Brain-Like Variational Inference

• Normative approaches in neuroscience aim to understand how brain-like behavior and mechanisms emerge from simple objectives.

• One such objective is the minimization of variational free energy (F), equivalent to negative ELBO from machine learning (used for many models like VAEs).

• Building on this unification potential, we introduced FOND,  a framework for deriving brain-like inference algorithms from first principles. 

• We then applied the FOND framework to derive a family of iterative VAE models, including the spiking iterative Poisson VAE (iP-VAE).

• iP-VAE is a stochastic, spiking, sparse coding model with principled theoretical justification and strong empirical performance, including (for the first time) on deep decoders.

Available at https://openreview.net/pdf?id=573IcLusXq.