Our understanding of galaxy evolution is derived from large surveys designed to maximize efficiency by only observing the minimum amount needed to infer properties for a typical galaxy. However, for a few percent of galaxies in every survey, these observations are insufficient and derived properties can be catastrophically wrong. Further, it is currently difficult or impossible to determine which objects have failed, so that these contaminate every study of galaxy properties. We develop a novel method to identify these objects by combining the astronomical codes that infer galaxy properties with the dimensionality reduction algorithm t-SNE, which groups similar objects to determine which inferred properties are out of place. This method provides an improvement for the COSMOS catalog, which already uses existing techniques for catastrophic error removal, and therefore should improve the quality of large catalogs and any studies that are sensitive to large redshift errors.