Motivated by the efficiency of biological systems in processing information, we develop biologically inspired, spiking, models for edge orientation selectivity, with computational primitives that are implementable on low-power subthreshold neuromorphic hardware. We use the iCub capacitive skin as front-end, stimulated by pressing bars at different orientations. Its analog output is encoded with spike trains, to mimic the slow adaptive mechanoreceptors in the human skin. Layers of Leaky and Integrate and Fire neurons endowed with global inhibition, Spike-Driven Synaptic Plasticity and homeostasis can learn the orientation of the stimuli. If the receptive fields of the tactile neurons are interleaved, the system shows hyperacuity, with respect to the expected resolution from the placement of the sensing areas.
The receptive fields created by randomly selecting sensitive points perform better than structured receptive fields with uniform distribution in discriminating small angles (down to 5°) and short bars.
Unsupervised spike-driven learning exploits the temporal coincidence of spatio-temporal spikes to learn a suitable connectivity pattern for edge orientation selectivity, if coupled with Winner-Takes-All competition and homeostasis.
The system is fully implementable on subthreshold mixed-mode CMOS circuits.
SNN, unsupervised local learning, spike-driven synaptic plasticity, neuromorphic circuits, capacitive iCub’s skin.