LIMBO: Information measures
The
aim is to establish methods that allow the
correct quantification of the information content of neuronal activity
recorded
in neurophysiological experiments. An
analytical
procedure has been developed that allows to eliminate the limited
sampling
error which previously invalidated attempts to quantify neural
transmission of
information in mammals [1,2]. A joint
project has
compared the efficiency of the procedure with that of an alternative
one
proposed by the group of John Hertz [5].
Information
analyses of both
single neurons and, through decoding procedures, of
populations
of
neurons, have been applied to recordings from the lab of Edmund Rolls
(including
monkey visual cortex [6,7] and orbito-frontal
cortex
[4]), and seminal collaborations have been initiated with other labs,
including
that of Mathew Diamond [3,8].
Subsequent
work has focused
on the short time limit [3], that allows the accurate
quantification of
the effects of decoding [9,17] and of correlations among simultaneously
recorded cells [10]; interactions among different neurons have also
been analyzed
in terms of firing coincidences (VdP
with Laura Martignon). We have then
analyzed
how the information about a set of discrete stimuli depends on the size
of the
neuronal population, both in the short time limit for correlated units
[16] and
over finite times with uncorrelated units [12], leading to a validation
of the
simple Gawne & Richmond model
(reviewed in [11]: the
figure to the left shows how the ceiling alone captures most of the
redundancy
when a population of IT units has to discriminate between 4, 9 or 20
face
stimuli). Stimulated by a collaboration with Eilon
Vaadia, we have explored the coding of more
structured
stimuli, with both a discrete and a continuous dimension, at the level
of
theoretical modelling [13,15]
and in real neurons recorded from the motor cortex of monkeys
performing
bimanual movements [14]. Recently we have extended decoding procedures
to deal with
the contribution of correlations [18,19].
The most recent developments are in connection with the hippocampal recordings in the lab of May-Britt and Edvard Moser, which required information measures obtained without explicit reference to the salient correlate, the position of the rat in the environment. We developed measures based on the distribution of population vectors, obtained from the spike counts over short time bins by simultaneously recorded units [20]. Such measures can be used to assess information content about spatial position, spatial context, etc.
Work in progress applies simplified network models and replica techniques [12, 15] to estimate measures of information content in newly stored spatial representations (EC, AT).
References:
- AT
& SP, Neural Comp 7:399-407 (1995)
- SP
& AT, Network
7:87-107 (1996)
- SP, G
- ET Rolls, HD Critchley
& AT, J Neurophysiol 75: 1982-1996 (1996)
- D Golomb,
J Hertz, SP, AT & BJ
- ET Rolls, AT &
MJ Tovee, Exp Brain Res 114:149-162 (1997)
- ET Rolls, AT, MJ Tovee
& SP, J Comput Neurosci 4:
309-333 (1997)
- RP,
AT, M Lebedev & ME Diamond, Soc
Neurosci abs 23:913.17 (1997)
- SP,
SRS, AT & ET Rolls, Neural Comp 11: 1553-1577
(1999)
- SP,
SRS, AT & ET Rolls, Proc Roy Soc B 266:
1001-1012 (1999) (with electronic appendices)
- AT,
you are welcome to ftp an early draft.ps
of a chapter from the Handbook of Biological Physics, vol.4,
803-829 (2000), edited by F Moss & S Gielen
- IS
& AT, Phys.
Rev. E 63, art. 011910 (2001)
- VdP &
AT, (cond-mat/0103286 ), Phys. Rev. E 64 art.
021912 (2001)
- VdP,
O.Steinberg, AT
& E. Vaadia, Neurocomputing 38-49: 1181-1184 (2001)
- VdP &
AT, (cond-mat/0202315 ), Phys. Rev. E 65 art.
041918 (2002)
- MB,
ME Diamond & AT, J Comput Neurosci
12:165-174 (2002)
- IS,
(physics/0110074),
Neural Comp. 14: 771-779 (2002)
- LF, ET Rolls, NC Aggelopoulos
& AT, Exp Brain Res 155:370-384 (2004)
- ET
Rolls, NC Aggelopoulos, L Franco & AT, Biological Cybernetics
90:19-32 (2004)
- S Leutgeb,
JK Leutgeb, AT, M-B
Moser & EI Moser, Science
305:1295-1298 (2004)