Table 2 Summary of GLIF models and results

From: Generalized leaky integrate-and-fire models classify multiple neuron types

 

Num. cells

Variables

Model parameters

Parameters in clustering

Explained variance Δt = 10 ms

Num. clusters

GLIF 1

645

\(V(t)\)

R, C, E L , Θ∞, δt

R, C, E L , Θ∞, δt

70.2%

10

GLIF 2

254

\(V(t)\), \(\Theta _s(t)\)

R, C, E L , Θ∞, δt, f v , δV, b s , δΘ s

R, C, E L , f v , δV, Θ∞, δt

67.7%

15

GLIF 3

645

\(V(t)\), \(I_1(t)\), \(I_2(t)\)

RASC, C, E L , Θ∞, δt, k1, δI1, k2, δI2

RASC, C, E L , Θ∞, δt,\(\delta I_1/k_1\), \(\delta I_2/k_2\)

72.4%

18

GLIF 4

254

\(V(t)\), \(\Theta _s(t)\) \(I_1(t)\), \(I_2(t)\)

RASC, C, E L , Θ∞, δt, f v , δV, b s , δΘ s , k1, δI1, k2, δI2

RASC, C, E L , Θ∞, δt, f v , δV, \(\delta I_1/k_1\), \(\delta I_2/k_2\)

75.9%

16

GLIF 5

253

\(V(t)\), \(\Theta _s(t)\) \(I_1(t)\), \(I_2(t)\), \(\Theta _v(t)\)

RASC, C, E L , Θ∞, δt, f v , δV, b s , δΘ s , k1, δI1, k2, δI2, a v , b v

N/A: additional parameters are not available for all 645 neurons.

77.6%

N/A

All features

645

N/A

N/A

τ m , R i , Vrest, Ithresh†, Vthresh†, Vpeak†, Vfasttrough†, Vtrough† up:downstroke†, up:downstroke*, sag, f–I curve slope, latency, max. burst index

N/A

16

Sub-thr features

645

N/A

N/A

τ m , R i , Vrest, Ithresh†, Vthresh†, Vtrough†, sag, f–I curve slope, latency, max. burst index

N/A

16

  1. The “Num. cells” column reports the number of cells for which a model was constructed or the paradigm was clustered. Note that for the GLIF models clustering was only performed on parameters that were available for all 645 cells. Therefore, the “Parameters in clustering” list is a subset of the total “Model Parameters” available for any level. The variables for each model level are listed in the “Variables” column. Note that resistance was fit along with after-spike currents in models where after-spike currents were implemented. R denotes the resistance fit without ASC and RASC denotes the resistance fit along with after-spike currents. GLIF5 does not report clusters because there are no additional parameters available for all 645 neurons, i.e., it would be the same clustering paradigm as in GLIF4 as the only new parameters associated with GLIF5 (a v and b v ) are only fit for the reduced set of cells. We are unable to cluster on the time scales of every after-spike current alone as there were five discrete possible values but only two were chosen for each neuron. Therefore, we cluster on the total charge deposited over short \(\delta I_1/k_1\) and long \(\delta I_2/k_2\) time scales (continuous numbers) for the model levels that contain after-spike currents. The average explained variance at a time resolution of 10 ms for all neurons at each level is reported as well as the number of clusters that were found using the aforementioned clustering parameters via the hierarchical clustering technique. As in Supplementary Tables 6 and 7, * denotes features measured during a short square stimulus, and † represents features measured during a long square stimulus (Fig. 2)
Back to article page