Statistical models of synaptic transmission evaluated using the expectation-maximization algorithm.

Amplitude fluctuations of evoked synaptic responses can be used to extract information on the probabilities of release at the active sites, and on the amplitudes of the synaptic responses generated by transmission at each active site. The parameters that describe this process must be obtained from an incomplete data set represented by the probability density of the evoked synaptic response. In this paper, the equations required to calculate these parameters using the Expectation-Maximization algorithm and the maximum likelihood criterion have been derived for a variety of statistical models of synaptic transmission. These models are ones where the probabilities associated with the different discrete amplitudes in the evoked responses are a) unconstrained, b) binomial, and c) compound binomial. The discrete amplitudes may be separated by equal (quantal) or unequal amounts, with or without quantal variance. Alternative models have been considered where the variance associated with the discrete amplitudes is sufficiently large such that no quantal amplitudes can be detected. These models involve the sum of a normal distribution (to represent failures) and a unimodal distribution (to represent the evoked responses). The implementation of the algorithm is described in each case, and its accuracy and convergence have been demonstrated.