Computational Neuroscience – 10 neocortex
Introduction
These notes are about memory in the neocortex; we will look at two type of neocortical memory,
recognition memory, which is believed to occur in the perirhinal cortex, and the sort of long
term memory which is stored in what are called the association cortex, which is a general term
for that part of the neocortex that has no other name.
Memory
We have looked at memory, medium term declarative memory, in the specialized circuits that
make up the hippocampus. This is only part of memory, here we will briefly examine some
other memory systems in cortex, specifically, recognition memory and the way some declarative
memories are stored in cortex.
Recognition memory
In situations where we might fail to name the items that make up the contents of a room, we
are often still able to spot a new or unfamiliar item, an item that has been added since the
room became familiar to us. In fact, we are believed to have a separate system for recognition
memory : there was for some times a debate between the idea that familiarity is a weak form
of recognition, rather than a separate type of memory, it is now believed that it is the latter,
with evidence from psychology [1] and neuroscience [2]. It is believed that recognition memory
is stored in the perirhinal cortex, an area of cortex immediately beside the entorhinal cortex
we discussed as part of the hippocampus.
Perhaps the most striking thing about recognition memory is its almost limitless capacity,
see Fig. 1. This is thought to be the consequence of the simpler task it performs, as illustrated
in Fig. 2, recognizing familiarity takes far fewer connections than recalling a memory. In fact,
we saw before that the memory capacity for an auto-associative network such as that attributed
to CA3 is
P =
0.035
a
cN (1)
where a is the sparseness, c is the fraction of pairs that are connected and N is the number of
neurons. The comparable formula for a recognition network is
P = 0.023cN2. (2)
The N2 rather than N gives it a vastly larger capacity [4].
This doesn’t resolve how recognition is done; in fact there is a diversity of approaches used.
One interesting application is found in [5] where recognition memory is used as a mechanism
for visual navigation in ants. It is likely that any memory is stored using a small subset of its
features.
Declarative memories stored in cortex
We have already examined the way declarative memories are stored in the hippocampus; there
are also declarative memories stored in cortex, but, it seems, the cortical memories are stored
in a different way. While the hippocampus seems capable of storing memories very quickly
with little chance of overlap, the cortex appears to learn memories slowly, in a way that
discovers overlaps and exploits them for learning. The idea is that memories are first stored
1
Computational Neuroscience – 10 neocortex
10
100
1000
10000
10 100 1000 10000
M
S
Figure 1: Recognized images plotted against presented images. S is the number of images
presented, M the number recognized as familiar in a forced-choice task. The data
points are plotted from data presented in [3] and the line represents log10M =
0.93 log10 S + 0.08..
A
Frog Pond Plop
Basho dsh Keats
B
Frog Pond Plop
Figure 2: Recognition circuits are simpler. In A is a schematic for recognizing dsh’s translation
of Matsuo Basho’s haiku, with different nodes for the different concepts and, as a
consequence, a large number of connections. In contrast, in B the goal is just to
recognize the poem as familiar, so there is one familiarity neuron and far fewer
connections.
2
Computational Neuroscience – 10 neocortex
in hippocampus and, over time, if they are important, they are learnt from their to neocortex,
where they are stored in a rich, interconnected, way.
Consider a simple model neuron with output
xi =
∑
wijxj (3)
where wij are connection strengths. Now the perceptron learning rule is
∆wij = η(hi − xi)xj (4)
where hi is the desired output and xi the actual output. With small η this should converge
over time to give hi = xi. The key here is that η is small, so convergence is slow. Consider
learning a series of birds
eagle
wren
parrot
This is the sort of disastrous correlation we discussed in the hippocampal context, hypothesising
that random connections related to neurogenisis are employed to lift the degeneracy. Here,
however, it allows properties to be deduced and stored.
eagle
can fly
The problem is that if a new bird is stored, this might lead to an erroneous conclusion
penguin
can fly
However, if the sequences are presented enough times and the learning rate is low enough, the
perceptron should be able to store the correct information
penguin
can fly can swim
3
References
but
wren
can fly can swim
In this way hippocampus and neocortex are thought to solve different memory problems, a
helpful analogy is given in [6]: imagine you drive to work each day, the hippocampus would
store the location of where you parked that morning so you could get back to your car in the
evening, your neocortex would learn which places are good for parking.
References
[1] Yonelinas AP. (2002) The nature of recollection and familiarity: A review of 30 years of
research. Journal Memory and Language. 46: 441–571.
[2] Brown MW and Aggleton JP. (2001) Recognition memory: What are the roles of the
perirhinal cortex and hippocampus. Nature Review Neuroscience. 2: 51–6
[3] Standing L. (1973) Learning 10000 pictures. The Quarterly Journal of Experimental
Psychology. 25: 207–22.
[4] Bogacz R, Brown MW and Giraud-Carrier C. (2001) Model of familiarity discrimination
in the perirhinal cortex. Journal of Computational Neuroscience. 10: 5–23.
[5] Baddeley B, Graham P, Husbands P and Philippides A (2012) A model of ant route
navigation driven by scene familiarity. PLoS Computational Biology. 8: e1002336.
[6] McClelland JL, McNaughton BL and O’Reilly RC. (1995) Why there are complementary
learning systems in the hippocampus and neocortex: insights from the successes and
failures of connectionist models of learning and memory. Psychological Review 102: 419.
4