In their paper, Pozzorini et al. introduce a very elegant and methodologically sophisticated approach to automated single-cell model estimation. Approaches like these allow for fast, 'high-throughput' computational characterization of physiological recordings. This may be an area of increasing importance for translating large and steadily growing physiological data bases efficiently into data-driven computational neuron models.

We would like to point out that an alternative recent approach to fully automated, predictive, 'high-throughput' model estimation and validation for large physiological data sets was introduced in

Hertäg et al. (2012), An approximation to the adaptive exponential integrate-and-fire neuron model allows fast and predictive fitting to physiological data. Frontiers in Computational Neuroscience, Vol. 6, Article 62.
http://journal.frontiersi...

This latter approach differs methodologically from the one introduced by Pozzorini and colleagues in various ways: First, it is based on closed-form expressions for instantaneous and steady-state f/I curves derived from an approximation to the AdEx (Brette & Gerstner 2005) model. Estimation in this case does not require numerical integration of the model's differential equations, making the process efficient and fast. Second, physiological training data in this case indeed consist of conventional step-current protocols. This has the advantage that a large pool of already existing in-vitro datasets could potentially be harvested, although we agree with Pozzorini et al. that step-current protocols may not necessarily be the best way to capture a neuron's dynamical properties.

Nevertheless, although parameter estimation is based on standard f/I curves, it is shown that the derived model can accurately predict spike times, within the limits of physiological reliability, on independent validation sets produced by fluctuating (in-vivo-like) current injections. Since these validation traces have a statistical structure very different from the training data, these results also firmly rule out potential issues with over-fitting, as explicitly examined in this work by comparing model to real-cell vs. real-cell to real-cell predictions. Using summary statistics like f/I curves furthermore avoids potential problems in parameter estimation resulting from fractal/rugged optimization landscapes that could plague models with latent variables and possibly chaotic dynamics (e.g. Wood 2010, Nature). In Pozzorini et al., the mathematical definition of the model results in a concave log-likelihood function, thus circumventing this problem as well.