We designed a VLSI chip to implement a neuroprosthesis which substituted synaptic plasticity subserving a learning function of the cerebellum using classical conditioning eye-blink protocol, using inputs from and outputs to brain stem nuclei. Recording electrodes were implanted in the pontine nucleus and the inferior olive of anesthetized rats, to detect activity elicited by a tone conditioned stimulus (CS) and a peri-orbital air-puff unconditioned stimulus (US) respectively. A stimulation electrode was placed in the facial nucleus such that an eyeblink conditioned response (CR) could be elicited by the CS. Amplified signals were passed to the custom VLSI chip (Bamford et al., 2012), which implemented signal processing pathways leading to detection of CS and US events, that converged on a model of cerebellar learning (Verschure and Mintz, 2001), in which timing of a conditioned response is established by interplay of LTP and LTD at parallel-fiber to Purkinje cell synapses. We applied repeated paired CS and US trials, where CS onset preceded US by several hundred ms and the two stimuli coterminated. During these trials, the efficacy of a neuromorphically emulated synapse reduced, which would induce a pause in the inhibitory activity of the Purkinje cell at a certain delay following CS onset. This caused a CR to be produced and subsequently its delay to be reduced with respect to the CS towards a physiologically appropriate timing. Prolonged exposure to CS-only trials resulted in the extinction of the conditioned response, consistently with behavioral and neurophysiological evidence.

Anesthetized rats cannot learn this time-adapted conditioned response (due to silencing of granule cells). Thus our results show that a learning function of the cerebellum can be substituted by a neuroprosthetic system. Notably, the VLSI chip, with its innovative field-programmable mixed-signal array design, acted as the substrate for the adaptation underlying learning of response timing. Our system stands apart from other neuroprostheses, based on black-box functional characterization, by its model-based approach; we believe it constitutes a promising pilot example of neuroprosthesis based on knowledge of the involved neural machinery, and biologically grounded models thereof.

This work is a follow-up of the EU project RENACHIP.