Information Technology Reference
In-Depth Information
changing the units' excitability (via the leak current) af-
fect the success of the network, and why?
.projects/Remove/Project_0
. Or, if you wish to
stop now, quit by selecting
Object/Quit
.
Letter Inputs
3.3.2
Localist versus Distributed Representations
The digit network we explored in the previous section
used very simple detectors with weight patterns that di-
rectly matched the input patterns they were intended to
detect, often referred to as
template matching
repre-
sentations. Another aspect of these units was that only
one unit represented each input pattern, referred to as a
local
or
localist
representation, which often goes hand
in hand with template matching ideas. Such representa-
tions are also often referred to as
grandmother cell
rep-
resentations, because their use must imply that some-
where in the brain is a neuron that (uniquely) represents
one's grandmother.
Though useful for demonstration purposes, localist
representations are not particularly realistic or power-
ful. Thus, it is unfortunate that the detector model of the
neuron is often associated with this type of representa-
tion — as we made clear in the previous chapter, the
detector
function
can apply to very complex and diffi-
cult to describe nonlocalist
contents
or representations.
Recordings of cortical neurons have consistently
shown that the brain uses
distributed
representations,
where each neuron responds to a variety of different
input stimuli, and that many neurons are active for
each input stimulus. For example, researchers have
presented visual stimuli that are systematically varied
along a number of different dimensions (size, shape,
color, etc.), and recorded the activities of neurons in
the areas of the cortex that process visual inputs (see
Desimone & Ungerleider, 1989; Tanaka, 1996, for re-
views). In all cases that we know, these studies have
shown that cortical neurons exhibit measurable
tuning
curves
, which means that they respond in a graded fash-
ion to stimuli within a range of different parameter val-
ues (figure 3.12). Furthermore, it is clear both from the
overlap in these tuning curves and from multiple paral-
lel recording that many neurons are simultaneously en-
coding the same input stimuli.
It is often useful to think about distributed units as
representing
features
of input stimuli, where a given
stimulus is represented by a collection of units that each
Now, we will see how the network responds to letter
inputs instead of digits.
First, set the
g_bar_l
leak conductance back to
6(or
Defaults
) and make sure
act
is selected in
the NetView window. Then set
env_type
to
LETTERS
(
Apply
) and press
Run
.
Notice the letters being presented over the input layer
on the network.
,
!
Use the VCR-like controls at the top of the Grid-
Log to scroll the display back to the start of the letter
presentations, because these have scrolled off the “top”
of the display (the single
<
button is a good choice for
fine-grained scrolling).
The only significant response came to the “S” letter
input from the “8” hidden unit — note that “S” is very
similar to the “8”.
Press the fast-forward-to-the-end button (
>|
)on
the grid log, so that it will continue to scroll when you
next
Run
it. Press
Cluster
,
CLUSTER_LETTERS
.Next,
press
Cluster
and
CLUSTER_HIDDEN
.
The first cluster plot should look like figure 3.10a,
and the second should look like figure 3.10b. You
should be able to see in these cluster plots that these
digit units do not respond very informatively to the let-
ter stimuli.
,
!
Question 3.4 (a)
Based on your experiences in the pre-
vious question, what would you expect to happen to the
cluster plot of hidden responses to letter inputs as you
lowered the
g_bar_l
leak current to a value of 4? Do
this — were you right?
(b)
Would you say that this
hidden representation is a good one for conveying letter
identity information? Why or why not? (Hint: Pay par-
ticular attention to whether any letters are collapsed in
the cluster plot — i.e., having no distance at all between
them.)
(c)
Can you find any setting of
g_bar_l
that
gives you a satisfactory hidden representation of letter
information? Explain.
Go to the
PDP++Root
window. To continue on to
the next simulation, close this project first by selecting
,
!
Search WWH ::
Custom Search