A long standing problem in neuroscience, both in modeling
and in data analysis, is the one of inferring synaptic
couplings from correlations of the sampled neural
activities. The recent availability of techniques allowing
simultaneous recording from several tens electrodes,
in-vitro as well as in-vivo, gave new momentum to research
in this direction. In particular, much effort has been
devoted to develop and refine inference methods inspired by
the statistical mechanics of spin systems (so called
‘inverse Ising’ methods). In the original proposal
(Schneidman E, Berry MJ, Segev R and Bialek W, Nature 440,
1007-1012, 2006), simultaneously recorded data are binned in
time, discretized and interpreted as successive
configurations of a spin system with pairwise interactions
at equilibrium. Inference proceeds then as the solution of a
constrained optimization problem: determine the spin
couplings providing the maximum entropy (Gibbs) distribution
compatible with the observed mean activities and pair
spatial correlations, used as constraints. ‘Brute-force’
solutions can be obtained by iterative procedures akin to
learning algorithms in Boltzmann machines; the need to
reduce the computational load for large networks motivated
the use of various forms of mean-field estimates of the
correlations from the measured mean activities. The interest
in relaxing the assumption of equilibrium later led to the
development of inference methods based on kinetic Ising
models; for a review of the state of the art see Hertz J,
Roudi Y, Tyrcha J, in "Principle of Neural Coding" S.
Panzeri and R. Q. Quiroga eds, CRC Press 2013.
In this work, using simulations of networks of
integrate-and-fire neurons, we incorporate in kinetic
inverse Ising inference methods the important notion that
spikes are transmitted between neurons with delays, which
are estimated from the profile of the cross-correlation
function prior to the inference procedure, and suggest the
right choice of the time bin used in the inference
algorithm. A method is also developed to take into account a
finite time of integration of the synaptic input. Finally,
we analytically and numerically study the relationship
between the inferred and the real synaptic efficacies, and
how the choice of the time bin affects it. Such
relationships turns out to be quadratic both for excitatory
and inhibitory synapses, but it depends critically on the
time bin for the excitatory synapses only, while being
essentially independent of the time bin for the inhibitory ones.
Work in collaboration with C. Capone, P. Del Giudice, C.
Filosa, G. Gigante.