Patch-clamp recordings are useful to investigate the electrophysiological dynamics of neurons at single synapse level with high time resolution. However, subthreshold membrane potentials (Vm) recorded from in-vivo animals include complex synaptic inputs from thousands of presynaptic neurons, and it is technically difficult to remove artifacts induced by respiration and blood vessel pulsation. The goal of this research is to quantify the amplitude of a synaptic input in in-vivo synaptic bombardments using deep learning, which is designed to recognize natural images with high accuracy. We recorded spontaneous Vm fluctuations from hippocampal CA1 pyramidal cells in anesthetized mice and a pseudo-ideal excitatory post synaptic potential (EPSP) from a CA1 pyramidal cell in a hippocampal slice. The in-vivo Vm fluctuations were randomly phase-shifted to distort the waveform of intrinsic synaptic inputs and were superimposed with a single EPSP with various amplitudes, then we yielded surrogate Vm images with and without EPSPs. We trained ResNet, a deep learning model, with this dataset to estimate the amplitudes of EPSPs embedded in Vm fluctuations. We succeeded in reducing the mean error of the prediction to a level of the standard deviation of spontaneous Vm fluctuations.