1.  ({ri}) ri (x,t) r (x,t) r (t)
A Simplified History of Neural Complexity
Symbolic
 ({ri})
(Noam Chomsky) 3
106
107
108
109
yrs
infinite recursion
        mammalian species        
echidna
CA1
CA3 DG
platypus
Memory
ri (x,t)
(David Marr)
lizard 2 1
Spatial
r (x,t)
(Hubel & Wiesel)
Chemical
r (t)
(e.g. Peter Dayan)

2. Let us concentrate on one step on our way out of the primordial neural tube
first, a process of evagination

3. some slight change in our chips:
and then…
some slight change in our chips:

4. some slight change in our chips:
and then…
some slight change in our chips:

5. and then… some slight change in our chips: into the hippocampus
the medial wall reorganized
into the hippocampus

6. the reorganization includes a spatial migration...
...it does not lead to a
new type of cortex...

7. but the reorganization of the hippocampus...
….occurs, fundamentally, through the detachment and
granulation of the dentate gyrus
(diagrams taken from the
book by P. Gloor)

8. it is, at its core, a granulation!
…self-similar
across species

9. watch it frozen in
its development,
in the opossum

10. can we understand this granulation?

11. or even the whole circuitry in detail?

12. David Marr, over 30 years ago, suggested to start from the function
In humans,
the hippocampus
had long been
implicated
in the formation
of episodic and
autobiographical
memories
(here, data by
Graham & Hodges)

13. Over the last few
years, imaging
evidence has
corroborated
traditional
neuropsychological
evidence
(here, fMRI study
of verbal encoding
into episodic memory
by Fernandez et al)

14. In rats, the evidence
from neurophysiological
recordings indicates
a primary role in
spatial memory
(here, data from
simultaneous
recordings by
Matt Wilson &
Bruce McNaughton)

15. (although a
minority view
has emphasized
a more active
role in spatial
computation;
here, data by
Neil Burgess &
John O’Keefe)

16. In monkeys,
Edmund Rolls
et al have found
spatial view cells,
suggestive of a
hippocampal
role intermediate
between the
human and the
rat description

17. The extensive
extrinsic
connections of the
hippocampus with
the neocortex are
consistent with
a role in the
formation of
memories

18. (diagram by Jaap Murre, 1996)
David Marr’s perspective was the same later adopted by most of his followers...
(diagram by Jaap Murre, 1996)

23. Yet, birds use their hippocampus in a similar way...

25. So, let us follow the same functional hypothesis...
…but let us try to be quantitative

26. What sort of device can….
generate, on line, appropriate (compressed) representations of each “snapshot”
store these representations on line, in a single “shot”
hold multiple representations simultaneously
retrieve each representation from partial cues
send back the retrieved information in a robust format ?

27. Icue=log2 p << Iitem  Niunit
generate, on line, appropriate (compressed) representations of each “snapshot”
store these representations on line, in a single “shot”
hold multiple representations simultaneously
to retrieve each representation from partial cues
send back the retrieved information in a robust format
requires a content addressable memory
CAM
Icue=log2 p << Iitem  Niunit

28. The analysis of large-
scale recordings
(here, by Skaggs &
McNaughton) shows
that the information
content of hippocampal
representations grows
linearly with population
size, before saturating
at the ceiling set by
the experiment.
Francesco Battaglia has
quantified the full Iitem
for place cells, using
an analytical model.

29. (Francesco has shown that with maps it is just the same)
generate, on line, appropriate (compressed) representations of each “snapshot”
to store representations on line, in a single “shot”,
hold multiple representations simultaneously
retrieve each representation from partial cues
send back the retrieved information in a robust format
(with neuronally plausible mechanisms)
is something Hebbian associative networks can do
...
p ~ 0.2 C / [a log(1/a)]
(Francesco has shown that with maps it is just the same)

30. If LTP modifies the same synapses affected by learning...
one may predict
the effect of
lower C values
higher dp/dt

31. Carol Barnes found, contrasting young and old rats

32. multiple representations
generate, on line, appropriate (compressed) representations of each “snapshot”
to store representations on line, in a single “shot”
multiple representations
retrieve each representation from partial cues
send back the retrieved information in a robust format
are held most efficiently in a free autoassociator, which minimizes the components required for a given information content

33. CA3 is dominated by recurrent collaterals

34. the read-out of the retrieved information
generate, on line, appropriate (compressed) representations of each “snapshot”
store these representations on line, in a single “shot”
hold multiple representations simultaneously
retrieve each representation from partial cues
the read-out of the retrieved information
is greatly facilitated by expansion recoding with additional associative “polishing”

37. CA3 CA1 Analytical models predict an optimal plasticity level for
CA3->CA1 (Schaffer)
collaterals, but are not yet
constrained enough to
predict the observed memory
activation differences
CA3
CA1

38. …but, why do we need CA1, then?
the answer may lie in the predictive ability that several models assign to the hippocampus (as well as to any associative network with time-asymmetric plasticity)
although CA3 may predict future “contexts” as well as CA1, this may conflict with devoting its recurrent collaterals to retrieve the current “context”.
We shall get back to this issue tomorrow…

39. generating information-rich compressed representations
store these representations on line, in a single “shot”
hold multiple representations simultaneously
retrieve each representation from partial cues
send back the retrieved information in a robust format
requires a dedicated preprocessor that sparsifies and decorrelates input activity

40. Why the separation of a dentate gyrus?
MF inputs (from DG)
force an informative
representation at storage
and are irrelevant for
retrieval
PP inputs (from EC)
modify during
storage
and relay the cue at
retrieval

41. The crucial prediction is consistent with recordings from normal rats

42. but it is difficult to test it in dentate lesioned rats
(Tucson data
by Jim Knierim)
but it is difficult to test it in dentate lesioned rats

45. Double dissociation Acquisition Index: Errors (T1-5 D1) - (T6-10 D1)
Retrieval Index: Errors (T6-10 D1) - (T1-5 D2)
Double dissociation