Abstract | We show that MID is a maximum-likelihood estimator for the parameters of a Iinear—nonlinear—Poisson (LNP) model, and that the empirical single-spike information corresponds to the normalized log-likelihood under a Poisson model. |
DKL<p(x|r = 0) p(x)) is the information (per spike) carried by silences, and | It is straightforward to show that empirical Bernoulli information equals the LNB model log-likelihood per spike plus a constant: |
Equivalence of MID and maximum-likelihood LNP | We will now unpack the details of this estimate in order to show its relationship to the LNP model log-likelihood . |
Equivalence of MID and maximum-likelihood LNP | This allows us to directly relate the empirical single-spike information (Equation 8) With the LNP model log-likelihood , normalized by the spike count as follows: |
Equivalence of MID and maximum-likelihood LNP | Where Llnpwo, D) denotes the Poisson log-likelihood under a “null” model in Which spike rate |
Introduction | This equivalence follows from the fact that the plugin estimate for the single-spike information [26], the quantity that MID optimizes, is equal to a normalized Poisson log-likelihood . |
Models with Bernoulli spiking | To show this, we derive the mutual information between the stimulus and a Bernoulli distributed spike count, and show that this quantity is closely related to the log-likelihood under a linear-nonlin-ear-Bernoulli encoding model. |
minimum information loss for binary spiking | Thus, the log-likelihood for projection matrix K, having already maximized With respect to the nonlinearities by using their plugin estimates, is |
minimum information loss for binary spiking | pluginComparison With the LNC model log-likelihood (Equation 30) reveals that: where H [r] = — 2133" logNVU) is the plug-in estimate for the marginal entropy of the ob-served spike countsobserved. |
NEMix inference | To do so, we approximate the marginal likelihood (10) by the eXpectation of the complete data log-likelihood |
NEMix inference | A derivation of the expected hidden log-likelihood and the maximum likelihood estimates is given in ‘Estimating the hidden signal’ of 81 Text. |
Simulation study | To assess the impact that pathway disruption has on the cell population level, we ran the simulations on a standard NEM using the log-likelihood model introduced in [23]. |
Supporting Information | The first column gives the log-likelihood for each model, showing that the true network is much less likely than the inferred networks. |
Parameter estimation | Minimization of Eq 4 is equivalent to maximizing the log-likelihood given by |
Parameter estimation | Model statistics With respect to the objective function, the log-likelihood and the AIC are summarised in 81 Table. |
Supporting Information | The values of the best fit and the average over 100 fits are given for the objective value (WRSS), the log-likelihood (ln(L( p ) )) and the Akaike Information Criterion (AIC), as defined in Materials and Methods. |