Search

Fitting models to data is an important part of the practice of science, made almost ubiquitous by advances in machine learning. Very often however, fitted solutions are not unique, but form an ensemble of candidate models -- qualitatively different, yet with comparable quantitative performance. One then needs a criterion which can select the best candidate models, or at least falsify (reject) the worst ones. Because standard statistical approaches to model selection rely on assumptions which are usually invalid in scientific contexts, they tend to be overconfident, rejecting models based on little more than statistical noise. The ideal objective for fitting models is generally considered to be the risk: this is the theoretical average loss of a model (assuming unlimited data). In this work we develop a nonparametric method for estimating, for each candidate model, the epistemic uncertainty on its risk: in other words we associate to each model a distribution of scores which accounts for expected modelling errors. We then propose that a model falsification criterion should mirror established experimental practice: a falsification result should be accepted only if it is reproducible across experimental variations. The strength of this approach is illustrated using examples from physics and neuroscience., Comment: v2: Fixed correspondence email 29 pages, 9 figures, 3 tables An online version of this paper is available at https://alcrene.github.io/emd-paper/abstract.html

Neuroscience and artificial intelligence are closely intertwined, but so are the physics of dynamical system, philosophy and psychology. Each of these fields try in their own way to relate observations at the level of molecules, synapses, neurons or behavior, to a function. An influential conceptual approach to this end was popularized by David Marr, which focused on the interaction between three theoretical 'levels of analysis'. With the convergence of simulation-based approaches, algorithm-oriented Neuro-AI and high-throughput data, we currently see much research organized around four levels of analysis: observations, models, algorithms and functions. Bidirectional interaction between these levels influences how we undertake interdisciplinary science., Comment: neuroscience, cognitive neuroscience, computational neuroscience, machine learning, artificial intelligence, philosophy

Mounting evidence in recent years suggests that astrocytes, a sub-type of glial cells, not only serve metabolic and structural support for neurons and synapses but also play critical roles in regulation of proper functioning of the nervous system. In this work, we investigate the effect of astrocyte on the spontaneous firing activity of a neuron through a combined model which includes a neuron-astrocyte pair. First, we show that an astrocyte may provide a kind of multistability in neuron dynamics by inducing different firing modes such as random and bursty spiking. Then, we identify the underlying mechanism of this behavior and search for the astrocytic factors that may have regulatory roles in different firing regimes. More specifically, we explore how an astrocyte can participate in occurrence and control of spontaneous irregular spiking activity of a neuron in random spiking mode. Additionally, we systematically investigate the bursty firing regime dynamics of the neuron under the variation of biophysical facts related to the intracellular environment of the astrocyte. It is found that an astrocyte coupled to a neuron can provide a control mechanism for both spontaneous firing irregularity and burst firing statistics, i.e., burst regularity and size., Comment: 11 pages, 11 figures; accepted for publication in Chaos: An Interdisciplinary Journal of Nonlinear Science, 2022

To understand how rich dynamics emerge in neural populations, we require models exhibiting a wide range of activity patterns while remaining interpretable in terms of connectivity and single-neuron dynamics. However, it has been challenging to fit such mechanistic spiking networks at the single neuron scale to empirical population data. To close this gap, we propose to fit such data at a meso scale, using a mechanistic but low-dimensional and hence statistically tractable model. The mesoscopic representation is obtained by approximating a population of neurons as multiple homogeneous `pools' of neurons, and modelling the dynamics of the aggregate population activity within each pool. We derive the likelihood of both single-neuron and connectivity parameters given this activity, which can then be used to either optimize parameters by gradient ascent on the log-likelihood, or to perform Bayesian inference using Markov Chain Monte Carlo (MCMC) sampling. We illustrate this approach using a model of generalized integrate-and-fire neurons for which mesoscopic dynamics have been previously derived, and show that both single-neuron and connectivity parameters can be recovered from simulated data. In particular, our inference method extracts posterior correlations between model parameters, which define parameter subsets able to reproduce the data. We compute the Bayesian posterior for combinations of parameters using MCMC sampling and investigate how the approximations inherent to a mesoscopic population model impact the accuracy of the inferred single-neuron parameters., Comment: 1st revision: 48 pages, 13 figures Improved statistical validation of results. Rewrite of Section 4.2 to clarify the link between the mesoscopic model and a transport equation. Multiple small improvements to the presentation Original: 46 pages, 12 figures

We study temporal correlations of interspike intervals (ISIs), quantified by the network-averaged serial correlation coefficient (SCC), in networks of both current- and conductance-based purely inhibitory integrate-and-fire neurons. Numerical simulations reveal transitions to negative SCCs at intermediate values of bias current drive and network size. As bias drive and network size are increased past these values, the SCC returns to zero. The SCC is maximally negative at an intermediate value of the network oscillation strength. The dependence of the SCC on two canonical schemes for synaptic connectivity is studied, and it is shown that the results occur robustly in both schemes. For conductance-based synapses, the SCC becomes negative at the onset of both a fast and slow coherent network oscillation. Finally, we devise a noise-reduced diffusion approximation for current-based networks that accounts for the observed temporal correlation transitions., Comment: 16 pages, 18 figures, 3 appendices. Accepted for publication in Physical Review E

The theoretical description of non-renewal stochastic systems is a challenge. Analytical results are often not available or can only be obtained under strong conditions, limiting their applicability. Also, numerical results have mostly been obtained by ad-hoc Monte--Carlo simulations, which are usually computationally expensive when a high degree of accuracy is needed. To gain quantitative insight into these systems under general conditions, we here introduce a numerical iterated first-passage time approach based on solving the time-dependent Fokker-Planck equation (FPE) to describe the statistics of non-renewal stochastic systems. We illustrate the approach using spike-triggered neuronal adaptation in the leaky and perfect integrate-and-fire model, respectively. The transition to stationarity of first-passage time moments and their sequential correlations occur on a non-trivial timescale that depends on all system parameters. Surprisingly this is so for both single exponential and scale-free power-law adaptation. The method works beyond the small noise and timescale separation approximations. It shows excellent agreement with direct Monte Carlo simulations, which allows for the computation of transient and stationary distributions. We compare different methods to compute the evolution of the moments and serial correlation coefficients (SCC), and discuss the challenge of reliably computing the SCC which we find to be very sensitive to numerical inaccuracies for both the leaky and perfect integrate-and-fire models. In conclusion, our methods provide a general picture of non-renewal dynamics in a wide range of stochastic systems exhibiting short and long-range correlations., Comment: 14 pages, 12 figures, 1 appendix. Accepted for publication in Physical Review E

In two recent papers, Rudolph and Destexhe (Neural Comp. {\bf 15}, 2577-2618, 2003; Neural Comp. in press, 2005) studied a leaky integrator model (i.e. an RC-circuit) driven by correlated (``colored'') Gaussian conductance noise and Gaussian current noise. In the first paper they derived an expression for the stationary probability density of the membrane voltage; in the second paper this expression was modified to cover a larger parameter regime. Here we show by standard analysis of solvable limit cases (white-noise limit of additive and multiplicative noise sources; only slow multiplicative noise; only additive noise) and by numerical simulations that their first result does not hold for the general colored-noise case and uncover the errors made in the derivation of a Fokker-Planck equation for the probability density. Furthermore, we demonstrate analytically (including an exact integral expression for the time-dependent mean value of the voltage) and by comparison to simulation results, that the extended expression for the probability density works much better but still does not solve exactly the full colored-noise problem. We also show that at stronger synaptic input the stationary mean value of the linear voltage model may diverge and give an exact condition relating the system parameters for which this takes place., Comment: 22 pages, 6 figures; new results have been added including a criticism of a recent result for the probability density of the membrane voltage by Rudolph and Destexhe