COSYNE

Log in

Detailed neuronal models are increasingly being used to address computational questions at the level of the single neuron as well as that of small and large-scale networks. Such models contain from but a few to several dozen and more parameters that must be constrained by experimental data. A good model is one that correctly predicts the experimental response to stimuli that were not used during the parameter constraining process. In order to explore this notion of predictability we record the in-vitro firing response of cortical inhibitory interneurons to both suprathreshold current steps of different amplitudes and current ramps of different magnitudes. Using this data, we constrain the model parameters by stochastic optimization of feature based distance functions (Druckmann et al. 2007, Frontiers in Neuroscience, 1: pp 7-18) We show that when comparing model to experiment within a given stimulus type (ramp or step), the larger the set of stimuli used to constrain the model, the greater the average error of the parameter constraining process (training error). Yet at the same time the error of the model on stimuli that were not used during the training (generalization error) decreases. In addition, we show that models trained on step currents generalize well to ramp currents but the reverse is not necessarily true. We conclude that the predictability of neuron models can be quantitatively measured in the aforementioned manner and that models constrained by feature based distance functions indeed predict quantitatively even stimuli of different dynamical nature (step vs. ramp).