Place cells fire when an animal occupies a specific location. Time cells fire in sequence during a temporal interval. The standard interpretation treats these as different computations — spatial mapping and temporal tracking — that happen to share hippocampal real estate. Yu, Wang, and Balasubramanian collapse this distinction.
Their CA3 model is a recurrent network trained as a predictive autoencoder: it receives partially occluded experience vectors and reconstructs what is missing. When the input contains spatial patterns, the network settles into stable attractor-like place fields. When the input contains temporal patterns, it produces sequentially broadening fields that recapitulate time cells. Mixed inputs produce a smooth transition between the two representations.
The network does not switch modes. It has no place-cell circuit and no time-cell circuit. The same weights, the same dynamics, the same objective function. What changes is the input statistics. Spatial inputs have rich positional structure and the network finds stable attractors. Temporal inputs have sequential structure and the network tracks elapsed intervals. The representation is not computed by the architecture. It is selected by the data.
The through-claim: what looks like functional specialization — two distinct neural populations doing two distinct jobs — can be a single computation whose output is shaped entirely by what it is asked to predict. The system does not need to know whether it is encoding space or time. It only needs to reconstruct what is missing from its current input, and the nature of what is missing determines the nature of the representation.
This reframes the question of how a brain region can do multiple things. It does not do multiple things. It does one thing — prediction under partial observation — and the multiplicity lives in the input, not the mechanism.