Researchers extend predictive coding to exponential-family distributions, bridging FEP and neural reality
New math shows how brains could use nonlinear, heterogeneous neurons for perception...
The free-energy principle (FEP) proposes the brain performs variational Bayesian inference, but earlier implementations relied on Gaussian assumptions and Laplace approximations, limiting biological realism (e.g., no negative firing rates, linear responses). Kataoka and Doya generalize FEP to the exponential family of distributions (EFD), which includes Poisson, Bernoulli, and many others used in neuroscience. They prove FEP's correspondence with predictive coding (PC) holds up to the second cumulant of the posterior under EFD.
This generalization naturally introduces nonlinear input-output properties, heterogeneous neuron types, and non-negative firing rates—all missing from Gaussian-only models. The authors also derive local, Hebbian-like plasticity rules that can train the network without global error signals. The work strengthens the theoretical foundation for understanding perception as inference in biological brains, offering a bridge between high-level variational principles and low-level neural dynamics.
- Framework works for exponential-family distributions (e.g., Poisson, Bernoulli), not just Gaussians
- Enables nonlinear, heterogeneous neural responses and avoids negative firing rates
- Training uses biologically plausible local plasticity rules, no global signals needed
Why It Matters
Brings theoretical neuroscience closer to capturing real neural properties, potentially inspiring more brain-like AI architectures.