Table 6.

Proposed correspondences between turbo coding in artificial neural networks and biological neural dynamics.

Turbo codes in artificial neural networks	Proposed correspondences in brains
Take data and distribute bits over two encoder–decoder networks.	Each sensory modality can be modeled as a noisy channel.
Generate a posterior probability estimate of the signal in one of the networks.	Within modalities, bottom-up updated states of deeper hierarchical levels calculate local posteriors (possibly taking the form of locally synchronized fast beta complexes).
Take the posterior from this network and propagate that belief as a prior to inform the calculation of a joint posterior for the other network.	Between modalities, auto-associative linkages across deeper hierarchical levels allow posteriors to be shared as empirical priors (possibly taking the form of larger and slower beta complexes).
Pass this message back to the original network as priors to inform the calculation of a new posterior.	Modalities are likely to be reciprocally connected, particularly in proximity to association cortices.
Repeat steps 3 and 4 until loopy belief propagation converges.	The formation of cross-modal synchronized complexes (at slower beta, alpha, and theta) frequencies may entail loopy message passing across modalities via self-organizing harmonic modes (SOHMs).
Result: Highly reliable data transmission even under highly noisy circumstances.	Result: Highly reliable perceptual inferences from noisy and ambiguous sensory information.