An Electro-optical Connectome Prototype for Eight Neuron

Representations in FPGA Technology

Lorenzo Ferrara, Alexey Petrushin and Axel Blau

Dept. of Neuroscience and Brain Technologies (NBT), Italian Institute of Technology (IIT), 16163 Genoa, Italy

Keywords: Brain-inspired Computation, Nervous System Emulation, Optical Connectome, Parallel Information Flow,

Structured Illumination, Replica-casting, Field-programmable Gate Arrays.

Abstract: In nature, interneural signaling is highly parallel and temporally precisely structured. It would require equal

parallelism and temporal accuracy to faithfully mimic neural communication in hardware representations.

Light-based communication schemes fulfil this prerequisite. We report on a prototype of an optical

connectome implementation for a neuromorphic system eventually consisting of eight neurons. The

platform is based on field-programmable gate arrays (FPGAs) that run neuron-specific response models.

Their axons are represented by light-emitting diodes (LEDs) with axonal arbors in the form of micro-

patterned transparencies. They distribute membrane voltage threshold crossings, which are represented by

light pulses, onto synapse-specific photodiodes of postsynaptic neurons. This contribution sketches out the

overall system design and discusses its prospective application in replicating the connectome of the

nematode C. elegans in the framework of the Si elegans project.

1 INTRODUCTION

Surprisingly, even simple biological neural

networks can outperform today’s fastest

computational systems in tasks such as pattern

recognition and locomotion control. Nervous

systems are complex, highly parallel information

processing architectures made of seemingly

imperfect and slow, yet exceptionally adaptive and

power-efficient components to carry out

sophisticated information processing functions.

However, despite the rapidly growing body of

knowledge on almost every aspect of neural

function, currently no computational model or

hardware emulation exists that is able to describe

or even reproduce the complete behavioural

repertoire of the nematode Caenorhabditis elegans,

an organism with one of the simplest known nervous

systems. C. elegans, a soil-dwelling worm with a life

span of a few days, 1 mm long and 80 µm in

diameter, is one of the five best characterized

organisms. It is multicellular and develops from a

fertilized egg to an adult worm similar to higher

organisms. The morphology, arrangement and

connectivity of each cell including its neurons

have been completely described

and are found to

be almost invariant across different individuals.

Initially, 6393 chemical synaptic connections, 890

electrical junctions, and 1410 neuromuscular

junctions were identified (White et al., 1986). Recent

revisions of the original electron microscopy datasets

suggest that these numbers may actually be higher.

All of this data including the connectome, the

detailed interconnectivity map of the 302 neurons

through synapses, is publicly available through the

Worm Atlas (Achacoso and Yamamoto, 1992;

Oshio et al., 2003; Varshney et al., 2011). Despite its

simplicity, the nervous system of C. elegans does

not only sustain vital body function, but generates

a rich variety of behavioural patterns in response

to internal and external stimuli. These include

associative and several forms of nonassociative

learning that persist over several hours (Hobert,

2003). Interestingly, many processes of learning and

memory in C. elegans were highly conserved across

different species during evolution, which

demonstrates that there are universal mechanisms

underlying learning and memory throughout the

animal kingdom (Lin and Rankin, 2010).

To replicate the parallel information processing

pathways in nervous systems as faithfully as

possible, an equally parallel information

Ferrara, L., Petrushin, A. and Blau, A..

An Electro-optical Connectome Prototype for Eight Neuron Representations in FPGA Technology.

In Proceedings of the 3rd International Congress on Neurotechnology, Electronics and Informatics (NEUROTECHNIX 2015), pages 127-132

ISBN: 978-989-758-161-8

127

transmission scheme would be required. However,

classical 2D interconnectivity designs are based on

serial data transmission protocols, which are prone

to temporal jitter when simultaneously distributing

signals to more than one target receiver. Parallel

wire-based approaches will likely encounter

interconnect bottlenecks upon their upscaling to the

simultaneous addressing of a high number of target

synapses (Cangellaris, 1998). The Si elegans project

therefore pursues the development and

implementation of a 3D electro-optical free-space

interconnectivity scheme for the parallel

transmission and precise temporal processing of

neuronal information. The general concept and its

variants have been sketched out and discussed in

previous publications (Petrushin et al., 2014;

Petrushin et al., 2015).

In this contribution, we report on the elements

and working principle of an 8-neuron prototype of a

static light-based connectome and its integration

with the FPGA representations of their pre- and

postsynaptic neurons.

2 OPTICAL EMITTER AND

LIGHT DISTRIBUTION

ELEMENTS

Membrane potential threshold crossings of nerve

cells, here represented by field-programmable gate

array (FPGA) boards, are transmitted from pre- to

postsynaptic neurons by an optical connectome.

When the neural response model in a presynaptic

FPGA reaches that threshold, it sets its axonal

output, one of its freely addressable I/O pins, to a

high state. This triggers the light source, thereby

initiating the optical communication. The light-

emitter module is composed of a printed circuit

board (PCB) with a light-emitting diode (LED) and

its driver, a collimator in front of the LED to reduce

the divergence of the beam, a transparent mask that

structures the projected light into permissive and

non-permissive pathways, a lens that focuses the

mask pattern on the target and a box to contain all of

these elements (

Figure

1).

An XLamp® XB-H LED was selected as the

light source. This LED has sufficient light intensity

(230 lm). Its emitted color matches the spectral

sensitivity of the photodetector. An LED was

preferred to a laser diode for its longer lifetime and

lower cost. The LED modulator is shown in Figure

2. It works as follows: the input voltage, V

, appears

at the non-inverting input of the operational

amplifier (op-amp) U1. U1 forms a feedback loop

that drives the transistor Q1 in such way that the

voltage at the current-limiting resistor R4 is equal to

. The input voltage can range from 0 to 1.8 V.

Figure 1: Light-emitter module.

At 0 V, the LED turns off. At 1.8 V, the LED turns

on and draws a maximum current of 0.82 A. The

trimmer, R1, is used for adjusting the light intensity.

Additional resistors and capacitors were added for

stability purposes.

Figure 2: LED modulator schematics.

The collimator is an ASMT-M015 component

(Avago). It is glued in front of the LED to collimate

the light to an angle of 15 °, thereby concentrating

the light and reducing optical losses.

The light-distribution mask is composed of a

transparency film, which is printed with black toner

from a laser printer and then cut to fit the box. The

mask is patterned to project the light only on the

target photodiodes of those postsynaptic FPGA

neurons that a presynaptic neuron establishes

connections with.

NeBICA 2015 - Symposium on Neuro-Bio-Inspired Computation and Architectures

128

The lenses are plastic replica of a plano-convex

lens with a diameter and focus of 25 mm. The

fabrication process (

Figure 3

) consists of creating a

polydimethylsiloxane (PDMS) mold of the original

lens with clay (A), filling the mold with a

photoresist (SU8, MicroChem; B and C) with the

desired optical characteristics (e.g., refractive index)

and curing it to give a replica-lens (D). This strategy

results in affordable lenses of good optical quality.

Figure 3: Replica-molding of lenses.

The housing for the optical elements is made from

acrylonitrile butadiene styrene (ABS) by resorting to

a 3D printing technique. The computer-aided design

(CAD) (Figure 4) features holes for attaching the

PCB, a hole for the fixation of the module mounting

column and its alignment, the correct focal distance

between the mask and the lens, a support structure

for the mask and an aperture for the lens.

Figure 4: CAD view of the housing of the light-emitter

components.

3 PHOTORECEIVER AND

SIGNAL CONDITIONING

ELEMENTS

The optical signals are converted into electrical

signals by a photodiode (Everlight, PD70-01C). This

opto-electrical transducer was selected for its small

footprint, good sensitivity, high efficiency in the

visible spectrum and its relatively small parasitic

capacitance. Once light strikes the active area of the

photodiode, a current flows from the cathode to the

anode. The generated current is on the order of

microamperes and needs to be amplified before it

can be processed by subsequent electronics. This

amplification is performed by a current-to-voltage

converter configuration called transimpedance

amplifier (Figure 5).

Figure 5: A) Transimpedance Amplifier Schematics. B)

Differential Amplifier Schematics.

The negative input of the amplifier, U1A, senses

the current generated by the photodiode. Voltage on

the output of the amplifier will be equal to the

photocurrent times the value of the feedback

resistor, R1. A small feedback capacitor, C1, is

necessary to maintain stability. The photodiode is

operated in photoconductive mode with a -15 V

reverse bias voltage. The advantage of such biasing

is an increase of the system’s dynamic range. As a

drawback, biasing will result in a higher dark

current. However, this increase in the dark current is

not critical for the application. The transimpedance

op-amp should have a large gain bandwidth, a low

input capacitance and a low bias current. In addition,

it should come in a dual op-amp surface mounted

device (SMD) package with a small footprint,

feature a rail-to-rail output and operate on a power

supply as low as 3.3 V. Among the available op-

amps, the model MCP662 (Microchip) was selected.

An Electro-optical Connectome Prototype for Eight Neuron Representations in FPGA Technology

129

The illumination pattern is defined by a mask

with black and transparent regions. Ideally, the black

areas of the mask should block the light completely.

However, in practice, some light passes through the

black areas of the mask, thereby creating some

background illumination. In order to separate the

signal from the background, one of the photodiodes

is used as a background reference sensor. This

photodiode generates a background level signal,

which is subtracted from the received signal. This

operation is performed by the differential amplifier,

U1B (Figure 5). The obtained differential signal is

compared to the reference voltage of 1.1 V (Figure 6).

Figure 6: Comparator schematics.

If the reference voltage is higher than the signal, the

output of the comparator, U2A, stays low. If the

reference voltage is lower than the differential

signal, the output of the comparator, U2A, changes

to high. The comparator output is connected to the

respective pin of the postsynaptic FPGA. Resistors

R9 and R10 were added to suppress any undesired

toggling of the output. The termination resistor, R5,

dampens any over- or undershoot to increase the

signal quality.

In this work, we installed a synaptic panel

composed of 8 synaptic boards (SBs), each carrying

the driving circuitry for 8 photodiodes (Figure 7).

The relative position of the photodiodes with

Figure 7: Synaptic panel with 8 synaptic boards. Each SB

controls 8 photodiodes. In the figure, only 4 SBs are fully

populated.

respect to each other equals their dimensions

(approximately 8 x 8 mm

). The pitch and the active

area of the photodiodes define the mask structure

because the projected light from an LED has to shine

on the correct postsynaptic target photodiodes, but

avoid those that do not participate in a particular

synaptic connection configuration.

4 INTEGRATION WITH THE

FPGA HARDWARE

INFRASTRUCTURE

To test the connection, the optical connectome was

integrated with remotely controlled FPGA boards.

Each board (Terasic, Altera DE4) drove a single

axonal LED as a pre-synaptic neuron and received

the output from one of the photodetectors on each

SB when acting as a postsynaptic neuron. A simple

connectome scheme is depicted in

Figure

Figure 8: Schematics of the neural connectome. For

simplicity, only 3 out of 8 neurons are depicted.

Eight light-emitter modules were aligned to shine

their light patterns onto the synaptic panel, each

carrying a different mask for a neuron-specific

projection scheme (Figure 9).

Figure 9: Eight light-emitter modules of an 8-neuron

communication connectome.

NeBICA 2015 - Symposium on Neuro-Bio-Inspired Computation and Architectures

130

In one validation experiment, four FPGAs were

activated to simulate the communication between

four neurons. Figure 10 demonstrates the projection

through two masks of two firing LEDs.

Figure 10: Test projection of two LED illumination

patterns driven by their respective pre-synaptic FPGAs.

The test was performed at different frequencies to

evaluate the delay between the triggering signal

from the FPGA and the response of the targeted

photodiode (yellow and red traces in Figure 11,

respectively). The response with a delay of 500 ns

and a maximum operation frequency of 200 kHz is

sufficiently accurate and fast to emulate biological

spiking.

Figure 11: PD response to an LED illumination. The

maximum modulation frequency that keeps a 50% duty

cycle is around 200 kHz.

5 CONCLUSIONS

A parallel electro-optical setup that faithfully

mimics neural communication and its timing has

been implemented. Although the hardware is

currently limited to the emulation of 8 neurons, it

demonstrates the proof of principle for the emulation

of more complex neural networks such as the

complete connectome of the nematode C. elegans

(302 neurons, ~8,000 connections). A platform

based on FPGAs has sufficiently extensive

computation capability for the emulation of very

complex neural response algorithms and network

connectivities. To date, proof-of-concept tests have

been performed to demonstrate the correct synaptic

connection through precise light addressing. The

setup will be employed to further investigate

phototaxis in C. elegans, where 8 neurons are

involved.

ACKNOWLEDGEMENTS

The Si elegans project 601215 is funded by the 7

Framework Programme (FP7) of the European

Union under FET Proactive, call ICT-2011.9.11:

Neuro-Bio-Inspired Systems (NBIS). Kind loans of

electronic equipment by the IIT robotics workshop

team were very much appreciated. Many thanks to

our collaboration partners Martin McGinnity, Pedro

Machado, Alicia Costalago Meruelo and Kofi

Appiah for their help and fruitful discussions. We

are very grateful for FPGA board donations by

Altera.

REFERENCES

Achacoso, & Yamamoto. (1992). AY's Neuroanatomy of

C. elegans for Computation (pp. 304): CRS Press,

Boca Raton, FL.

Cangellaris. (1998). The interconnect bottleneck in multi-

GHz processors; new opportunities for hybrid

electrical/optical solutions. Fifth International

Conference on Massively Parallel Processing, Las

Vegas, Nevada, USA.

Hobert. (2003). Behavioral plasticity in C. elegans:

Paradigms, circuits, genes. Journal of Neurobiology,

54(1), 203-223.

Lin, & Rankin. (2010). Nematode Learning and Memory:

Neuroethology. In Breed & Moore (Eds.),

Encyclopedia of Animal Behavior (pp. 520-526).

Oxford: Academic Press.

Oshio, Iwasaki, Morita, et al. (2003). Database of

Synaptic Connectivity of C. elegans for Computation.

Retrieved from http://ims.dse.ibaraki.ac.jp/ccep/.

Petrushin, Ferrara, & Blau. (2015). The Si elegans

connectome: A neuromimetic emulation of neural

signal transfer with DMD-structured light. SPIE

Photonics West, Emerging Digital Micromirror

Device Based Systems and Applications VII, Session

6: Novel and Emerging DMD Applications, San

Francisco, USA.

Petrushin, Ferrara, Liberale, et al. (2014). Towards an

electro-optical emulation of the C. elegans

connectome. 2nd International Congress on

Neurotechnology, Electronics and Informatics, Rome,

Italy.

An Electro-optical Connectome Prototype for Eight Neuron Representations in FPGA Technology

131

Varshney, Chen, Paniagua, et al. (2011). Structural

Properties of the Caenorhabditis elegans Neuronal

Network. PLoS Computational Biology, 7(2),

e1001066.

White, Southgate, Thomson, et al. (1986). The Structure

of the Nervous System of the Nematode

Caenorhabditis elegans. Philosophical Transactions

of the Royal Society of London. B, Biological Sciences,

314(1165), 1-340.

NeBICA 2015 - Symposium on Neuro-Bio-Inspired Computation and Architectures

132