Dynamics of neural motifs realized with a minimal memristive neurosynaptic unit
Résumé
Spiking neural networks (SNNs) require both neurons and synapses. Here we introduce a synaptic unit with both excitatory and inhibitory biomimetic dynamics (Rall’s function). Combined with our previously introduced memristive spiking neuron, they complete a minimal model platform to build general SNNs. Our hardware neurosynaptic (NS) unit is fully analogue and implements a textbook theoretical model. It avoids the mismatch problem and is easily tunable at biocompatible timescales. We demonstrate the modularity of our NS-unit platform by building various basic neurocomputing primitives produced by basic motifs and their combination. With our novel hybrid theoretical and experimental approach we make three notable contributions: (1) We demonstrate that recursive inhibition (adaptation) and excitation can both be formulated as self-consistent dynamical problems. (2) We demonstrate that a minimal dynamical memory can be implemented with just one recursive single neuron. (3) We show that adopting a biomimetic membrane current in adaptive neuron models reveals an unknown connection to a cornerstone of physics, the harmonic oscillator model. Our NS unit realizes a building block for midsize SNNs of arbitrary geometry. Its simple design, along with the wide accessibility of ordinary electronic components, make our methodology an attractive platform to build neural interfaces for biological neuroscience, medical devices including deep brain stimulation, robotics, and artificial intelligence systems such as reservoir computing. We provide a bill of materials and printed-circuit-board designs to implement the device.
Domaines
| Origine | Fichiers éditeurs autorisés sur une archive ouverte |
|---|---|
| Licence |