Efficient "Shotgun" Inference of Neural Connectivity from Highly Sub-sampled Activity Data
Fig 4
Network connectivity can be reasonably estimated, even with model mismatch.
The panels (A-Q) are the same as Fig 3, where instead of a logistic GLM (Eq 1), we used a stochastic leaky integrate and fire neuron model (in discrete time). In this model, Vi, t = (γVi, t−1+(1−γ)Ui, t+εi, t)𝓘[Si, t−1 = 0] (U defined in Eq 2), Si, t+1 = 𝓘[Vi, t > 0.5]. We used εi, t ∼ 𝓝(0,1) as a white noise source. Also, we set γ = 20ms−1, similar to the inverse of the membrane’s voltage average integration timescale [60]. The weights were estimated up to a global multiplicative constant (resulting from the model mismatch), which was adjusted for in the figure. We conclude that our estimation method is robust to modeling errors, except perhaps the diagonal weights—their magnitudes were somewhat over-estimated due to the reset mechanism (which effectively increases self inhibition).