Details of derivation and implementation of the SPP model.

<p>First, we review the main idea behind predictive processing (PP) and how it is utilized here on the level of single synapses and related mathematical derivations. Then, we describe details about our main theoretical result showing that the synaptic efficacy updates minimize the free energy of the synaptic efficacy with respect to the back-propagating action potentials. Furthermore, we show that the same learning rules also emerge if the SPP is applied to a learning scenario for recurrent neural networks with arbitrary numbers of neurons and synapses.</p> <p>(PDF)</p>