Fig. 2
From: Generalized leaky integrate-and-fire models classify multiple neuron types
Overall, 645 different neurons from 16 transgenic lines containing all the required stimuli on the Allen Cell Types Database are considered in this study. (Left) Illustrated colors correspond to the different transgenic lines in all figures. “n” describes the number of neurons for which the lowest level model (GLIF1) could be generated. Transgenic lines can identify either or both inhibitory (I) or excitatory (E) cells, which reside in layer 1 (L1) through layer 6 (L6). Note that most Chrna2-Cre-positive neurons are inhibitory and thus are labeled here as inhibitory. (Right) A minimal set of stimuli were required for training and testing different GLIF models. GLIF models were trained using a at least two repeats of pink noise stimuli (3 s each, 1/f distribution of power, 1 – 100 Hz) with amplitudes centered at 75, 100, and 125 percent of action potential threshold, b a short (3 ms) just supra-threshold pulse to fit the instantaneous threshold, Θ∞, c a long square (1 s) pulse just below threshold to estimate the intrinsic noise present in the voltage traces (used in the post hoc optimization step), and d a series of three peri-threshold short pulse sets for any model with reset rules (GLIF2, GLIF4, and GLIF5). GLIF models were then tested using a hold-out stimulus set e of at least two sweeps of a second pink noise stimuli generated in an identical manner to the training but initialized with a different random seed. Representative data shown from an Htr3a-Cre-positive neuron (resting membrane potential (RMP) −67 mV) and a Ctgf-Cre-positive neuron (RMP = −74 mV). Scale bar in a corresponds to 40 mV (a, e), 50 mV (b, d), or 20 mV (c)
