{"title": "An Analog VLSI Saccadic Eye Movement System", "book": "Advances in Neural Information Processing Systems", "page_first": 582, "page_last": 589, "abstract": null, "full_text": "An Analog VLSI  Saccadic Eye Movement \n\nSystem \n\nTimothy K.  Horiuchi \n\nBrooks  Bishofberger  and  Christof Koch \n\nComputation and Neural  Systems Program \n\nCalifornia Institute of Technology \n\nMS  139-74 \n\nPasadena, CA 91125 \n\nAbstract \n\nIn  an effort  to understand saccadic eye  movements and their rela(cid:173)\ntion  to visual  attention and other forms  of eye  movements,  we  -\nin  collaboration with  a  number of other laboratories -\nare  carry(cid:173)\ning out a  large-scale effort  to design  and build a  complete primate \noculomotor  system  using  analog  CMOS  VLSI  technology.  Using \nthis technology, a low power, compact, multi-chip system has been \nbuilt  which  works in  real-time  using  real-world  visual  inputs.  We \ndescribe  in  this paper  the  performance  of an early  version  of such \na  system  including  a  1-D  array of photoreceptors  mimicking  the \nretina, a circuit computing the mean location of activity represent(cid:173)\ning  the  superior  colliculus,  a  saccadic  burst  generator,  and  a  one \ndegree-of-freedom  rotational  platform  which  models  the  dynamic \nproperties  of the  primate oculomotor plant. \n\n1 \n\nIntroduction \n\nWhen  we  look  around  our environment, we  move  our  eyes  to  center  and stabilize \nobjects of interest onto our fovea.  In order to achieve this, our eyes  move in  quick \njumps with short  pauses in between.  These  quick jumps (up  to 750  deg/sec in hu(cid:173)\nmans)  are known as saccades and are seen  in both exploratory eye movements and \nas  reflexive  eye  movements  in  response  to sudden  visual,  auditory,  or  somatosen(cid:173)\nsory stimuli.  Since  the intent of the saccade is to bring new  objects of interest onto \nthe  fovea,  it  can  be  considered  a  primitive  attentional  mechanism.  Our  interest \n\n582 \n\n\fAn Analog VLSI Saccadic Eye Movement System \n\n583 \n\nlies  in  understanding  how saccades  are  directed  and how  they  might interact with \nhigher  attentional processes.  To pursue  this goal,  we  are  designing  and building a \nclosed-loop hardware system based on current models of the  saccadic system. \n\nUsing  traditional  software  methods  to  model  neural  systems  is  difficult  because \nneural  systems  are  composed  of large  numbers  of elements  with  non-linear  char(cid:173)\nacteristics  and  a  wide  range  of time-constants.  Their  mathematical  behavior  can \nrarely  be  solved analytically and simulations slow  dramatically as  the  number and \ncoupling  of elements  increases.  Thus,  real-time  behavior,  a  critical  issue  for  any \nsystem  evolved for  survival  in  a  rapidly changing world,  becomes  impossible.  Our \napproach  to  these  problems  has  been  to  fabricate  special  purpose  hardware  that \nreflects  the  organization of real  neural  systems  (Mead,  1989;  Mahowald  and  Dou(cid:173)\nglas,  1991; Horiuchi  et  al.,  1992.)  Neuromorphic analog VLSI technology has many \nfeatures  in common with nervous  tissue such as:  processing  strategies that are fast \nand reliable, circuits  that are robust against noise  and component variability, local \nparameter storage for the construction of adaptive systems and low-power consump(cid:173)\ntion.  Our  analog chips  and the nervous system both use  low-accuracy components \nand  are significantly constrained  by wiring. \n\nThe design of the analog VLSI saccadic system discussed  here is part of a long-term \neffort  of a  number of laboratories  (  Douglas  and  Mahowald  at  Oxford  University, \nClark at Harvard University, Sejnowski at UCSD and the Salk Institute,  Mead and \nKoch  at  Caltech)  to  design  and  build  a  complete  replica  of the  early  mammalian \nvisual system in  analog CMOS  VLSI. \n\nThe  design  and  fabrication  of  all  circuits  is  carried  out  via  the  US-government \nsponsored silicon service  MOSIS,  using  their 2 J.1.m  line process. \n\n2  An  Analog  VLSI Saccadic System \n\nFigure  1:  Diagram of the current  system. \n\nThe system obtains visual  inputs from  a  photoreceptor  array, computes the  target \nlocation  within  a  model  of the  superior  colliculus  and  outputs  the  saccadic  burst \ncommand to drive  the  eyeball.  While  not  discussed  here,  an  auditory  localization \n\n\fS84 \n\nHoriuchi, Bishofberger, and Koch \n\ninput is  being  developed to trigger  saccades  to acoustical stimuli. \n\n2.1  The Oculomotor Plant \n\nThe oculomotor plant is a one degree-of-freedom turntable which is driven by a pair \nof antagonistic-pulling motors.  In  the  biological system  where  the  agonist  muscle \npulls  against  a  passive  viscoelastic  force,  the  fixation  position  is  set  by  balancing \nthese  two forces.  In our system, the viscoelastic  properties of the oculomotor plant \nare simulated electronically and the fixation point is set  by the shifting equilibrium \npoint of these forces.  In order maintain fixation off-center, like the biological system, \na  tonic signal  to the motor controller must be maintained. \n\n2.2  Photoreceptors \n\nThe  front-end  of the  system  is  an  adaptive  photoreceptor  array  (Delbriick,  1992) \nwhich amplifies small changes in light intensity yet adapts quickly to gross changes in \nlighting level.  The current system uses a  1-D array of 32 photoreceptors 40  microns \napart.  This  array  provides  the  visual  input  to  the  superior  colliculus  circuitry. \nThe gain  control  occurs  locally at  each  pixel of the  image  and  thus the  maximum \nsensitivity is maintained everywhere in the image in contrast to traditional imaging \narrays which  may  provide  washed  out  or  blacked-out areas  of an  image  when  the \ncontrast within  an  image is  too large.  In  order  to trigger  reflexive,  visually-guided \nsaccades,  the output of the photoreceptor array is coupled to the superior  colliculus \nmodel by a  luminance change detection  circuit.  A  change in luminance somewhere \nin  the  image sends  a  pulse  of current  to  the  colliculus circuit  which  computes  the \ncenter of this activity.  This coupling passes a current signal which is proportional to \nthe  absolute-value of the  temporal  derivative of a  photoreceptor's  voltage output, \n(i.e.  IIdI(x, t)/dtll where  I(x,t)  is  the output of the  photreceptor  array).  While we \nare  initially  building  a  1-D  system,  2-D  photoreceptor  arrays  have  been  built  in \nanticipation of a  two degree-of-freedom  system.  While these  photoreceptor  circuits \nhave  been  successfully  constructed,  we  do  not  discuss  the  results  here  since  the \nperformance of these  circuits  are  described  in the literature (Delbriick  1992). \n\n2.3  Superior Colliculus  Model \n\nThe superior colliculus,  located on the dorsal surface  of the midbrain, is  a key  area \nin  the  behavioral  orientation  system  of mammals.  The  superficial  layers  have  a \ntopographic  map  of visual  space  and  the  deeper  layers  contain  a  motor  map  of \nsaccadic  vectors.  Microstimulation  in  this  area  initiates  saccades  whose  metrics \nare  related  to  the  location  stimulated.  This  type  of representation  is  known  as  a \npopulation coding.  Many neurons  in  the  deeper  layers of superior  colliculus are \nmultisensory  and  will  generate  saccades  to  auditory  and somatosensory  targets as \nwell  as visual  targets. \n\nWhile  it  is  clear  that  the  superior  colliculus  performs  a  multitude  of integrative \nfunctions  between  sensory  modalities  and  attentional  processes,  our  initial  model \nof superior  colliculus  simply  computes  the  center  of activity  from  the  population \ncode seen  in  the superficial layers (i.e.  the  photoreceptor  array) using the weighted \naverage  techniques  developed  by  DeWeerth  (1991)  for  computing  the  centroid  of \n\n\fAn Analog VLSI Saccadic Eye Movement System \n\n585 \n\nCentroid Circuit Output vs. Target Error \n\n10 \n\n2.. \n\n2.6 \n\n2.0 \n\nI.' \n\n.\u00abJ \n\n~ V \n\nV\u00b7 \n\n/ \n./ \n\nr:r'/ \n\n.. / \n\n/ \n\nV \n\n.:J) \n\n\u00b7211 \n\n\u00b710 \n\no \n\n10 \n\n20 \n\nFigure  2:  Output  of the  centroid  circuit  for  a  flashed  red  LED  target  at  different \nangles  away  from  the  center  position.  Note  that  the  output  of the  circuit  was \nsampled  1 msec.  after stimulus onset  to account for  capacitive delays. \n\nbrightness.  The results  of the  photoreceptor/centroid circuits  are  shown in Fig.  2. \nIn the  case  of visually-guided saccades,  retinal error  translates directly  into motor \nerror and thus we  can use  the photoreceptors  directly as our inputs.  This simplified \nretina/superior  colliculus  model  provides  the  motor error  which is  then  passed  on \nto the  burst generator. \n\n2.4  Saccadic  Burst Generator \n\nThe  burst  generator  model  (Fig.  3)  driving  the  oculomotor  plant  receives  as  its \ninput, desired  change in eye position from the superior colliculus model and creates \na  two-component  signal,  a  pulse  and  a  step  (Fig.  4).  A  pair  of these  pulse/step \nsignals drive the two muscles of the eye which in turn moves the retinal array, thus \nclosing the loop.  The burst generator model is  a  double  integrator model  based on \nthe work by Jurgens,  et  al (1981) and Lisberger  et  al (1987) which uses initial motor \nerror as the input to the system.  This motor error is injected into the  \"integrating\" \nburst  neuron  which  has  negative  feedback  onto  itself.  This  arrangement  has  the \neffect  of firing  a  number of spikes proportional to the initial value of motor error.  In \nthe circuit, this integrator is implemented by a 1.9 pF capacitor.  This burst of spikes \nserves to drive the eye rapidly against the viscosity.  The burst is  also integrated by \nthe  \"neural integrator\"  (another  1.9 pF capacitor) which holds the local estimate of \nthe current eye position from which the  tonic,  or  holding signal is generated.  Figs. \n4  and  5  show  output  data from  the  burst  generator  chip  and  the  response  of the \nphysical mechanism to this output.  The inputs to the burst generator  chip are  1)  a \nvoltage indicating desired  eye  position and  2)  a  digital trigger signal.  The outputs \nare  a  pair of asynchronous  digital  pulse  trains  which  carry  the  pulse/step  signals \nwhich drive the left  and right  motors. \n\n\fS86 \n\nHoriuchi, Bishofberger, and Koch \n\n3  Discussion \n\nAs  we  are  still  in  the  formative  stages  of our  project,  our  first  goal  has  been  to \ndemonstrate a  closed-loop system which can fixate  a  particular stimulus whose  im(cid:173)\nage  is  falling onto its photoreceptor  array.  The first  set  of chips represent  dramat(cid:173)\nically simplified  circuits  in  order  to  capture  the first-order  behavior  of the system \nwhile using known  representations.  Owing to the large number of parameters that \nmust be set,  and their sensitivity to variations, we have begun a study to investigate \nbiologically plausible approaches to automatic parameter-setting. \n\nIn  the  short  term we  intend  to dramatically refine  the  models  used  at  each stage, \nmost notably the superior colliculus which is involved in the integration of non-visual \nsources  of saccade  targets  (e.g.  memory or  audition),  and in  the  mechanisms used \nfor  target  selection or fixation.  In the longer run,  we  plan to model  the interaction \nof this system with other oculomotor processes such as smooth pursuit, VOR, OKR, \nAND vergence eye movements. \n\nWhile  the  biological  microcircuits  of the  superior  colli cui us  and  brainstem  burst \ngenerator  are not well  known,  more  is  understood  about the representations found \nin  these  areas.  By  exploring  the  advantages and  disadvantages of various compu(cid:173)\ntational  models  in  a  working  system,  it  is  hoped  that  a  truly  robust  system  will \nemerge  as well as better  models to explain the biological data.  The construction  of \na  compact  hardware system  which  operates  in  real-time  can  often  provide  a  more \nintuitive understanding of the  closed-loop system.  In addition, a  visually-attentive \nhardware system which is physically small and low-power has numerous applications \nin the real  world such  as in mobile robotics or remote surveillance. \n\n4  Acknowledgements \n\nMany thanks go to Prof.  Carver Mead and his group for developing the foundations \nof this  research.  Our  laboratory  is  partially  supported  by grants from  the  Office \nof Naval Research  and the  Rockwell International Science  Center.  Tim Horiuchi is \nsupported by a  grant from the  Office  of Naval Research. \n\n5  References \n\nT. Delbriick and C.  Mead,  (1993)  Ph.D. Thesis,  California Institute of Technology. \n\nS.  P.  DeWeerth,  (1991)  Ph.D. Thesis,  California Institute of Technology. \n\nT.  Horiuchi, W.  Bair,  B.  Bishofberger,  A.  Moore,  J.  Lazzaro,  C.  Koch,  (1992)  Int. \nJ.  Computer  Vision  8:3,203-216. \n\nR.  Jiirgens,  W.  Becker,  and H.  H.  Kornhuber,  (1981)  BioI.  Cybern.  39:87-96. \n\nS.  G.  Lisberger,  E.  J.  Morris,  and L.  Tychsen,  (1987)  Ann.  Rev.  Neurosci.  10:97-\n129. \n\nM.  Mahowald, and R.  Douglas, (1991)  Nature 354:515-518. \n\nC.  Mead,  (1989)  Analog  VLSI and Neural Systems,  Addison-Wesley. \n\n\fAn Analog VLSI Saccadic Eye Movement System \n\nS87 \n\nLeft Burst Neuron \n\nLeft Motor Neuron \n\nMotor Error < 0 \n\nMotor Error> 0 \n\nOther inputs: \nVOR/OKR  ---t~ \nSmooth Pursuit \n\nNeural \n\nIntegrator \n\nI~II.IIII \n\nRight Burst Neuron \n\nIIIIII \n\nRight Motor Neuron \n\nFigure  3:  Schematic diagram of the  burst  generator.  The  burst  neuron  \"samples\" \nthe  motor  error  when  it  receives  a  trigger  signal  (not  shown)  and  begins  firing \nas  a sigmoidal function  of the  motor error.  The spikes feedback  and  discharge  the \nintegrator and the burst is shut down.  This \"pulse\"  signal drives the eye against the \nviscosity.  This signal  is  also  integrated by  the  neural  integrator which  contributes \nthe  \"step\"  portion of the motor command to hold  the eye in its final  position.  The \nneural integrator has  additional velocity inputs for  other oculomotor behavior such \nas smooth pursuit, VOR and OKR. Note that the burst neuron for the other muscle \nis  silent in  this  direction. \n\n\f588 \n\nHoriuchi, Bishoiberger, and Koch \n\n-0.25 \n\n0..00 \n\n0..25 \n\n0.50. \n\n0..75 \nlime (Iecondl) (10..2 ) \n\n1.00 \n\n1.25 \n\n1.50. \n\n1.75 \n\n2.00 \n\n2.25 \n\nFigure  4:  Spike signals  in  the  circuit  during  a  small  saccade.  (7.5  degrees  to  the \nright, starting from  4.8  degrees  to  the  right.)  Top:  Burst  neuron,  Middle:  Neural \nIntegrator,  Bottom:  Motor  neuron.  (one  of  the  outputs  of the  chip)  Note  that \nthe  \"neuron\"  circuit  currently  used  increases  both  its  pulse  frequency  and  pulse \nduration for  large input currents,  causing the voltage saturation seen  in the bottom \ntrace. \n\nEye Position VB. Time \n\n80. \n\n60. \n\n40. \n\nI  20. \n\n0. \n\nI!! \n:ii \n~ \ni \n\n-20. \n\n-40. \n\n-60. \n\n~D+---~----~----+----+----;---~~--~----+---~----~ \n0.,225 \n\n-0.025 \n\n0..200 \n\n0.,000 \n\n0..100 \n\n0..125 \n\n0..050. \n\n0..075 \n\n0.,150. \n\n0..175 \n\n0.,025 \n\nlime (lecondl) \n\nFigure  5:  Horizontal  position  vs . \ntime  for  21  different  saccades.  Peak  angular \nvelocity  achieved  for  the  60  degree  saccade  to  the  right  was  approximately  870 \ndegrees  per  second.  The  input  command  was  changed  uniformly from  -60  to  +60 \ndegrees. \n\n\fAn Analog VLSI Saccadic Eye Movement System \n\n589 \n\nFinal Eye Position VS.  Burst Command Voltage \n\n80 \n\n60 \n\n40 \n\n20 \n\n0 \n\n\u00b720 \n\n-40 \n\n-60 \n\nI \nI \n:1 \n~ \n~ \ni \nit; \n\n\u00b780 \n\n1.25 \n\n1.50 \n\n1.75 \n\n2 .00 \n\n2 .25 \n\n2.75 \nl,.,ul Voha. to  BUI'II Generator (center - 2.5v) \n\n2.50 \n\n3.00 \n\n3.25 \n\n3.50 \n\nFigure 6:  Linearity of the system for  the position data given in the previous figure. \nFinal eye position was computed as the average eye position during the last 20 msec. \nof each  trace. \n\nAverage of 10 Saccades from  \"center\" to 30 deg. R \n\n.. -..  .'.- .. \n\n35 \n\n30 \n\n25 \n\nI  20 \n\n8 \n:-e \n! \n~  10 \n\n15 \n\n5 \n\n0 \n\n\u00b75 \n-0.025 \n\n0.000 \n\n0.025 \n\nO.OSO \n\n0 .075 \n\n0.100 \n\n0 .125 \n\n0 .150 \n\n0.175 \n\n0.200 \n\n0.225 \n\nlime (Iecondl) \n\nFigure 7:  Repeatability:  The solid line shows averaged eye position (relative to cen(cid:173)\nter)  vs.  time for  10  identical saccades.  The dashed lines show a standard deviation \non  each  side  of the  mean.  Most  of the  variability is  attributed  to  problems  with \nfriction. \n\n\f", "award": [], "sourceid": 776, "authors": [{"given_name": "Timothy", "family_name": "Horiuchi", "institution": null}, {"given_name": "Brooks", "family_name": "Bishofberger", "institution": null}, {"given_name": "Christof", "family_name": "Koch", "institution": null}]}