Figure 3: Multimodal physiological measurements can be reconstructed from a shared latent manifold.

A. The delay embedding framework extended to model multimodal observables. We used the same encoder previously trained to predict calcium images. We coupled this encoder to decoders trained separately for FAD and hemoglobin. B. Example reconstructions of neuronal calcium (jRGECO fluorescence), metabolism (FAD), and blood oxygen (concentration of oxygenated hemoglobin). C. Reconstruction performance for each observable (as in Fig. 2A). See Fig. 2 for results with calcium.