0018-9162/03/$17.00 © 2003 IEEE January 2003 69

C O V E R F E A T U R E

P u b l i s h e d b y t h e I E E E C o m p u t e r S o c i e t y

Computer Electronics Meet Animal Brains

U ntil recently, neurobiologists have usedcomputers for simulation, data collection,and data analysis, but not to interactdirectly with nerve tissue in live, behavinganimals. Although digital computers and

nerve tissue both use voltage waveforms to trans-mit and process information, engineers and neu-robiologists have yet to cohesively link the elec-tronic signaling of digital computers with theelectronic signaling of nerve tissue in freely behav-ing animals.

Recent advances in microelectromechanical sys-tems (MEMS), CMOS electronics, and embeddedcomputer systems will finally let us link computercircuitry to neural cells in live animals and, in par-ticular, to reidentifiable cells with specific, knownneural functions. The key components of such abrain-computer system include neural probes, ana-log electronics, and a miniature microcomputer.Researchers developing neural probes such as sub-micron MEMS probes, microclamps, microprobearrays, and similar structures can now penetrateand make electrical contact with nerve cells with-out causing significant or long-term damage toprobes or cells.

Researchers developing analog electronics suchas low-power amplifiers and analog-to-digital con-verters can now integrate these devices with micro-controllers on a single low-power CMOS die.Further, researchers developing embedded com-puter systems can now incorporate all the core cir-cuitry of a modern computer on a single silicon chipthat can run on miniscule power from a tiny watch

battery. In short, engineers have all the pieces theyneed to build truly autonomous implantable com-puter systems.

Until now, high signal-to-noise recording as wellas digital processing of real-time neuronal signalshave been possible only in constrained laboratoryexperiments. By combining MEMS probes withanalog electronics and modern CMOS computinginto self-contained, implantable microsystems,implantable computers will free neuroscientistsfrom the lab bench.

INTEGRATING SILICON AND NEUROBIOLOGYNeurons and neuronal networks decide, remem-

ber, modulate, and control an animal’s every sen-sation, thought, movement, and act. The intimatedetails of this network, including the dynamic prop-erties of individual neurons and neuron popula-tions, give a nervous system the power to control awide array of behavioral functions.

The goal of understanding these details motivatesmany workers in modern neurobiology. To makesignificant progress, these neurobiologists needmethods for recording the activity of single neuronsor neuron assemblies, for long timescales, at highfidelity, in animals that can interact freely with theirsensory world and express normal behavioralresponses.

Conventional techniquesNeurobiologists examine the activities of brain

cells tied to sensory inputs, integrative processes,and motor outputs to understand the neural basis

By enabling better study of animal behavior’s neural basis, implantable computers may revolutionize field biology and eventually lead to neuralprosthetics, hardware-based human-computer interfaces, and artificial systems that incorporate biological intelligence principles.

Chris DiorioJaideepMavooriUniversity of Washington

70 Computer

of animal behavior and intelligence. They alsoprobe the components of neuronal control circuitryto understand the plasticity and dynamics of con-trol. They want to know more about neuronaldynamics and networks, about synaptic interac-tions between neurons, and about the inextricablelinks between environmental stimuli and neuronalsignaling, behavior, and control.

To explore the details of this biological circuitry,neurobiologists use two classes of electrodes torecord and stimulate electrical signals in tissue1:

• intracellular micropipettes to impale or patch-clamp single cells for interrogation of the cell’sinternal workings, and

• extracellular wires or micromachined probesfor interrogating multisite patterns of extra-cellular neural signaling or electrical activityin muscles.

Neurobiologists use amplifiers and signal generatorsto stimulate and record to and from neurons throughthese electrodes, and signal-processing systems toanalyze the results. They have used these techniquesfor decades to accumulate a wealth of understand-ing about the nervous system.2 Unfortunately, todate, most of these experiments have been performedon slices of brain tissue or on restrained and immo-bilized animals, primarily because the electronicinstruments required to run the experiments occupythe better part of a lab bench.

This situation leaves neurobiologists with a nag-ging question: Are they measuring the animal’s nor-mal brain signals or something far different?Further, neurobiologists want to understand howanimal brains respond and react to environmentalstimuli. The only way to truly answer these ques-tions is to measure a brain’s neural signaling whilethe animal roams freely in its natural environment.

Salient objectivesThe solution to these problems lies in making the

test equipment so small that a scientist can implantit into or onto the animal, using materials andimplantation techniques that hurt neither computernor animal. Recent developments in MEMS, semi-conductor electronics, embedded systems, bio-compatible materials, and electronic packagingfinally allow neuroscientists and engineers to beginpackaging entire neurobiology experiments intohardware and firmware that occupy less space thana human fingernail.

Researchers call these bioembedded systems neu-rochips. Scientists from the University of Washing-

ton, Caltech, and Case Western Reserve Universityhave teamed to build these miniaturized implantableexperimental setups to explore the neural basis ofbehavior.3 This research effort has developed or is inthe process of developing the following:

• miniaturized silicon MEMS probes for record-ing from the insides of nerve cells;

• biocompatible coatings that protect theseprobes from protein fouling;

• a stand-alone implantable microcomputer thatrecords from and stimulates neurons, sensorypathways, or motor control pathways in anintact animal, using intracellular probes, extra-cellular probes, or wire electrodes;

• neurophysiological preparations and tech-niques for implanting microchips and wireelectrodes or MEMS probes into or onto ani-mals in a way that does not damage the probesor tissue;

• firmware that performs real-time biologyexperiments with implanted computers, usinganalytical models of the underlying biology;and

• software to study and interpret the experi-mental results, eventually leading to reverse-engineered studies of animal behavior.

As the “Neuroscience Application Examples”sidebar shows, the first neurochip experiments usesea slugs and moths in artificial environments, butbroad interest has already arisen for using im-plantable computers in many other animals.

DESIGNER NEUROCHIPSLike their benchtop experimental counterparts,

neurochips use amplifiers to boost low-voltagebiological signals, analog-to-digital converters(ADCs) to digitize these signals, microcomputersto process the signals, onboard memory to storethe signals, digital-to-analog converters (DACs)to stimulate nerves, and software to control theoverall experiment.

Figure 1 shows a neurochip’s basic elements. Thekey requirements are that the neurochip be smalland lightweight enough to fit inside or onto the ani-mal, have adequate signal fidelity for interactingwith the millivolt-level signals characteristic ofnerve tissue, and have sufficient processing powerto perform experiments of real scientific value.

The basic components of a neurochip are com-mercially available today. They include instrumen-tation amplifiers, ADCs/DACs, reconfigurablemicrocomputers, and high-density memory. For

January 2003 71

Currently we are implanting neurochips intwo types of animals to study their neuronalbehavior.

Wiring a sea slugBeneath a research vessel anchored in the

Puget Sound, two scientists clad in scuba gearhover over the bright orange sea slug shownin Figure A. From the outside, this slug lookslike any other. But this particular slug has abattery-powered microcomputer implantedin its brain and minuscule silicon needles com-municating with its neurons. The microcom-puter faithfully performs a biology experi-ment as the animal goes about its normalbehavior.

Meanwhile, the scientists videotape theslug’s feeding, fleeing, and social behaviorswhile measuring water currents and geomag-netic fields. Later, these scientists will studythe environmental measurements and elec-tronic recordings in an attempt to decode howthe slug’s brain patterns correlate with behav-ior. The anticipated outcome: groundbreak-ing findings in behavioral neurobiology.

Monitoring a moth’s flight controlsIn a small, dark zoology lab, a giant moth

performs an aerial ballet as it feeds from arobotically controlled artificial flower, un-aware that the flower’s movements are pro-grammed to test the moth’s flight dynamics.The ultra-high-speed infrared video recordertapes the moth’s every movement.

But the special part of this experiment isneither the flower nor the videotaping. It isthe tiny battery-powered microcomputerattached to the moth’s thorax that recordselectrical signals from the flight muscles andsense organs and stores this data in onboardmemory.

Suddenly, the moth appears to struggle tokeep up with the flower. But the flower’smovements haven’t changed. Rather, theonboard microcomputer has begun stimulat-ing the nerves that enervate the moth’s wingmuscles, adjusting the wing-stroke phase insubtle ways that will let scientists measure theimpulse response of the moth’s flight-controlloops.

These experiments and others like them willsoon be played out in biology laboratoriesacross the country, if the multidisciplinaryteam of computer engineers and zoologiststhat is developing these implantable comput-ers has its way.

Neuroscience Application Examples

Tritonia diomedea MEMS probe tip,amplifier Brain

Microcontroller,A/D, cache

BatteryMemory

TetherVisceralcavity

(1) (2)

Figure A. Monitoring a giant sea slug. (1) Tritonia diomedea—shown with its prey, anorange seapen, in the background—is typically 20 cm in length, has a readily accessi-ble brain with large and well-characterized neurons, and is extraordinarily tolerant ofsurgical insult. (2) Artist’s conception of a neural recording setup implanted inside Tri-tonia.

1 cm

(1) (2)

Figure B. Tracking a moth’s movements. (1) Manduca sexta is typically 4 cm in lengthwith a wingspan of about 12 cm; at 2.5 gm, it is one of the largest flying insects. It cancarry loads up to a gram and fly at speeds up to 30 miles per hour, flapping its wings 25times a second. (2) Artist’s conception of a neurochip attached to Manduca’s thorax.

Realrhododendrons

Computer-controlledartificial flower

Figure C. High-speed infrared video captures Manduca hovering and sipping nectarfrom a moving flower. Visually guided flight control lets the animal compensate forrapid changes in wind direction and the flower’s swaying movements.

72 Computer

example, a Programmable System-on-a-Chip fromCypress MicroSystems integrates a microprocessor,variable-gain amplifiers, an ADC, a memory con-troller, and a DAC into a single integrated circuit.

First-generation neurochips integrate one ormore ICs, passive elements such as capacitors, bat-teries, and I/O pads on small micro-PCBs. The pro-totype neurochip shown in Figure 2 used packagedICs and button cells, and occupied a 1 cm × 3 cmprinted-circuit board. The “production” version,due out of processing in early 2003, uses die-on-board technology and thin-film batteries, and is

ADC

CPU

DACAmplifier

Amplifier

Memory

Accelerometer

Infrared LED forsynchronizing

with video camera

Nerve+

musclesignals

_+

Upload/download

link toexternal

computer

Figure 1. Neurochip functional block diagram. Solid lines show required compo-nents, dashed lines show some optional components.

Figure 2. Prototypeneurochips. (a, b) Afirst-generation neu-rochip comprisingdifferential ampli-fiers and batterieson a micro-PCBattached to theManduca moth’sthorax. The animal’sexoskeletonprovides a simpleattach point withoutbiocompatibilityissues. Manuallyimplanted bipolarrecordingelectrodes connectto recording sites.(c) A tethered in-flight recording fromthe thoracic flightmusculature. (d) Asecond-generationneurochip prototyperecords from twonerve or musclefiber sites, storingthe signals inonboard nonvolatilememory.

(a) (b)

(c)

(d)

smaller than 1 square centimeter. Future-genera-tion neurochips will integrate all the electronicsonto a single silicon chip, and will likely be smallerthan 10 mm on a side.

ProbesBuilding the probes that let a neurochip eaves-

drop on the electrical signaling in a nerve bundle,group of neurons, or single neuron presents adaunting task. Benchtop experiments on con-strained animals typically use metallic needles—often made of stainless steel or tungsten—tocommunicate with nerve bundles, micromachinedsilicon probes to record from groups of neurons,4

or glass capillaries filled with a conductive ionicsolution to penetrate and record from the inside ofindividual neurons. In unconstrained animals, flex-ible metallic needles, attached to the animal withsurgical superglue, and micromachined siliconprobes still work. However, replicating the perfor-mance of glass capillaries in flying, swimming, wig-gling animals is a different story entirely.

Several centimeters long and quite fragile, the glasscapillaries that neurobiologists use to probe theinsides of nerve cells typically have tip diameterssmaller than 0.3 microns. They impale neurons evenmore fragile than the probes themselves. Neuro-biologists use micromanipulators to painstakinglyand precisely drive single probes into single neurons.Fortunately, MEMS technology offers a possiblealternative to these glass capillaries. As Figure 3shows, University of Washington researchers aredeveloping silicon MEMS probes and flexible inter-connect structures to mimic the performance of glasscapillaries in an implanted preparation.5 Researchershave already recorded intracellular signals with earlyprototypes, and development is ongoing.

GlymeResearchers seek to implant both probes and

neurochips inside an animal’s brain. Unfortunately,an animal’s immune system rapidly and indiscrim-inately encapsulates all foreign bodies with pro-teins, without regard for the research value ofimplanted probes and neurochips. The adsorbedproteins not only attenuate the recorded electricalsignals, but can also jeopardize the animal’s sur-vival by causing abnormal tissue growth.

Researchers at the University of Washington’sCenter for Engineered Biomaterials have developedplasma-deposited ether-terminated oligoethyleneglycol coatings that inhibit protein fouling, asFigure 4 shows.6 Preliminary research indicates thatthese glyme coatings can reduce the protein foul-

ing of probes and neurochips to levels acceptablefor week-long experiments.

The power struggleNeurochips can derive power from onboard bat-

teries, external radiofrequency sources, a wire

January 2003 73

Figure 3. Micromachined silicon probes, flexible interconnect structures, andsea slug surgery. (a) Released, flexible silicon devices ready for implantation; (b)a sharp microelectrode on a flexible polyamide support; (c) the implantation pro-cedure places the needle on the exposed brain of a sea slug and the silicon basewith the external wires tucks under the slug’s skin; and (d) the postsurgery seaslug with implanted device can move freely in the water tank.

Figure 4. A fluorescence microscope image of a patterned 1,500 µm × 1,500 µmprotein-resistant plasma polymerized tetraglyme (pp4G) pad on a silicon-dioxidesubstrate, with additional 200 micron × 200 micron gold pads on and around thePP4G pad, after incubation in a solution containing fluorescently labeledproteins. The silicon-dioxide and gold areas adsorb protein and appear light,while the pp4G-coated areas resist protein adsorption and appear dark.

(a) (b)

(d)(c)

74 Computer

tether, or the nerve tissue itself. The ultimate deci-sion on the power source depends on the nature ofthe experiments and the animal’s environment.Batteries are attractive because they avoid the anten-nas and charge pumps required to capture RFenergy, operate in all environments, do not restrictthe animal’s movement the way a tether does, andprovide much more power than tapping nerve cellsfor energy.

Batteries have a weight disadvantage, but thin-film technologies using LiCoO2/LiPON/Li andNi/KOH/Zn promise flexible rechargeable batter-ies with peak current densities greater than 12 mAper square centimeter for short-duration experi-ments, and lifetimes measured in days or longer atlow-current densities.7,8

Batteries are ideal for the two sample prepara-tions shown in the “Neuroscience ApplicationExamples” sidebar. The typical hawkmoth flighttime is less than 60 seconds. The 12 mA providedby a 200 mg, one-square-centimeter battery easilypowers a neurochip for this experiment’s duration.The sea slug trolling methodically along theseafloor lies at the opposite end of the spectrum,needing only a few milliamps of current to powera neurochip for a week. The slug can easily accom-modate a large battery in its visceral cavity, allow-ing extended untethered experiments.

Out of memory?Once implanted, an embedded neurochip must

read its experimental procedure from memory, runthe experiment, acquire the neural spike trains, thenstore the results in memory. As with all computersystems, memory size is an issue for neurochips.Fortunately, the electrical spike trains generated bynerve tissue have a stereotyped shape as shown inFigure 2c, suggesting that neurochips should com-press the neural waveforms before storing them inmemory.

Compressing the signals has two advantages.First, it effectively increases the onboard storagecapacity. Second, it decreases the frequency ofmemory writes, reducing power consumption.Even simple compression algorithms such as run-length encoding can achieve better than 10 to 1compression ratios on neural signals.

Custom algorithms that apply vector quantiza-tion, run-length encoding, and Huffman encodingto different parts of the neural waveform canachieve up to 1,000 to 1 compression ratios.9 Giventhe limited computing power of an implantablemicrocomputer, simpler is better when it comes tocompression, but even simple RLE offers hugepower and memory-size benefits.

A STIMULATING WORLDPassive neurochips that do nothing more than

record will provide neurobiologists with a wealthof data. But even now, with the first neurochipsbarely in production, neurobiologists are alreadycalling for designs that stimulate nerve tissue as wellas record from it. Active neurochips will allow stim-ulus-response experiments that test models of hownervous systems control behavior, such as how sen-sory inputs inform motor-circuit loops and the logicor model behind the response.

Indeed, the neurochip project’s long-term goal isto develop a hardware and software environment inwhich a neurobiologist conceives a stimulus-re-sponse experiment, encodes that experiment in soft-ware, downloads the experiment to an implantedneurochip, and recovers the data when the experi-ment concludes. Figure 5 shows a model of integra-tive biology in which neurochips play a key part.

W ith advances in integrated circuit process-ing will come ever more capable andpower-efficient embedded computers. The

simple neurochips of today will become the com-plex embedded systems of tomorrow, when embed-ding in this ultimate sense will mean computerelectronics embedded in nerve tissue.

Neurochip Technology Forecast

Even as the first miniature neurochips record neuronal action poten-tials, researchers at the University of Washington are testing stimulus par-adigms to evoke controlled muscular extension and contraction.

Rather than driving the muscles directly using high-resolution voltagestimulus waveforms generated by digital synthesis and a digital-to-analogconverter, they tried stimulating nerve bundles instead, using simple dig-ital waveforms directly. They derived pulse-width modulated signalsdirectly from logic gates, and drove these waveforms into the nerve bun-dles that enervate the muscles.

Early results show great promise, not only because the technique actu-ally worked, but because a microcontroller can easily generate digitalpulses, and the drive currents needed for nerve stimulation are up to 100times smaller than those needed to drive muscle tissue directly. This powersavings will allow functional stimulation by miniature neurochips.1

Next on the research agenda: statistical machine learning. Researchersalready plan to use smart algorithms, smart software, and smart chips tointeract dynamically with nerve tissue. They suspect that machine learn-ing can help them study the cause-and-effect relationships involved in thebehavior of sensory motor circuits. Beyond that, they won’t speculate,but the applications of this neurochip research to robotics, medical pros-thetics, and a host of other applications seem obvious.

Reference1. M.R. Enstrom et al., “Abdominal Ruddering and the Control of Flight in the

Hawkmoth,” to be published, 2003; www.sicb.org/meetings/2003/schedule/abstractdetails.php3?id=758.

Enabling neuroscientists to better understand theneural basis of behavior is reason enough to developsuch devices. The long-term promise—hinted at inthe “Neurochip Technology Forecast” sidebar—ismuch greater, however, perhaps leading one day toneural prosthetics, hardware-based human-com-puter interfaces, and artificial systems that incor-porate principles of biological intelligence. �

AcknowledgmentsWe thank our collaborators in this research,

including Mike Tu, Tom Daniel, Russell Wyeth,Nathalia Peixoto, Dennis Willows, Yael Hanein,Gary Holeman, Karl Böhriger, Eric Chu, BuddyRatner, Denice Denton, Vickie Pan, Udo Lang, andMatt Richardson. We also thank the Office ofNaval Research and the David and Lucille PackardFoundation for funding this research.

References1. T.G. Smith et al., eds., Voltage and Patch Clamping

with Microelectrodes, Lippincott, Williams &Wilkins, 1995.

2. E.R. Kandel, J.H. Schwartz, and T.M. Jessell, Prin-ciples of Neuroscience, 4th ed., McGraw-Hill, 2000.

3. S. Combes, “Bugs, Slugs, ’n’ Chips: DownloadingBiology,” Northwest Science & Technology, Spring2000, pp. 32-36.

4. M.D. Gingerich, J.A. Wiler, and K.D. Wise, “Use ofan Active Microelectrode Array for Stimulation andRecording in the Central Nervous System,” Engi-neering in Medicine and Biology, 1999, Biomedical

Eng. Soc., 1999, p. 471.5. Y. Hanein et al., “Towards MEMS Probes for Intra-

cellular Neuronal Recording,” Sensors Update, vol.10, no. 1, 2002, pp. 1-29.

6. Y. Hanein et al., “Micromachining of NonfoulingCoatings for Bio-MEMS Applications,” Sensors andActuators B: Chemical, vol. 81, no. 1, 2001, pp. 49-54.

7. J.B. Bates et al., “Thin-Film Lithium and Lithium-Ion Batteries,” Solid State Ionics, vol. 135, 2000, pp.33-45.

8. L.G. Salmon, R.A. Barksdale, and B.R. Beachem,“Development of Rechargeable Microbatteries forAutonomous MEMS Applications,” TechnicalDigest, 1998, pp. 338-341.

9. M. Richardson, Data Compression of Neural Sig-naling in an Implantable Microchip, tech. report 02-08-01, Dept. Computer Science, University ofWashington, 2001.

Chris Diorio is an associate professor of computerscience at the University of Washington. Hisresearch interests lie at the interface of computingand biology. Diorio received a PhD in electricalengineering from the California Institute of Tech-nology. Contact him at diorio@cs.washington.edu.

Jaideep Mavoori is a PhD candidate in electricalengineering at the University of Washington. Hisresearch interests include biological intelligenceusing biocomputer interfaces. Mavoori received anMS in electrical engineering from Clemson Uni-versity. Contact him at jaideep@washington.edu.

January 2003 75

Figure 5. Integrative biology using neurochips. A neurochip can splice into neuromuscular pathways to unravel the internal biological cir-cuitry. Neurochips armed with recording and stimulation channels will help neurobiologists understand the complex interactions among themoth’s various neural control systems.