Visual prostheses are brain–computer interfaces that are implanted in early processing stages of the visual system of blind patients. In an effort to induce light sensations, visual prostheses inject, via arrays of stimulating electrodes, spatiotemporal trains of current pulses which excite the adjacent neural tissue. Human experiments with current state-of-the art retinal prostheses have revealed that, although visual percepts can be elicited by electrical stimulation, these percepts are not closely related to the spatial patterns of stimulation. One of the main reasons for this failure is that present methods of prosthetic stimulation result in non-specific activation of multiple retinal pathways. Recent evidence, however, suggests that the specificity of neural activation can be increased by manipulations of the spatiotemporal parameters of stimulation. Before these notions are evaluated in human experiments, which are subjective and prone to patient fatigue and frustration, it is imperative that they are assessed in animal models using cortical recordings. Toward this end, we have developed a computational method for analyzing the cortical multi-site local field potential (ms-LFP) evoked in response to electrical stimulation of a site presynaptic to where LFPs are recorded. This method applies a nonlinear decoding technique on the recorded ms-LFP signal to quantify the information transmitted downstream from the stimulation site. Validation of this method using an implant attached to the epiretinal surface of cats and ms-LFP recordings from layer 4 of cat primary visual cortex, demonstrates that the spatial origin, the duration and the amplitude of injected current pulses can all be decoded simultaneously from single-trial ms-LFP responses. Our findings indicate that the developed method is a highly sensitive probe for characterizing the efficacy of visual prosthetic stimulation.