Email updates

Keep up to date with the latest news and content from BMC Neuroscience and BioMed Central.

This article is part of the supplement: Eighteenth Annual Computational Neuroscience Meeting: CNS*2009

Open Access Open Badges Poster presentation

Fast inference of couplings between integrate-and-fire neurons from their spiking activity

Simona Cocco12, Stanislas Leibler23 and Rémi Monasson23*

Author Affiliations

1 CNRS-Laboratoire Physique Statistique de l'ENS, Paris, 75005, France

2 The Rockefeller University, New York, USA

3 CNRS-Laboratoire Physique Theorique de l'ENS, Paris, 75005, France

For all author emails, please log on.

BMC Neuroscience 2009, 10(Suppl 1):P128  doi:10.1186/1471-2202-10-S1-P128

The electronic version of this article is the complete one and can be found online at:

Published:13 July 2009

© 2009 Cocco et al; licensee BioMed Central Ltd.

Poster presentation

Multi-electrode recordings make available the simultaneous spiking activity of tens of neurons for hours [1]. An important issue is to reconstruct the network of connections between the cells from this activity. To be more precise, let us model cells as Leaky Integrate-and-Fire (LIF) neurons (see [2] and references therein) whose membrane potentials obey the differential equation (units are chosen so that the membrane capacitance equals unity),


where g is the leak conductance. Jij is the strength of the connection from neuron j onto neuron i and tj, k the time at which cell j fires its kth spike; we assume that synaptic inputs are instantaneously integrated i.e. the synaptic integration time is much smaller than 1/g and the typical inter-spike interval. Ii is a constant external current flowing into cell i, and ηi(t) is a fluctuating current, modeled as a Gaussian white noise process with variance σ2. Neuron i remains silent as long as Vi remains below some threshold potential (equal to1). If the threshold is reached at some time then a spike is emitted, and the potential is reset to its rest value of zero, and the dynamics resumes.

The above model implicitly defines the likelihood P of the spiking times {tj, k} given the currents Ii and synaptic couplings Jij. If we are given the spike times {tj, k} we will infer the couplings and currents by maximizing P. In principle, P can be calculated through the resolution of Fokker-Planck equations (one for each inter-spike interval) for a one-dimensional Orstein-Uhlenbeck process with moving boundaries. However, this approach, or related numerical approximations [3], are inadequate to treat data sets with hundreds of thousands of spikes (such as in recordings of populations of retinal ganglion cells) in a reasonable time.

In our approach, we choose to approximate P from the contribution coming from the most probable trajectory for the potential for each cell i, referred to as Vi *(t). This approximation is exact when the amplitude σ of the noise is small. The determination of Vi *(t) was done numerically by Paninski for one cell in [4]. We have found a fast algorithm to determine Vi *(t) analytically in a time growing linearly with the number of spikes and quadratically with the number of neurons, which allows us to process recordings with tens of neurons easily. The algorithm is based on a detailed and analytical resolution of the coupled equations for the optimal potential Vi *(t) and the associated optimal noise ηi *(t) through (1), and is too complex to be explained in this abstract.

Once the optimal paths for the potential and noise has been determined, we calculate the log-likelihood of the corresponding couplings and currents through the integral of the squared optimal noise [4]. This log-likelihood is clearly a concave function of the currents and couplings and can be easily maximized using the Newton-Raphson method. Our algorithm has been tested on artificially generated data, real experimental data (recordings of 32 ganglion cells in the salamander retina submiited to random flickering stimulus, courtesy of M. Meister) and compared to other inference methods based on the Ising model (see abstract by Cocco). To give a flavor of the computational effort required, it takes us about 20 seconds to process a set of 120,000 spikes fired by 32 cells on a personal computer.


  1. Taketani M, Baudry L: Advances in Network Electrophysiology Using Multi-Electrode Arrays. Springer-Verlag, New York; 2006. OpenURL

  2. Jolivet R, Lewis TJ, Gertsner W: Generalized integrate-and-fire models of neuronal activity approximate spike trains of a detailed model to a high degree of accuracy.

    J Neurophys 2004, 92:954. Publisher Full Text OpenURL

  3. Paninski L, Pillow JW, Simoncelli EP: Maximum likelihood of a stochastic integrate-and-fire neural econding model.

    Neural Computation 2004, 16:25-53. Publisher Full Text OpenURL

  4. Paninski L: The most likely voltage path and large-deviation approximations for integrate-and-fire neurons.

    J Comput Neurosci 2006, 21:71. PubMed Abstract | Publisher Full Text OpenURL