Fig. 2: Representation of the considered neuronal architectures.
From: Bioinspired spiking architecture enables energy constrained touch encoding
a Connectivity between primary tactile afferents and neural populations in the Cuneate Nucleus; b, c schematic representations of the proposed architectures for the considered spiking neural networks. Lines between layers depict unidirectional structural connectivity, flowing from left to right. In the SF-DIR architecture (b), the mechanoreceptor layer (L1) directly connects to the output layer (L2), with connections that can dynamically switch between excitatory and inhibitory states during training. In the SC-BIO architecture (c), the structural connectivity mirrors that of the early somatosensory system (see subpanel a), with mechanoreceptor-activated primary sensory neurons (L1) being exclusively excitatory and interneurons (L2) purely inhibitory. d–f Schematic representations of three biologically plausible variants of the SC-BIO architecture: d suppression of lateral inhibition via removal of the interneuron population; e replacement of feedforward inhibition with feedback inhibition from the output layer to the population of interneurons, forming a recurrent loop; f combined architecture incorporating both feedforward and feedback inhibitory connections.
