Information Technology Reference
In-Depth Information
specialized memory systems, instead of one system that
tries to achieve a compromise.
Going down a level of detail, the notion of pattern
separation fits quite well with the detailed biological
properties of the hippocampal formation, providing fur-
ther support for the idea that this structure is specialized
for the pattern separation function. In short, there is
considerable synergy here between computational prin-
ciples and behavioral and neuroscience data.
Although we find the present model of hippocampal
function compelling, there is considerable controversy
in the literature regarding the best way to characterize
the functional role of the hippocampus. For example,
a number of researchers support the notion originally
developed by O'Keefe and Nadel (1978) that the hip-
pocampus implements a spatial map, and is used pri-
marily for spatial navigation. One stimulus for this
model was the discovery that hippocampal neurons in
the rat seem to encode specific spatial locations (e.g.,
figure 9.9). We have seen that this kind of specific, con-
junctive firing can be explained in terms of the more
general function of forming sparse, pattern separated
representations. Furthermore, there is little evidence
that there is any kind of topography to the hippocampal
representations, calling into question its role as a map.
Instead, it might be better viewed as producing episodic
memories of spatial locations (i.e., binding together all
the features present in a given location), without nec-
essarily encoding much about the spatial relationships
among different locations (e.g., McNaughton & Nadel,
1990).
Another body of literature dealing with the func-
tion of the hippocampus comes from animal learn-
ing paradigms such as conditioning and discrimination
learning. Sutherland and Rudy (1989) proposed that
the hippocampus is necessary to learn nonlinear prob-
lems that require a configuration or conjunction of stim-
uli to be formed (we explored this kind of problem us-
ing the “impossible” task in chapter 5). For example, in
the
negative patterning
problem, rats have to learn that
stimulus
A
leads to reward (
A
+
), as does stimulus
B
(
B
+
), but that both
A
and
B
together do not (
AB
)
(this is the same as the XOR task studied by Min-
sky & Papert, 1969; Rumelhart et al., 1986a). Rats with-
out a hippocampus are impaired on learning this prob-
lem, suggesting that the hippocampus might be useful in
learning that the conjunction of A and B is different than
the two stimuli individually. This view is consistent
with the idea suggested above that the sparse, pattern
separated representations in the hippocampus are also
conjunctive
, in that they bind together different stimu-
lus elements.
Subsequent experimentation with cleaner hippocam-
pal lesions has shown that the hippocampus is not al-
ways (or even usually) necessary to learn nonlinear
problems (Rudy & Sutherland, 1995). These findings
are consistent with the idea that the cortex can learn
complex tasks through error-driven learning (as ex-
plored in chapter 5), but that this learning must be slow
— even intact rats take many iterations to learn com-
plex nonlinear problems, as would be expected of cor-
tical learning. The balance between pattern separation
and pattern completion is also relevant here, because it
is likely that an intact hippocampus will be doing task-
inappropriate pattern completion in these complex non-
linear tasks, with many epochs of error-driven learning
necessary to achieve the appropriate hippocampal rep-
resentations. O'Reilly and Rudy (in press) contains an
implemented model (based on the hippocampal model
we just explored) and fuller discussion of these phe-
nomena.
The hippocampus has been a popular target of com-
putational modeling, and there are a large number of
other models in the literature, many of which share ba-
sic ideas in common with our model (e.g., Treves &
Rolls, 1994; Hasselmo & Wyble, 1997; Moll & Mi-
ikkulainen, 1997; Alvarez & Squire, 1994; Levy, 1989;
Burgess, Recce, & O'Keefe, 1994; Samsonovich & Mc-
Naughton, 1997), but some that make very different
claims about the role of the hippocampus in learning.
For example, two models (Schmajuk & DiCarlo, 1992;
Gluck & Myers, 1993) claim that only the hippocam-
pus is capable of performing error-driven learning, and
is used to “teach” the cortex. Such models are incon-
sistent with recent data showing that children with very
early hippocampal lesions nevertheless acquire seman-
tic information normally, while still suffering from sig-
nificant episodic (i.e., hippocampal) memory deficits
(Vargha-Khadem et al., 1997).
Search WWH ::
Custom Search