Overview

Sim Bamford demonstrated the 'ReNaChip', developed at l'Istituto Superiore di Sanita' in Rome. This chip primarily contains a configurable array of "intimately"-mixed-signal components, designed with neuromorphic modelling and neural signal processing in mind. It was designed specifically to replace a cerebellar micro-circuit in a real-time closed-loop in-vivo system for classical conditioning, but it may have more general application. In addition, Mika Laiho and 'Hasler' presented the RASP FPAAs from Georgia-tech. The CyberOptic? FPAA Transducer (http://capocaccia.ethz.ch/capo/wiki/2011/cyber11) project was spawned from the latter discussion.

Regarding the 'Renachip':

Some aspects of the chip are described in this paper:

"Intimate mixing of analogue and digital signals in a field-programmable mixed-signal array with lopsided logic", Bamford SA, Giulioni M, IEEE Biomedical Circuits and Systems Conference (BIOCAS), 2010. (http://capocaccia.ethz.ch/capo/attachment/wiki/2011/fpma11/Bamford%20Giulioni%202010%20BIOCAS.pdf)

The overall aim of the project is described in this paper (where the role of the chip would be to replace some of the software components of the system).

“The Application of a Real-Time Rapid-Prototyping Environment for the Behavioral Rehabilitation of a Lost Brain Function in Rats”, Prückl R, Grünbacher E, Ortner R, Taub AH, Hogri R, Magal A, Segalis E, Zreik M, Nossenson N, Herreros I, Giovannucci A, Ofek Almog R, Bamford S, Marcus-Kalish M, Shacham Y, Verschure PFMJ, Messer H, Mintz M, Scharinger J, Silmon A, Guger C, IEEE Symposium Series in Computational Intelligence (SSCI), 2011.

During the workshop, 4 small example systems were prototyped to demonstrate some of the features of the device.

Example 1: Neuron membrane models.

Two neurons were created which received the same external spiking inputs and differed only in that one had an exponential leak whereas the other had a linear leak:

Exponential Leak:

Linear Leak:

Here is a visualisation of the result of auto-routing, laid over the SRAM switch matrix of which the chip is composed.

Example 2: Spike-Timing-Dependent Plasticity (STDP)

A pulse-triggered model of synaptic weight dynamics based on the relative timing of pre- and post-synaptic neuroal spikes (programmed as input).

Example 3: Recurrent attractor dynamics

9 neurons were programmed with fully recurrent connections. They are quiescent until strong input arrives. After the input is dropped the activity is self-sustaining, until it is quenched by incoming inhibition. This therefore demonstrate two stable states of the network. Three spiking outputs are shown at the bottom. The mismatch between them is apparent in the slow resetting of the traces. This mismatch was used to create a random distribution of axonal delays, to avoid synchronisation amongst the neurons of the network.

Example 4: Noise generator

A chain of amplifiers were configured to amplify random (thermal) noise; this may be useful as a source of stochastic events in a future design. Here below is shown different stages of the noise generator amplifier chain, for a sweep of different combinations of current in feedforward and feedback amplifiers at each stage:

Apparently random noise appears at certain sweet-spots in the sweep:

In other places in the sweep, a fundamental frequency becomes apparent:

In other places in the sweep, bursting dynamics emerge (the disruption in the middle is due to the reprogramming of the bias currents between trials):