This excerpt from

Neurocomputing.James Anderson and Edward Rosenfeld, editors.© 1989 The MIT Press.

is provided in screen-viewable form for personal use only by membersof MIT CogNet.

Unauthorized use or dissemination of this information is expresslyforbidden.

If you have any questions about this material, please contactcognetadmin@cognet.mit.edu.

(1987)Massimo A. Sivilotti , Michelle A. Mahowald , and Carver A. Mead

Real-time visual computations using analog CMOS processing arraysAdvanced Research in VLSI: Proceedings of the 1987 Stanford Conference, P. Losleben (Ed.),

Cambridge, MA: MIT Press, pp. 295- 312

Currently a number of groups in private industry and the universities are trying tomake devices to i~ piement neural networks . Some groups are using optical techniques(see paper 39); others are using VLSI . At present, there are design projects underwayto implement many of the models we have presented here: dynamical attractor networks

, associative networks , Boltzmann machines. For the last paper in our collection ,we have chosen a description of one of the most interesting and highly developedapproach es to using VLSI technology for realizing a neural network . We have doneso even though technology advances so quickly that any paper on hardware will be)

out of date by the time it is published .

A group at the California Institute of Technology , directed by Carver Mead, is

systematically trying to design VLSI devices to act as sensory systems. The circuitryis modeled as closely as possible on the structure of mammalian sensory systems. This

paper describes their progress in modeling the retina, but other work in this laboratoryis trying to model the cochlea, to be used as the first stage of an artificial auditory

system.

It is sometimes not appreciated what a complex information processing system is

contained in the retina of the higher animals . In the frog many cells are specialized to

detect objects of interest to frogs- most prominently , bugs. The visual world of the

frog must be nearly blank until it detects one of the few objects it was designed to see.

The retina of primates is less specialized. However , it is still a formidable information

processing device. Most cells in the retina respond most strongly to movingstimuli , and respond weakly or not at all to stationary objects . There is also stronglateral inhibition , which enhances contrast and provides some gain control . There are

mechanisms that shift the system from one class of receptors (rods , for night vision ) to

another class (cones, for day vision ) as the light level changes. The size of receptivefields of the retinal output cells changes with light intensity , so that at very low

intensities cells integrate light over a large spatial region and over a long period of

time, trading resolution for sensitivity . As intensities increase, the retina rewires itself

so as to give higher spatial resolution , but lower sensitivity . There are several different

classes of output cells in cats: somewhat linear cells with small receptive cells (X -cells),less linear cells with larger receptive fields ( Y -cells), and group of complex , highlynonlinear cells that look almost like a piece of specialized frog retina mixed into the.retina (W -cells).

What the cortical visual areas are receiving does not look anything like a point -forpoint

copy of light intensities ; rather , it has been highly preprocessed. This early

preprocessing is surely very effective at reducing the computational burden of higherlevels.

43Introduction

702Chapter43

Sivilotti , Mahowald, and Mead are trying to build devices that incorporate as muchof this peripheral computing power as they can. As one example, used by this groupas an important starting point in their work, most retinal cells respond very stronglyto motion. For retinal cells, changes in light intensity are the single most potent aspectof the image. If one is building a traditional computer vision system, motion could bedetected by looking for changes in intensity computed from a succession of stationaryimages. However, the computational problems become extremely difficult. It is a hard

correspondence problem to discover which points in the two different views of the

moving object are the same.The authors view the problem of building a neural network as a branch of analog

circuit design. This is certainly correct. The bulk of the paper is a description of theconstruction techniques that can be used to obtain a large dynamic range in the

receptor units, to obtain sensitivity to changes in light intensity, and the practicalproblems involved in putting a number of such devices on a single chip.

Figures 18 and 19 in their paper give an idea of their success in capturing some

aspects of retinal processing. Carver Mead shows a dramatic videotape of this demonstration at lectures. When the artificial retina is looking at a stationary set of vanes,

there is barely an image on the screen. When the vanes start to rotate, the retina

explodes into life, and its responses to the edges of the vanes become large, outliningthem clearly. Having higher layers analyze a dynamic image of this type is consider ablydifferent from having higher layers analyze a static image, and changes processingrequirements consider ably. The approach described in this paper has a great deal torecommend it if we want to make systems that are, like us (or so we flatter ourselvesto be), intelligent perceivers.

The high level associative networks discussed in many other places in this collection,if present in the brain, are operating with carefully and effectively preprocessed inputs.The highest levels can do only a few kinds of simple computation. In the brain mostof the work has been done for them by special-purpose hardware, tried and tested byevolution, at the lower levels. It is a good strategy for the brain to follow and will bea good strategy for artifical devices to use as well.

Sivilotti,

Real-time visual computations using analog CMOS processing arraysAdvanced Research in VLSI: Proceedings of the 1987 Stanford Conference, P. Losleben (Ed.),

Cambridge, MA : MIT Press, pp. 295- 312

Integration of photosensors and processing elements provides a mechanism to concurrently perform computations

previously intractable in real-time. We have used this approach to model biological early vision process es. A set of

VLSI "retina" chips have been fabricated, using large scale

analog circuits (over lOOK transistors in total). Analog processing provides sophisticated, compact functional elements,

and avoids some of the aliasing problems encountered inconventional sampled-data artificial vision systems.

Models for Retinal Computations

The retina perfonns the first step in visual processingand provides the data for all subsequent stages ofthe visual system. Although various species perfonnslightly different sets of retinal computations, there areseveral aspects of visual processing that are commonto many different organisms [6] . These ubiquitous features

include logarithmic compression of the incomingintensity at the detector level, and the extensive use oflateral and temporal inhibition in the retinal computations

. These functions are computed using smoothlyvarying (continuous) analog potentials, rather thanneuronal action potentials.

Several explanations have been proposed for whythe visual system perfonns these computations in theretina. As is often the case when investigating biological

systems, it is not possible to detennine thereason that the system adopted a particular strategy;rather, these systems are optimzied with respect tomultiple constraints [ 18] .

One particular set of advantages to retinal processing can be observed by assuming that the function of

the retina is the "neat packaging of infonnation " [ 2]

to be sent to higher visual areas. Retinal ganglioncells transmit information to the brain by propagatingaction potentials along their axons in the optic nerve.There are a finite number of discriminable signal levelscoming over the optic nerve due to intrinsic noise. Thevisual infonnation must be encoded in such a way thatthe full dynamic range of the neuron is used. Automatic

gain control mechanisms, such as logarithmiccompression of intensity and center-surround antagonistic

receptive fields allow the system to encodedetail over a large range of ambient light levels. Inaddition to the limited resolution of signal levels, thediscrete-time nature of the action potential limits thetemporal resolution of events. A large fraction of retinalprocessing is dedicated to extracting motion events. Ifretinal ganglion cells encoded simple intensity, thenany change in intensity would be encoded as a changein pulse rate. Even assuming that such an encodingallowed no statistical fluctuation in pulse interval, the

Integration of Photoreception and Processing

By their very two-dimensional nature, images constitutea high bandwidth interface with the real world. Themost powerful supercomputers are incapable of evenrudimentary analysis of static images. Real-time analysis

of motion information, requiring computationover several images, is completely infeasible. Yet biological

early vision process es are clearly able to performthese computations, by exploiting the inherent parallelism

of visual inputs in a truly concurrent fashion.Computations are spatially localized, and computingelements are replicated as required. Only significantinformation is transmitted along the optic nerve.

Artificial vision systems are further limited by theearly sampling performed by television camera front-ends. Because each point is sampled only every1/30 second, an object can easily move several pixelsbetween samples. Motion interpretation has thus beenconverted from a local problem to the much moredifficult co" espondence problem. In signal processingterms, high frequency information is irretrievably lostdue to aliasing.

Modeled on biological architectures, our approachis to spatially interleave integrated photo receptors andprocessing on a VLSI die [ 16] . To obtain a sufficientlyrich, yet compact, set of computing elements, analogmicropower CMOS circuits are used. The resultinghigh density permits complex systems to be built ,that demonstrate powerful collective behaviors [20] .Finally, by performing temporal operations on continuous

data, prior to sampling for transmission ofT-chip,susceptibility to aliasing is reduced.

(1987)Michelle A. Mahowald. and Carver A. MeadMassimo A.

time at which such a change had occurred could bedetermined only to the time between pulses. In signalprocessing terms, the derivative information wouldhave been aliased away by temporal sampling of theimage. For this reason, motion detecting ganglion cellsproduce action potentials that correspond to changesin intensity, rahter than intensity itself. In this way,a pulse burst corresponds to an important featuremoving over that particular place on the retina.Higher-level correlations among events can then bereconstructed without loss of information due to temporal

aliasing.

arrangement provides a larger output voltage swing,and sets the output voltage in the 1.0 to 2. 5 V rangebelow ~ d' making direct coupling to subsequentstages possible. The feedback current is generatedby mirroring the load current for the emitter of thephototransistor. If the feedback to the phototransistorbase is omitted (Figure lb ), the receptor is sensitive tolower light levels (by a factor of hre), but will saturateat bright levels, as the MOS loads leave subthreshold.These receptors operate at light levels comparable tothe useful range of cones in human retinas, and formthe basis for the RET10 chip discussed later.

A Simple Local Computation: Discrete-TimeDerivativeWe perceive motion when a point in an image displaysnon-zero spatial and temporal derivatives. In otherwords, an edge (space derivative) that is moving causesa change in brightness at that point in the image. Thus,a local time derivative is the simplest computationto highlight areas of an image that are moving. Because

this computation is purely local, no interpixelcommunication

'is required. The derivative can be approximated

by comparing the present photoreceptoroutput with some suitably delayed version of the

Fiaare 2 Photo~ ptor with discrete-time ditrerentiator.

704Chapter43

~~CIT'"Figure 1 Logarithmic photoreceptors.

Light-Level IndependenceA vision system intended for operation in an unconstrained

environment must include automatic gaincontrol (AGC) with respect to absolute ambient lightlevel. Taking the logarithm of the incident light intensity

is a simple local AGC mechanism. Receptors withlogarithmic response have the additional advantage ofproviding output voltages whose difference is proportional

to the contrast ratios within the image, which arethe perceptually important parameters.

An integrated photoreceptor with an output that islogarithmic over 5 orders of magnitude in light intensity

is shown in Figure la . Its operation is similarto that of one previously described [9] ; a large-areabipolar transistor is formed using the n-well for thebase and p-type diffusion as the emitter. The substrateforms the collector, and hence the device is operated ina common-collector configuration. The output voltageblases the gate of a p-channel MOS feedback transistoroperating in subthreshold. In this regime, the channelcurrent is exponential in the gate voltage with a slopeof about 1 decade per 100m V. If a second subthresholdtransistor is used for source degeneration, the slope canbe decreased to about 1 decade per 300 m V. This

output.A discrete-time derivative based on the circuit in

Figure 2 forms the core of the RET20 chip. Whenswitch 81 is closed, the amplifier forms a unity-gainfollower stage that stores the current state of the systemon capacitor C1. The switch is then opened, and anyevolution of the input away from the sampled value isamplified by the open-loop voltage gain of the (wide-

range) transconductance amplifier [ 21] .If switch 81 is implemented with a Ma S pass transistor, transient switching charge is injected onto C1 [ 15] .

To minimize this effect, a transconductance amplifierwas used for the switch (Figure 3). When the biascurrent in A2 is reduced to zero, the capacitor is effectively

isolated from the input . Less noise charge is

Fiaure 3 Transconductance amplifier used as low-noise switch.

Fiaure 5 RET30 tesselation.

Figwe 4 Discrete-time derivative: sample operation.

generated because the channel charge is symmetricallydivided between both branch es of the differential pair(which have identical operating points in the followerconfiguration), and because any capacitive clock feed-

through is decreased by the cascode connection of thedifferential pair. In addition, the clock need not berail-to-rail (hence decreasing dV/dt even further), andit can be single phase and unipolar. Figure 4 illustratesthe output of the circuit when presented with an asymmetrical

100 Hz triangle wave input (; clock frequency1000 Hz).

RET30: Continuous-Time Derivatives andLocal Space DerivativesBiological retinas contain horizontal cells that providelateral conductance and can be loosely thought ofas providing an average of the signal values in theneighborhood with which the local signal can be compared

. Inspired by this model, the RET30 (Figure 5)consists of an array of receptors, R, interconnectedby a hexagonal resistive network [ 11] . To providetemporal smoothing, a capacitor to ground is locatedat each junction of 6 neighboring horizontal resistors.

105Sivilotti, Mahowald, and Mead 1981

CIT'"

Each local processing element takes the differencebetween the potential of the horizontal network andthat of the receptor output, and drives the local potential

of the horizontal network toward the local receptoroutput potential. The "derivative"

computed is thedifference between the input signal and a spatiallyand temporally smoothed version of that signal. Thespatial part of the processing emphasizes areas in theimage containing the most information. The emphasiscorresponds to a discrete approximation to a Lapla-Clan operator applied to the image. The temporal partof the processing is a finite-gain, single time-constantdifferentiation.

Horizontal Resistors To construct a practical space-time derivative system, we must be able to create timeconstants of the same order as the time scale of motionevents, without using enormous area for capacitors.The horizontal network spreads the potential at onepoint outward through a resistive sheet. To keep thetime constant (th) of the spreading on the same scaleas others in the system, enormous resistor values arerequired, (lO I I_ lO I3 .a)- larger than the resistanceof any integrated device we can build. A circuit thatimplements a very high-value resistance in a control ledway is shown in Figure 6. V I and V 2 represent potentials

J'i of two neighboring locations in the network.The current 10 into the upper node is constrained to bethe same as that out of the lower node by acurrent -mirror arrangement. Hence, any current out of node 1must flow into node 2. By symmetry, the magnitude ofthis current must be zero when the voltages are equal,and will be monotonically related to their difference.However, it can never exceed 10. The I - V relationshipis shown in Figure 7. The limiting current 10 is set bythe current mirror input, and controls the value of tho

706Chapter

2

10

2

Fiaure 6 Horizontal resistor. Continun~ -time clifferenti Sltnr

As can be seen from the figure, the region of linearoperation of the horizontal circuits is about :t 100m V,corresponding to an illumination contrast ratio ofabout 2: 1 at the photoreceptors.

Time Differentiators The differentiator of the RET30chip is implemented using the same circuit as thatshown in Figure 3, except the A2 amplifier is notclocked. Because the t/J input on the second amplifiercontrols the maximum current that may flow into orout of the storage capacitor, it determines the rate atwhich the capacitor is charged, and hence the time-constant fr . The net current into the capacitor is of thesame form as that shown in Figure 7, for the samereasons. This circuit has unity gain at DC, and a gainat short times set by the open-circuit gain of the firstamplifier, as in the discrete-time case. Experimentaldata illustrating the operation of the continuous-timedif Terentiator are shown in Figure 8.

The advantage of this arrangement is that any inputoffset voltage of the first amplifier is not multiplied bythe gain of the amplifier in its effect on the outputvoltage.

It can be argued that the saturating characteristic ofall these circuits is desirable, as it prevents one extremeinput (or faulty circuit element) from paralyzing theentire network. Thus, even for these simple operations,

CMOS Design Frame for Scanning Arrays

Fundamental limits on the number of pads imposed byavailable VLSI packaging technologies, coupled withthe high area cost of dedicated wiring within theimaging array, necessitate a scalable communicationarchitecture. Time multiplexing of sampled data signals

requires a minimal number of pads, and simplifiesthe external system design by reducing the number ofexternal components needed to compensate for thedifferent electrical nature of the off-chip environment.

To facilitate experimentation with a wide variety ofprocessing core cells, a "design frame"

approach wasadopted, with a standard peripheral frame providingall the communication support functions, as well as thedata sampling mechanism. By separating the computation

and communication tasks, the responsibility of thecore cell designer is simplified, and consists of providing

, as the result of some computation, an analogvoltage to be sampled and transmitted. An additionalbenefit is the independence of the tiling topology fromthe computation topology. For example, our designframes support true hexagonal tiling (Figure 9), hexagonal

tiling using offset rectangles (Figure 10), and pure

Figure8

Figure 7 I -V characteristic.

many of the properties of collective circuits can be

preserved.As shown , maximum outputs will occur when high

contrast features move over the retina . If only timederivative information is desired, the horizontal network

is unnecessary, and can be disabled by setting the10 input current to zero. When enabled, the horizontalnetwork computes the extent to which the light receivedby an individual receptor differs from the average levelin its neighborhood . It is thus most sensitive to a point ,less to a comer , even less to an edge, and not at all toa uniform gradient . The system can be made to displaya sustained response to one of its preferred stimuli evenif that stimulus is stationary .

Figure 9 Hexagonal tiling with Boston geometry (HEXRET).

rectangular tiling; in all cases, the pixel stream is (offset)rectangular.

Considerable work has been done on charge-transfer

systems, particularly for CCD image arrays [5] . Recently, bipolar photo transducer arrays have received

some attention due to their more favorable saturationand antiblooming properties [ 7,1,3] . In general, thenon-CCD scanning techniques consist of switchingsome charge basin onto a column line, then switchingthat column's charge packet onto a global output line.Often, these switch es are implemented with speciallyfabricated low-threshold MOS devices [ 14] . The primary

source of fixed-pattern noise on the output isdue to inversion charge variation in the switchingtransistors [ 13,23] . Mechanisms to reduce this noiseinclude integration over the entire pixel time, and

sampling at some constant point during the pixel event.Our RE Txx chips have taken a somewhat different

approach to obtaining an acceptable signal-to-noiseratio (SNR). First, photocharge is integrated at each

pixel site. It is not then destructively dumped on theoutput line, but rather is stored on a local capacitor,which is nondestructively sampled using a single MOScharge- sense transitor. The current in this "bit line"

is sensed, eventually by an external amplifier. To minimize

signal propagation delays due to C(dV/dt) losses(i.e. charging/discharging the highly capacitive bit lines,output line, and output pad/off-chip wiring), current-

steering sensing is employed throughout.

Figure 12 Horizontal switch es.

Pixels are enabled on a row-by-row basis by aswitched pass transistor in series with the transductiontransistor (Figure 11). Two configurations are possible:

Figure Ila shows a conventional cascade arrangementthat minimizes the dependence of I on Vbus; Figure 11 bshows a configuration designed to maximize the linearrange of I with v, by operating M 2 in the ohmic regionto provide source degeneration for MI . In either case,linear operation can be guaranteed if Vbus is maintainedsufficiently close to ground, so M 1 is operating in itsohmic regime.

One entire row of the scanning array is enabledsimultaneously, at the line clock rate tf>y. Within eachscan line, the pixel clock t/JH sequences the connectionof the different "bit" lines onto the single output line(Figure 12). To maintain control of ~ us' and to keepthis value independent of the pixel current I , the N - 1bit lines not connected to the output bus are insteadconnected to a dummy bus, biased at the desired valueof Vbus (Figure 13). Thus, the pixel current I flows at alltimes.

OfT-chip, the current in the output line is convertedto a voltage by a current- sense amplifier (Figure 14).This configuration implicitly blases the output line at

707Sivilotti, Mahowald, and Mead 1987

Vbua Vbua

I I

t / >enable

V

2 1

t / >enable

V

1 2

. .

(a

) (b

)

FiIDfe 11 Current transducers.

H.. ~ H..+l

v..+l ..c/>vt

v..

Hexagonal tiling with Manhattan geometry (RET30).Figure 10

q , B

H

. .

- - - .

H"

+ l

Output

Dummy

1

"

bit

"

line

e

ad line

From

To

previous next

stage

stage

Figure 13

~

~ Vbua : : : : :

. .

r ~RT. multiplexer.

mentary devices depends on the availability of a clockand its logical complement with minimal skew betweenthese signals. In the presence of appreciable skew,the transmission gate switch actually can be noisierthan a single pass transistor switch. For this reason,the clock signal and its complement are generatedsimultaneously by a CMOS set- reset logic (CSRL)[ 12] shift register, in which both signals are propagatedtogether. This shift register is clocked with a two-phasenonoverlapping clock running at the pixel rate. Because

the CSRL outputs are not fully restored duringone clock phase, they are buffered by a pair of inverters.The full circuit, including multiplexing switch es, isshown in Figure 15. The input to the shift register chainis brought off-chip. During one line period, a single

" 1"

is shifted into the register, and is shifted sequentiallythrough all the pixel columns.

The same CSRL design is used in the vertical shiftregister to sequentially enable successive rows. Thebuffer inverters serte the additional function of drivingthe highly capacitive

"word line" (; enable in Figure 11).

A functional diagram of the entire system is shownin Figure 16. An objective of the design frame was tosimplify the system-level interface with the pixel array.There are only eight signal lines (four clock S- ; " I' ; "2'; Hl' ; 82; two shift register inputs- Hi D' "in; and twoanalog outputs- lout, ldummyload) plus power and anysignals specific to the core cells in the pixel array.

Implementation and Experimental Results

We fabricated three different chips, incoporating theRET10, RET20, and RET30 core cells. Pixel array sizesand chip dimensions are shown in Table 1. The first

708Chapter43

Dummy load line.

Fig. e 15 stage with

Fiaure 14 External current-sense amplifier.

Vbus. External compensation in the form of Ccomp isrequired to counteract the highly capacitive input seenby the opamp.

The last circuit elements to consider in the analogsignal pat~ are the multiplexing switch es in Figure 13.These can inject noise charge onto the output lines intwo ways: (1) as a MOS transistor shuts off, the mobilecharge in the channel region is divided (nearly equally)between the source and the drain, and (2) the switchingclock itself can couple through the overlap capacitanceof the gate with the source-drain regions. To minimizethese problems, CMOS transmission gates driven bycomplementary phases derived from the horizontalclock were used as switch es. Any injected parasitic

charge is offset by the opposite physical processoccurring in the complementary transistor, as well asby the same process occuring in reverse in an adjacentpass gate (as that column becomes enabled). More complicated

charge compensation schemes are possible[8,19] , but our simulation results indicated that theseschemes did not significantly improve the clock noisesuppression in this case.

Compensation of charge injection using comple-

Sivilotti, Mahowald, and Mead 1987

�

V ; n

tPv1

Vertical

Shift

Register

utput

versions were fabricated on MOSIS [4] run M57Q in

August 1985. Except for the HEXRET, they all sharethe same design frame, implemented in Manhattan

geometry using the WOL design tool set [ 10,17] ;the HEXRET implemented the RET10 circuit usingarbitrary angle geometry, and was layed out with then G G E Rj P OO H design system [22] .

The chips were tested with two different setups: (1)a work station based tester capable of ~ ing all the

chip clocks, and of digitizing and displaying the outputsignal at a rate of 60,000 pixelsjsec (approximately8 framesjsec for RET10), and (2) a lTL clock generator

board that also produced a raster on which the

chip output could be superposed and displayed on an

oscilloscope at rates of more than 400,000 pixelsjsec(over 50 framesjsec).

Figure 17 shows sample output from the RET10chip, and clearly illustrates the logarithmic compression

performed by the photoreceptors. Figures 18 and19 demonstrate the operation of the RET30 chip. Inboth cases, the stimulus is a dark cross mounted ona rotating axis. In Figure 18 the cross is stationaryand, with the horizontal resistor network disabled, noimage is seen. In Figure 19 the cross is rotating at

approximately 10 rpm, and is clearly visible.

Figure 18 RET30- stationary image.

709

Fia- e 16

Table 1

Design frame: system level interface.

Fiaure 17 RETIO sample output.627060

Performance of core cells (3 pm feature size)

Chip�

33 x 33 N/A92xSO 88x88

113 x 98 64 x 64164 x 143 48 x 48

PhototransistorRETIOR EnoRET30�

Conclusions

Although many models have been proposed for thevisual system, it is not possible to simulate enoughcases to gain real confidence in the model, even onour most powerful computers. For this reason, wewill not really understand visual processing, especiallywith respect to demanding problems such as motionanalysis, until we succeed in building a system that iscapable of doing visual processing in real-time. Untilrecently, we have not had a technology in which suchfundamental synthetic investigations could be carriedout. With the evolution of high-density VLSI technology

, we have a means for conducting these extremelyimportant investigations. The work will not be trivial .Previously, the most massive application of large-scalecircuits has been in digital systems. Although analogintegrated circuit techniques have developed alongwith digital ones, no comparable methods exist formanaging the complexity of extremely large analogsystems. This paper not only has described a prototypevision system, but also has illustrated an approach toproblems of this class.

Acknowledgments

This research was supported by a grant from theSystem Development Foundation , and an equipmentgrant from Hewlett -Packard Corp . The authors are indebted

to Dick Lyon , Lars Neilsen, Michael Emerling ,and John Tanner for many useful discussions.

17 Massimo A. Sivilotti. A User's Guide to the WOL Design Tools.Technical Report 5237: TR: 86, California Institute of Technology,1986.

References

1 M. Aoki, H. Ando, S. Ohba, I. Takemoto, S. Nagahara, T. Nakano,M. Kubo, and T. Fujita. 2f3-inch format MOS single-chip color

710Chapter 43

Figure 19 RET30- moving image.

18 M. V. Srivivasan, S. B. Laughlin, and A. Dubs. Predictive coding:a fresh view of inhibition in the retina. Proc. R. Soc. London B,216:427-459,198219 RE. Suarez. P. R. Gray, and D. A. Hodges. All-MOS chargeredistribution analog-to-digital conversion techniques- part II.IEEE Journal of Solid State Circuits, SC-l0:379-385, December1975.

the Y -system of mammAlian retina. California Institute of Technology, Internal Technical Memo 5144:DF:84, June 1984.

System. Master's thesis, California Institute of Technology, 1986.5225:TR:86.

12 Carver Mead and John Wawrsynek. A new discipline for CMOSdesign: an architecture for sound synthesis. In 1985 Chapel HillConference on Very Large Scale Integration, pages 87- 104, 1985.

13 S. Ohba, M. Nakal, H. Ando, S. Hanamura, S. Shimada, K. Satoh,K. Takahashi, M. Kubo, and T. Fujita. MOS area sensor: partII - low-noise MOS area sensor with antiblooming photodiodes.IEEE Journal of Solid State Circuits, SC- 15(4):747- 752, August1980.

14 Yoshio Ohkubo. An analysis of fixed pattern noise for MOS-CCDtype image sensors under quasi-stationary conditions. IEEE J Oflrnalof Solid State Circuits, SC-21(4): 555- 560, August 1986.

15 Bing J. Sheu and Chenming Hu. Switch-induced error voltageon a switched capacitor. IEEE Journal of Solid State Circuits, SC-19(4):519- 525, August 1984.

16 Massimo A. Sivilotti. Toward a Motion-Based VLSI Vision

imager. IEEE Journal of Solid State Circuits, SC-17(2):375- 380,April 1982.

2 H. B. Barlow. Three points about lateral inhibition. In W. A.Rosenblith, editor, Sensory Communication, pages 782- 786, M.I.T.Press, Cambridge Mass., 1961.

3 Savvas G. Chamberlain and Jim P. Y. Lee. A novel wide dynamicrange silicon photodetector and linear imaging array. IEEE Journalof Solid State Circuits, SC-19(1):41- 48, February 1984.4 D. Cohen and G. Lewicki. MOSIS- the ARPA silicon broker. InProceedings from the Second Caltech Conference on VLSI, pages29- 44, california Institute of Technology, Pasadena, CA, 1981.5 P. LP . Dilion, D. M. Lewis, and F. G. Kasper. Color imagingsystem using a single CCD area array. I EEE TransElectron Devices,ED-25: 102- 107, February 1978.

6 Akimichi Kaneko. Physiology of the retina. Ann. Rev. Neurosci,2: 169- 191, 1979.

7 N. Koike, I. Takemoto, K. Satoh, S. Hanamura, S. Nagahara,and M. Kubo. MOS area sensor: part I- design consideration andperformance of an n-p-n structure 484 x 384 element color MOSimager. IEEE Journal of Solid State Circuits, SC-15(4):741- 746,August 1980.

8 J. L. McCreary and P. R. Gray. AU-MOS charge redistributionanalog-to-digital conversion techniques- part I. IEEE Journal ofSolid State Circuits, SC-I0:371- 379, December 1975.

9 Carver Mead. A sensitive electronic photoreceptor. In 1985 ChapelHill Conference on Very Large Scale Integration, pages 463- 471,1985.

10 Carver Mead. The WOwy . Technical Report 5113:TR:84,Ca1ifornia Institute of Technology, June 1984.

11 Carver Mead and Michelle Mahowald. An electronic model of

711Sivilotti, Mahowald, and Mead 1987

20 John Edward Tanner. Integrated Optical Motion Detection. PhDthesis, California Institute of Technology, 1986. 5223:TR:86.

21 Eric A. Vittoz. Micropower techniques. In YanDis Tsividis andPaolo Antognetti, editors, Design of MOS VLSI Circuits for Telecommunications

, pages 104- 144, Prentice-Hall, Englewood Cliffs,NJ, 1985.

22 Telle E. Whitney. Hierarchical Composition of VLSI Circuits.PhD thesis, California Institute of Technology, 1985. 5189:TR:85.

23 W. B. Wilson, H. Z. Massoud, E. J. Swanson, R. T. George Jr.,and R. B. Fair. Measurement and modeling of charge feed throughin n-channel MOS analog switch es. IEEE Journal of Solid StateCircuits, SC-20(6): 1206- 1213, December 1985.

This excerpt from

Neurocomputing.James Anderson and Edward Rosenfeld, editors.© 1989 The MIT Press.

is provided in screen-viewable form for personal use only by membersof MIT CogNet.

Unauthorized use or dissemination of this information is expresslyforbidden.

If you have any questions about this material, please contactcognetadmin@cognet.mit.edu.