The November 13, 2016 Kaik?ura, New Zealand earthquake sequence is unusually complex with its mainshock involving at least 17 crustal faults within the Marlborough fault system at the southern end of the Hikurangi trench, along the Pacific-Australian plate boundary. Previously published coseismic slip models derived from seismological, geodetic, field and tsunami observations have major differences in the number of ruptured faults, their detailed geometries and dip angles, the rupture pathway and dynamics, and the contribution of the underlying subduction interface. Several studies have used aftershocks as independent constraints on these questions, but the resolution of the available data is very limited. In this study, we develop a deep learning method based on artificial intelligence image recognition and apply it to the Kaik?ura sequence from the time of the mainshock until the end of 2016. During the 48.5-day period, we locate 64,465 events, six times the number of 10,861 in the routine GeoNet catalog. Our catalog is also much more complete than all other existing catalogs, including one derived from template matching. We demonstrate that the aftershocks form continuous distributions beneath the segmented surface ruptures, signifying possible mainshock rupture paths both inland and offshore which may be connected, and which may even have propagated in tandem. Also, the seismicity is distributed more broadly than typical widths of off fault deformation, suggesting the involvement of additional unmapped, variably-oriented secondary faults within a complex shear zone.