to see that the lateral connectivity is responsi-
ble for developing topographic representations, you can
load a set of receptive fields generated from a network
trained with lat_wt_scale set to .01.

Do

View , RFIELDS ,

and select

v1rf_lat01.rfs.log .
You should see little evidence of a topographic orga-
nization in the resulting receptive field grid log, indi-
cating that this strength of lateral connectivity provided
insufficient neighborhood constraints. Indeed, we have
found that these patterns look similar to those with net-
works trained without any lateral connectivity at all.
There appears to be an interaction between the topo-
graphic aspect of the representations and the nature of
the individual receptive fields themselves, which look
somewhat different in this weaker lateral connectivity
case compared to the original network. These kinds of
interactions have been documented in the brain (e.g.,
Das & Gilbert, 1995; Weliky, Kandler, & Katz, 1995),
and make sense computationally given that the lateral
connectivity has an important effect on the response
properties of the neurons, which is responsible for tun-
ing up their receptive fields in the first place.

Go to the PDP++Root window. To continue on to
the next simulation, close this project first by selecting