Implicit neural representations (INRs) for time-varying volumetric data typically require dense sampling over spatiotemporal coordinates, where each observation corresponds to a single point in space and time. This coordinate-wise formulation necessitates extensive sampling during optimization, leading to high computational costs and inefficient use of temporal structure. This work revisits this design choice, demonstrating that dense spatiotemporal sampling is not necessary for learning time-varying fields.
Instead, we represent the data as a collection of spatially indexed time series and train INRs using sequence-level supervision at each spatial location, rather than coordinate-wise scalar samples. This reformulation eliminates the need for dense spatiotemporal sampling and allows learning from the full temporal evolution of each spatial location in a structured manner. We demonstrate that this representation is compatible with a range of existing INR architectures, consistently improving reconstruction quality while significantly reducing training costs.
Furthermore, we show that this formulation can be combined with mixture-of-experts architectures, and our MoE instantiation further enhances reconstruction quality compared to both the base reformulation and existing MoE-based INR methods, providing stronger capacity allocation under heterogeneous temporal dynamics.
Blogger's Review: This research redefines the application of implicit neural representations in time-varying data, offering new insights for efficient learning. Notably, it enhances reconstruction quality while reducing computational costs, which is significant for practical applications. Future explorations into combining this with more architectures could be promising.