Large-Scale Brain Area Functional Organization - Computational Explorations in Cognitive Neuroscience - page 211

1−

However, survival in the world demands that rapid
learning also occur, because specific, arbitrary informa-
tion is also very important (e.g., remembering where
you parked your car today, or remembering which cave
you hid your food in and distinguishing that from the
cave you saw the bear family enter last night). Notice
also that in this rapid form of learning, the memories
of individual episodes should be kept separate instead
of integrating across them (e.g., today's parking spot
should not be confused with yesterday's, nor should one
cave be confused with the other). Because learning can-
not be both fast and slow (or both integrating and sep-
arating), there is a basic tradeoff between the demands
of slow integration and rapid separating.
It seems that the brain has resolved this tradeoff by
allowing the cortex to learn slowly and integrate over
experiences, while the hippocampus provides a comple-
mentary rapid, separating learning system. This idea is
consistent with a large amount of data on people and an-
imals with hippocampal lesions (see McClelland et al.,
1995, for a review). For example, a famous patient,
known by his initials “HM,” had large chunks of his
medial temporal cortex (including the hippocampus) re-
moved bilaterally to prevent epilepsy that was originat-
ing there. HM was subsequently unable to learn much
of any new information about the people he met, the
events that occurred, and the like. However, he was
able to learn a number of relatively complex perceptual-
motor tasks, and showed other forms of intact learn-
ing that is characteristic of posterior-cortical learning.
Perhaps the clearest distinction between the role of
the hippocampus and that of the cortex comes from a
group of people who had relatively pure hippocampal
lesions early in life, and yet have acquired normal lev-
els of long-term knowledge about the world and tasks
like reading (Vargha-Khadem, Gadian, Watkins, Con-
nelly, Van Paesschen, & Mishkin, 1997). Their primary
deficit is on tasks that require rapid learning of arbitrary
information. Chapter 9 explores these ideas further.

0.8−

0.6−

0.4−

lrate = 1

lrate = .1

0.2−

lrate = .005

0−

0 5 10 15 20 25

Figure 7.6: Effect of learning rate lrate on the ability of
a weight to represent underlying conditional probability of
the input unit activity given the output unit activity (i.e., the
CPCA Hebbian learning objective). This conditional proba-
bility was .7, as is reflected by the .005 lrate case. With lrates
of .1 or 1, the weight bounces around too much with each
training example (which is binary), making it impossible to
represent the overall probability that is only apparent across a
number of individual examples.

ing mechanisms described in chapters 4-6 to develop
representations of the important underlying structural
and statistical characteristics of the world, to process
perceptual inputs effectively, and to produce systematic
and useful motor outputs. Given this, learning in the
cortex must necessarily be slow to integrate over many
individual experiences and extract the general underly-
ing regularities of the environment (McClelland, Mc-
Naughton, & O'Reilly, 1995; White, 1989b).
Figure 7.6 shows a simple example of this point,
taken from the exploration described in section 4.6,
demonstrating that a slow learning rate enables the
weight to converge on the actual underlying conditional
probability of an event occurring in the environment. If
somehow the world were to provide underlying statis-
tical regularities all in one experience , then you might
be able to get away with a faster learning rate. How-
ever, because each experience is typically just a small
and probably noisy fragment of the overall picture, slow
learning must be used to blend this fragment smoothly
together with all the others. Note that by virtue of in-
tegrating each episode together with previous ones, the
unique details specific to these episodes are lost, with
only some faint residue remaining.

7.4.2 Active Memory versus Overlapping

Distributed Representations

A different type of tradeoff can be used to understand
the difference between the posterior and frontal cortex,