The ability to efficiently and reliably estimate the presence and magnitude of cortical synapses is critical to understanding how synaptic connectivity relates to neural computation, behaviour, and disease. Two-photon optogenetic stimulation paired with whole-cell electrophysiology allows for potential presynaptic partners to be tested for a given patched cell by measuring optogenetically evoked postsynaptic currents (PSCs). However, existing optogenetic mapping approaches fail to reach their potential for three critical reasons: (1) they rely on stimulating neurons one at a time, (2) they use prolonged interstimulus intervals to allow postsynaptic conditions to return to a baseline, and (3) they typically do not explicitly account for variability in photoactivability and synaptic transmission. Failing to account for such variability introduces systematic biases in connectivity estimates since postsynaptic responses may be incorrectly associated with cells that were stimulated but failed to spike. Here we develop a high-throughput synaptic mapping system that simultaneously addresses each of these limitations. We combine 3D holographic optogenetics with a temporal demixing neural network and model-based compressed sensing algorithm for rapid learning of synaptic connections. Our holographic system simultaneously stimulates ensembles of neurons, thereby testing for many synaptic partners per trial. Our computational methods provide two key advances. First, our temporal demixing network allows us to stimulate in quick succession with minimal information loss, without having to wait for postsynaptic conditions to return to a baseline before stimulating the next ensemble. Second, our model-based compressed sensing algorithm provides highly scalable parallel inference of synapses while accounting for variability in photoactivation and synaptic transmission. We validate our methods on biological data using paired patch-clamp experiments that provide ground truth presynaptic spiking and postsynaptic responses. Together, our mapping system infers synaptic connectivity an order of magnitude faster than previously possible, enabling large-scale in vivo experiments where mapping time is highly constrained.