Spacetime Autoencoders Using Local Causal States
This work addresses forecasting challenges in spatiotemporal systems, but it appears incremental as it builds on existing local causal state concepts by adding decoding and dynamics.
The authors tackled the problem of forecasting complex spatiotemporal systems by framing local causal states as spacetime autoencoders, introducing a stochastic decoding method that maps latent fields back to observables and leveraging Markovian properties for dynamics, resulting in a new forecasting approach.
Local causal states are latent representations that capture organized pattern and structure in complex spatiotemporal systems. We expand their functionality, framing them as spacetime autoencoders. Previously, they were only considered as maps from observable spacetime fields to latent local causal state fields. Here, we show that there is a stochastic decoding that maps back from the latent fields to observable fields. Furthermore, their Markovian properties define a stochastic dynamic in the latent space. Combined with stochastic decoding, this gives a new method for forecasting spacetime fields.