Abstract
We introduce a novel, probabilistic binary latent variable model to detectnoisy or approximate repeats of patterns in sparse binary data. The model isbased on the "Noisy-OR model" (Heckerman, 1990), used previously for diseaseand topic modelling. The model's capability is demonstrated by extractingstructure in recordings from retinal neurons, but it can be widely applied todiscover and model latent structure in noisy binary data. In the context ofspiking neural data, the task is to "explain" spikes of individual neurons interms of groups of neurons, "Cell Assemblies" (CAs), that often fire together,due to mutual interactions or other causes. The model infers sparse activity ina set of binary latent variables, each describing the activity of a cellassembly. When the latent variable of a cell assembly is active, it reduces theprobabilities of neurons belonging to this assembly to be inactive. Theconditional probability kernels of the latent components are learned from thedata in an expectation maximization scheme, involving inference of latentstates and parameter adjustments to the model. We thoroughly validate the modelon synthesized spike trains constructed to statistically resemble recordedretinal responses to white noise stimulus and natural movie stimulus in data.We also apply our model to spiking responses recorded in retinal ganglion cells(RGCs) during stimulation with a movie and discuss the found structure.