We analyzed action-potential activity from populations of Tritonia neurons with independent component analysis (ICA). ICA automatically sorted action potentials, artifacts, and noise into separate channels. The method was not sensitive spike overlap or changing spike shape. Individual neurons were located on the detector array by reconstructing raw data with single independent components. Independent components were also used as classifiers, and agreement with human experts indicated that the dimensionality of population data could be reduced in a meaningful way with ICA. The infomax algorithm we used to determine the independent components only separated as many action potential trains and artifacts as there were detectors, which has implications for the design of multi-unit recording experiments.