T 7 .5 W2477 ENGI UC-NRLF C 3 D31 33b i 3fii UCB H /.) HS Stochastic Control in Eye Movement Tracking By r -* Shean Wang i t- + RESEARCH PROJECT Submitted in partial satisfaction of the requirements for the degree of MASTER OF SCIENCE in Electrical Engineering and Computer Sciences in the GRADUATE DIVISION of the UNIVERSITY OF CALIFORNIA, BERKELEY Approved: Lawrence Stark, Research Adf/i so r Edwin R. Lewis, Research Advisor T7 INTRODUCTI^II 1 sxp3T.i::.::T r j: JIIOD 3 033EHVA-TIOU ?^Ci: OJHZ EXP2S1MEIITAL RZ3UL23. 6 DISCUSSION 9 POSSIBLE ?U2URE Din2CO?lCi; 24 SUI'2>iASY 26 RZFSREKCI 28 PIGUHSS 29-72 range of fr-jcuer.cy , the eye can even anticipate the target. This leads numerous models of a predictor for the eye movement systen. Consequently, an effort is made to compare the experimental results of this report with those of existing proposed models. II. The pursuit and saccadic movements of eyes. Many established experiments have shown that the afferent signals are continuous but the efferent signals are either continuous or discrete or both. Thus the respective eye movement can be pursuit, saccadic, or both pursuit and saccadic. Therefore, questions are raised about the discrete nature of eye movements. Where does the discrete nature come from? How does it work? As of now, there is no physiological explanation. However, from the engineering point of view, there can be numerous possibilities. The most satisfactory model is a stochastic sample-data control system with a uniformly distributed inter-sampling time. This system v/ill be reviewed in this reoort. . -i. . :---. _ ;-. The experimental procedure for the measurement of horizontal eye movements has been well established and therefore only a brief description v.-ill be presented here. In a dark room, the subject v;as asked to fixate on a light spot which, is controlled by a target notion generator. A small beam of infra-rad light was directed to the left eye ball, and its reflection was collected by a pair of photocells aimed at the iris-scleral border as shown in Figure 3a. 2ye positions could be detected by the difference in output of these two photocells. Then an operational amplifier was used to amplify the difference and send it to a Sanborn recorder. In order to minimize the interference introduced by motion, the subject leaned his forehead against =i fir b?.r ?.nd bit on a piece of dinLal iia^ression wax as shown in Figure 3t>. Experimental data were collected from three subjects. Three different kinds of stimuli were presented: regular square wave, irregular square wave (or staircase wave), and ramp input. The latency of response to each stimulus could be defined in three different ways as shown in Figure 4a, 4b, and 4-c. However, in Figure 4-c, it was very hard to decide to which particular input the eye did respond. The time required for eye to respond upon a stimulus (reaction time) was about 150 msec, therefore, the tine larger than but closest to 150 msec was choosen. For each experimental run, a histogram of latency of response was generated. The binwidth was 25 msec, and the middle point of each bin was used as the latency within its + 12 msec range (see Figure 5 to Figure 11). The average and standard, deviation of latency were computed by IBM 6A-00 for each experimental run of both square wave and staircase wave stimuli. A set of experimental data v;as collected from the same subject with the sane stimulus by changing the frequency alone. The frequencies used are listed in Table 1 (see next page). For each set of square wave and staircase wave experimental runs, the inedian(H), average(Av), mode(Hd), and standard deviation(s) of latency were ploted vs. frequency as shown in Figure 5 to Figure 10. In the ramp input case, only median vs. frequency was ploted. Set I. Regular square v/ave input! 1. 2. 3. 4. 5. 6. 7. 8. 9. 10. 0.1 Hz. 0.3 Hz. 0.5 Hz. 0.9 Hz. 1.1 Hz. 1.3 Hz. 1.5 Hz. 1.7 Hz. 1.9 Hz. 2.3 Hz. Set II_. Irregular square v/ave inputs: 1. .150 Hz, .235 Hz, .488 Hz 2. .255 Hz, .488 Hz, .785 Hz 3. .488 Hz, .785 Hz, 1 .230 Hz 4. .785 Hz, 1 .230 Hz, 1 .480 Hz 5. 1 .230 Hz, 1 .480 Hz, 1 .710 Hz 6. 1 .480 Hz, 1 .710 Hz, 1 .950 Hz 7. 1 .710 Hz, 1 .950 Hz, 2 .210 Hz 8. 1 .950 Hz, 2 .210 Hz, 2 .450 Hz 9. 2 .210 Hz, 2 .450 Kz, 2 .730 Hz 10. 2 .450 Hz, 2 .730 Hz, 3 .010 Hz Set III. Samp inputs: 1. 2 /sec. 2. 1 /sec. 3. .7/sec. 4. o?'j /sec. 5. ,5/sec. 6. 0.33/sec. 7. O.l33/sec. 8. 1 6/sec. Table 1 . 'T ;1 -lj_. (A) Regular square v;ave inputs: The results from subject C (Figure 5) and subject W ( Figure 7) give similar conclusior. for the values of the median, the average, and the mode of the latency distribution curve. They shov; that these three values are all frequency dependent. But at the meantime, subject C and subject S shov/ a similarity in the curve of the standard deviation of latency distribution vs. frequency. That is, the standard deviation of the latency distribution is also frequency dependent. For the latency distribution curve, Let M = median Av = average lid = mode s = standard deviation then the above similarities can be summarized as follows: (1) 0^ frequency 1.5 Hz. M, Av, Ivd, and s are increasing as frequency increasing. Therefore the distribution of latency is shifted as frequency is changed. The distribution starts at the large time l-.g region, then shifts to the tine lead region. As frequency continue 3 J ;o increase, it shifts back tov:ards the time region. Subject 3 (Figure 8) yields a set of different results for ~~:ic.n, average, and node. These three values are all betv;een 100 nsec and 175 msec. The distribution curves do not have any obvious shift as frequency changed. The standard deviation curve from subject W shows that the value varies betv;een +_ 35 msec to + 88 "sec, and it is frequency independent. Despite the differences among the above discussed three subjects, they do have seme similarities. When frequency was increased, the number of abnormal responses of all three subjects would increase, and their distribution curves all were normal form. ( r- +- Q n -r"/->or*^N T.TOTr^ i f TJ*-? i~n*-*\s\ Q 4- V^-; v_-^ ^ ->._ ~ ..^Vt-y x - J-j-j* ^ %_; ow Figure 10) : The median, average, mode, and standard deviation of the latency distribution are not all frequency dependent. The values of median, average, and mode varied within the range of 150 msec to 250 msec. But the shape of the latency distribution curve seems to be frequency dependent. As frequency increases the distribution curve tends to change its shape from a nor~al distribution to a square form distribution. This change can be observed by comparing Figure Sa to 8c Figure 9a to 9c, and Figure 10a to 10c. (C) Ramp inputs (Figure 11): (1) The speed of target 0.33/sec: Latency decreased as speed increased. (2) 0.53/sec the speed of target 0.83/sec: As speed increased, the latency of response to the continuous portion increased (Figure 11), and the 8 latency of response to the discontinuous portion decreased (Figure 11). O) Tiie speed of the target O.G5/3ec: Latency of response decreased as speed decreased. The latency of response to the discontinuous portion is much less than the latency of response to the continuous portion as shown in Figure 11d. In this report, the discussion was divided as tv/o parts: one emphasises the pattern recognition associated with the eye movement, the other one concentrates on the discrete nature of eye movement. In either case, different models were discussed. Comparison was made between models or between those models and my experimental results. There were three models discussed in the pattern recognition phenomenon, namely, Dallos and Jones' model, Susie's model Cyr and Fender's model. In the discrete phenomenon, Forster's model was reviewed, and Latour's experimental results were mentioned. (A) Pattern Recognition (1) Dallos and Jones' model: From the above presentation, it is obvious that the distribution of latency is very much dependent upon both the wave form and the frequency of the stimulus. In fact, this is the property that led Dallos and Jones'" to theorize about the "acquisition of learning". From their theor;/, a model is proposed. Three properties of this model are worth discussing, namely, learning speed, prediction, rnd detection. (a) Learning Speed By examining the first few cycles of tracking responses to sinusoidal and square wave inputs (Figure 12), Dallos and Jones believed that the rate of learning was considerably greater for continuous than for discontinuous input with the same frequency. But from my experimental observation (see previous section and Figure lid), the response of the ramp tracking shov/ed that it is not so. When "both continuous and discontinuous changes were presented in a repetitive manner, namely the ramp stimulus, the rate of learning in the discontinuous portion was faster than that of continuous portion (Figure 11d), thus contradicted their conclusion. Therefore, the correlation, if any, between the rate of learning and the continuity of stimuli remains still a question. (b) Predictor For each input stimulus, an output Bode plot v;as ploted. Then the corresponding closed-loop transfer function was obtained by curve fitting. Basing on the assumption of linearity and unit feedback, the corresponding open-loop transfer function for each stimulus can be computed froa its closed-loop trancf c-r function. Let G I) (ov/)= closed-loop transfer function for sinusoidal input . G(jw) = closed-loop transfer function for random input . gr>(d w )= open-loop transfer function for sinusoidal input . g(jv;) = open-loop transfer function for random input . then gp(jv/) and g(j'.;) can be computed from the following equations: E G p (jw)= 11 By cor.parins the tv;o open-loop transfer functions, G and gCo'w), a hypothetical predictor transfer function P(jw) was defined as follov;s: Since this predictor v:as derived aia thematic ally, its validity should be examined by various ways. Firstly, hov; reliable is this predictor? According to Dallos and Jones, the predictor should give a phase lead starting fron 100 decrees increasing to 250 degrees as frequency increased iron 0.5 cps to 2.5 cps. In our experiment, the subject S was presented with square wave stimuli (as shown in Figure 6d). The frequency distribution of latency for the subject showed that the existence of prediction is question- able. As frequency changed, there was no significant shifts in the dlslrlouljion curve. ( The latency of the response fall into the interval of 100 msec and 1pO msec) It therefore implied that the response of subject S failed to predict any of the stinuli within the applied range of the frequency of the square wave. Secondly, under a specific frequency, what kind of distri- bution curve of latency would the predictor have? Because by examining this curve, I believe that the efficiency of the predictor could be obtained. A sharp and narrow curve (eg. a nornal distribution) could represent an effective predictor. On the other hand, a flat and wider curve (like rectangular shape) could mean a weaker one. The efficiency of the proposed predictor was not discussed. Thirdly, since learning was an unconventional process, it should not be inconceivable that the predictor ni.-ht not be 12 physically reclined. Hut from a physiological point of vie;;, it would be an interesting and meaningful thins to know the possible mechanism or location of the predictor. But Dallos and Jones made no attemp to suggest it. (c) Detector Since the operation of the predictor depended upon the wave forn of stimulus, there should be another element, namely, detector, which could exhibit the predictor under periodic stimuli and could inhibit it under aperiodic stimuli. A simple detector was suggested by Dallos and Jones. The detector included a memory unit M with a finite decay time, and a comparator D (Figure 12). The comparator would continuously compare the output signal from the memory unit and the present error signal. If the two signals were dislike, then the error signal would bypass the predictor ^ *>* -I- r^ i stimuli. If the tv/o signal v;ere alike, then the error information would be channelled through the predictor. Based on this model, Dallos and Jones tried to explain the frequency response of square wave stimuli. There were three cases: 1. For low frequency stimuli: Since the nemori sable information for a square wave appeared to be the time duration betv;een the transitions, therefore, when the time duration was longer than the memory trace, then no information could be stored, consequently, no learning process could proceed. 2. For high frequency stimuli: Since the target motion was too fast, even with a periodic input, the error signal could not be a periodic one. 3o learning failed. 3. For intermediate frequency G : ;:.auli: A steady state tracking could well be established. The amplitude of distortion in eye motion was not significant. V.'ithin this range of frequency, there was almost complete synchronism between input and output, therefore, the memory trace could not be continuously reinforced. Since the memory trace was assumed to decay with time, then after some period of tine, the loss of synchronism should be expected. Dallos and Jones observed that in the steady state tracking, the average time duration for good tracking period was approximately 5 to 7 seconds. Following those good tracking periods, there were a few cycles without compensation. From the above explanation of the square wave response, it seemed to me that there were some points open for argument: (a) It is questionable, that whether the large value of latency during a high frequency target tracking was cauccd by the aperiodic error signal or noo. If we took eye muscle into consideration, it should be easy to demonstrate that the eye could not move as fast as one wished to. Suppose a subject was asked to count repeatedly "one, two, one, two, one, ", and simult- aneously, the subject's eye moved back and forth between two points in front of him, the counting of number and the movement of the eye were synchronized. If the counting started at a low frequency, the eye could keep the synchronisation very well. If the speed of counting was increased, the eye gradually become unable to follow. The back and forth movement of the eye would get worse and worse as the speed of counting increased. In this demonstration, the brain knew the speed and the exact positions of the two fixation points, The speed c.nd position were not known by "learning" but by "given", the failure of the eye showed that there 14- was a speed, lir.itc.tior. for the eye. It should not be confused with the arbitrary eye movement, namely, the eye was asked to cove bad: and forth between left and right without specific position to fixate. In this case, because no accuracy was required, the eye could respond faster than in the former case. In the case of eye tracking movement , the speed of the eye and the positions for fixation v;ere not given. If the eye could "learn", as suggested by many models, it should be expected that the response would not be better than the "given" case. Thus the bad tracking response of a high frequency square wave stimulus could be caused simply by the limitation of the eye muscle movement. I was one of the subjects. According to my own experience, at high frequency.' tracking, I could "feel" or "sense" the frequency of the target, over. v.'hcr. my eye failocl to follov. 7 it. Therefore, it seemed to me that the "learning behavior" should belong to the brain, not to the eye. Thus when the target frequency was high, and the eye failed to track properly, the brain still could "learn" the frequency of the target, if the target motion was periodic. (b) There were many mechanical models for the muscle. No matter which one was the best, at least they implied that the muscle possessed mechanical properties. Therefore, it would be proper to assume that there was a nature frequency for the eye muscle mechanism. If the frequency of the target motion was close to the natural frequency of the eye muscle mechanism, then a resonance phenomenon might happen. This was probably the reason for the good tracking responses for the intermediate frequency stimuli. The above argument was just induced, from ''common cense" and "experimental feclir .j" , no strong evidence or experimental results supported it. However, it seemed to me that using mechanical (or mathematical) model of physiological properties to approach the eye tracking s-stem might be more realistic and practical than a pure mathematical derivation or curve fitting. (2) Sugie's model: A different approach to the stud;/ of the "predictive control" of the eye movement v;as carried out by Sugie. After studying the response to regular square v:ave inputs, he proposed a mathematical model of estimation based on stochastic optimal control concept (Figure 14- and Figure 15). The variable to be estimated was the period of the target notion, and the statistical parameter was the variance of the estimated period. Four assumptions were made, (i) The estimation depended upon the target frequency and the number of target cycles, (ii) The estimated period of the target motion was accom- panied v;ith some stochastic randomness caused by the uncertainty of human memory, (iii) The mean of the estim- ated period coincided with the period of the target motion, because the probability function had no reason to be biased, (iv) The variance of the estimated period should depend on the target frequency and the number of target cycles (Figure 15). A brief reason for the last assum- ption Y/as worth mentioning: (a) The dependence of the variance upon target frequency: As Dallos and Jones suggested, for a square wave input, the useful information could be obtained only twice a cycle. V/hen frequency was low, the memory dec.'.y would cauce uncertain information, therefore, the variance of the estimated r>eriod could be assumed 16 large. However, when the tnrget frequency was very high, the sample data phenomenon was taking into account. Accourding to L. R. Young and L. Stark-"*, the sampling period was 250 msec. Therefore, when the input period became compatible with the sampling period, the change of the target position could not be sensed very accurately. Consequently, the variance of the estimated period could be assumed large. Thus the variance should have its minimum value at the intermediate input frequency range. (b) The dependence of the variance upon the number of cycles of target motion: As more information became available, the variance of the estimated period should be decreased. Consequently, the variance of estimated period should be a inonotonically decreasing function of the number of cycles. For simplification, he assumed a rectangular probability density function (Figure 18), that is, / i probability density 2 = otherwise a The estimation was assumed to optimise a "performance index". This performance index PI was defined as a mean-squared value of the length of time by which a stimulus preceded or laged behind a response. ( 91 O Cy' + "kri-r.-*- ~ X ' p-r , +.<- . fl-f- 4. TT Ot)o .r X = I w LLUJt ' 2y< where x = constant reaction time without prediction. t = ontinr.1 reaction tine. opt The response of this r.odel (7i~ure 1?) rave a .3000. suggestion for the possible range of the mode an'l average for the distribution of latency. But it failed to jive a desired distribution curve. As Dallos ar.d Jones, Sugie assumed the memory decayed as the only explanation for the low frequency square wave response. However, he had a different explanation for the higher frequency response. He considered the discrete nature of the saccadic eye movement which, I believe, was more realistic than the explanation given by Dallos and Jonse. Sugie assumed that the variance also depended upon the cycle number of the target notion, that is, the variance was a nonotonically decreasing function of cycle nunber, independent of target frequency. Statistically, it should be true. But practically, it would be doubltful. When conducting experiments cf a longer time duration, the eye night be fatigue, which could give unreliable responses. Therefore the variance could be larger as sore experimental runs were recorded. 2n the derivation of his model, Sugie assumed the number of cycles of the target motion to be infinite. Since the data he used was fron several other models, therefore the cycle numbers would not necessary be equal nor infinite. It seemed to me that the dependence of the variance upon the cycle number was not a proper assumption. Another interesting assumption worth mentioning was concerning the unbiased probability density function. A rectangular function was assumed for simplification. But Sugie did not suggest any other suitable functions. Whether a normal distribution function could give a desirable distribution curve for latency or not was worth to know. 18 (3) Cyr and Fender's model: After a number of predictive models (lender and ITye, 1961; Young and Stark, 1963 i Hobinson, 1955) for the eye tracking system were proposed, the existence of the predictive ability of the eye movement was challenged by Cyr and Fender. Based on the study of human eye movement in two dimensional tracking task, Cyr and Fender found that the system was non-linear. Therefore, a transfer function could, be derived for the oculomotor system. Thus it should not be possible to predict the response to one class of target motion by linear combinations of the responses to other classes of stimuli /\ (as in the Dallos and Jon model ) Consequently, the computation of a minimum phase lag (Dallos and Jones, 1963^; Young and Stark, 1963 ) should not be possible. Accordingly, the latency was only a simple delay which depended on the class of target motion. Since all of my experimental results of this report were from one dimensional eye tracking movement, therefore no comparision between Cyr and Fender's results and mine could be carried out. However, there was a point which might be worth mentioning. In the one-dimensional eye tracking movement, the ability to predict was based not only on the frequency of the stimulus, but also on the shape of the target motion. In other words, the eye could predict, if the stimulus was of periodic form with a intermediate frequency. Thus for disproving the predictive two-dimensional input with a intermediate frequency should be used. But Cyr and Fender only used two kinds of stimuli, one was random signal, the other one was a sum of four small sinusoids in both vertical and horizontal directions. Therefore, no periodical input was used by Cyr and Fender. Consequently, their conclusion was not strong enough. (B) Sirro^o ITature Of .77 e Movement The discrete nature of the saccadic eye movement control is well established. Since the introduction of the sample-data model by Youn;~ and Stark, the analysis of this discrete nature has developed numerous refined models (see reference 6, 7, 8). There are three different types of sample-data systems. A model has either target-synchronized or target- asynchronized properties. A target-synchronized system is c ock-asynchronized. But a target-asynchronized system can be either clock-synchronized or clock-asynchronized. These definitions for the three types of sample-data systems are summarised in Table 2. desk- clock- targe t- STnchronlzed impossible possible target- asynchronized possible possible Table 2 20 (1 ) Latour's experimental results: In the early studies of horizontal eye movement by Latour q and Bouraan , the frequency distribution of reaction time for both short run ( Figure 24- ) and long; run ( Figure 25 ) were obtained. Latour noticed that, in short sessions the reaction time distribution was multimodal, such that: t = ka where t = reaction time = most frequent t value ** ' 2 j 3j ~ - ~ a = 20 to 40 msec. In longer sessions, this phenomenon faded out* But the distribution of the "difference between the reaction time and its succeeding movements" showed a similar frequency distribution curve. Let *t = t n - t n+1 where t = reaction time at the nth reaction t n+ -p reaction time at the (n+1 )th reaction At = difference between the reaction time and its succeeding movement. then the distribution of *t ( Figure 26 ) could be represented t = T - where ^ = some positive value By comparing the frequency distribution of reaction time t and the frequency distribution of the At ( Figure 26 ), Latour concluded that there was a continuous decrease in the " most 21 frequent reaction tiae " , follov.od by a relatively fact increase of within a period of about 20 seconds ( Figure 27 ). In other words, for a long session tracking, the eye could graduatly shorten its reaction tine then rapidly drift back to the starting value of reaction tine. This cycle could be continued as the experiment went on. Latour's experimental results and observations could be in the discussion of the following .burster's model. (2) ii'orsber's stochastic sample-data model: n The studies by Lang' showed that the model for the discrete nature of the eye movement was more satisfactory if it was o randomly sampled. Forster reexanined the three possible types of interosample time distribution function ( see Table 2, page 19 ) with his revised sample data model ( Figure 19 ). Started from target asynchronized system, he assumed that: L = t 1 + D where L = latency D = dealy time from occurrence of the input to the next occurrence of a sampling instant (-from this definition D should be strictly non-negative ). t* = constant delay time, as in Figure 20. f v (x) - probability density function of the variable 3C x ( any of the above variables ). Since D was a random variable, L should also be a random variable, thus the density function would be related by: = fy( a 22 .based on the assumptions that: (a) the input occurcd in as interval f , (b) the time between inputs were greater than f , the DloLjC maximum value of f_ , C (c) the input v/as uniformly distributed Forster derived out f. '!) /fmax i) clt , for Di ' max J i) for D*. Since fy(-U) was a non-increasing function, thus it should give a non- increasing latency histogram ( Figure 21 ). But the actual latency distribution increased for low latencies. Therefore, Forster suggested that for asynchronized sample, the eye sometimes was able to shorten the sampling intervals, when a step v/as observed. This hypothesis could be supported if the function v/as as follev/ed: 0.0133 60 . G. J. 3t-Cyr and D. H. Fender, " Nonlinearities of The Human Oculomotor System Time Delay". Vision Res. vol. 9, pp. 14-91-1503. Pergamon Press 1959. 4-. L. Stark, G. Vossius, and L. R. Young, IEEE, Trans, on Human Factor in Electronics HFE-5, pp. 52-57. 1962. 5. L. R. Young, and L. Stark, IEEE Trans, on Human Factors in Electronics, IIFE-4-, pp. 32-51. 1963. 6. L. Stark, "Neurological Control System, Studies in Eiocngincering" . Plenum Press, Nev/ York, pp. 240-24 7 !-, 351-357. 1953. 7. G. V. r . Lang, " Representation of The Human Operator As A Sampled- Data System", PhD Thesis, University of London, 1967. 8. 3. M. Forster, 2. M. Thesis Report MIT-6S-2, Man Veliical Laboratory, Center for Space Research, HIT, Cambridge, Mass., 1968. 9. P. L. Latour, and M, A. Bouman, "A Non-analog Time Component in Eye Pursuit Movement". Proc. MIT Symposium, 1959. 10. P. L. Latour, "The Eye and Its Tinirng" . 1961. 29 Muscle Afferent Si en Efferent Signals Visual Corte:: 17, 13, 19 Afferent Signals : continuous. Efferent Signals : continuous and discret Fig. 1 sampler target position eye detector controller and plant eye position Fig. 2 iris > solera nhotocell subtraction moving li^ht spot (target) oscilloscope; . 3a. Fir:, ob . (a) Signal Normal Response (b) Abnormal Response (c) n v L=min. (L i >150insec. ) , 1=1 Signal 1 1 1 i 1 1 1 1 Response 33 0.1 or 3. of occurer.ce. M: . : ;ns. -. 277.15ns. s: 75.75 - 5 2 - 1' I 200 2 35 ; -' : .')0 us 2 . 0.3 cr: s . of occurenc . 51 4- 3 M: 14-Oras. Hd: 14-Or.s. Av: 135.56ms. s: 33.82 50 100 150 200 ms 3. 0.5 ~ps fr Of OCC. M: 100:as. Md: 100ms. Av: 77.71ms. s: 63.04- 5- '- 4. -. ' 1 ML , INI , . I , . . I ~i5o so 50 5 100 150 200 ms 4. 0.7 cvs. ;,'-of occ . M: -10ms. Md: -130.-80,-70,-30,0,30,50n3 Av: -7.5ms. s: 91. 06 3 - - 2 ' | j 50 Q 50 100 200 500 4-00 n ms 5. 0.9 cps. 2 of occ. M: -60ns. Md: -70, -60ms. Av: -4-9. 60ms. s: 52.56 c " s 4- 2- 1 "' \ I ! . 1 1CO. 50 50 100 1GO Predictable square '..av/o; Sub; ^'j; 1.1 c-os. .IS . Av: 9.43ns. s: 57.39 ox ucc; . 5 5 2 1- j 100 50 50 100 130 ms 7. 1.3 cps. II: 50ns. Md: -10,30ms. ': Of OCC . Av: 37.04nis. s: 48.89 6 5- 3- 2 1 I i 50 6 50 100 150 ms 3. 1.5 cps. M: 70ms. Md: 50ins. V of occ. Av: 63.6833. s: 40ms. 8 n 7 G 5' 4 . 3 2 1 j 50 "100 150 ms 9. 1.75 cps. M: 90ms. Md: 100ms. of occ. Av: 85. 00ms. s: 53.49 12 11 10 9 8 7 6 5" i I ^> 3' 2 1 1 1 ; i : 1 | f i , J_ 35 ^- . . s: 59. 80" .-!-5as. or occ. 11 10- 9 8- 7 6- t 5' 4-- 3" 2" f j X 50 100 150 200 250 300 ms \ M. vs. freo. \ Average x K \ Mode Q D >00 \ ^ Median * 0.1 0.3 0.5 Hz 80 70 60 50 Standard deviation X1 O.'j 0.5 0.7 0.9 1 1 1.3 1.5 1. 75 2 Figure 5c . 36 ( res . ; : : . ' . . 2 - of occ;. . 50 100 150 200 ms 2. 0.3 cjs. (res. 6 of occ I-:-. 113ns. Md 125ras, Av.: 153.3; : i : +55.21 50 0.5 cps. i' Of OCC. 8 6 -f 100 (res. 22) 150 200 ms M: 125ms. Md: 150ms. Av. : 73.8ms. SOJDV: -120.79 200 150 100 50 50 100 150 200 250 ins 14 - 12 ' 10 * 8 6 ' 0.9 cps. (res. 32) ~ of occ. II: 100ms. Md 100ns. Av.: 103.67 S'IDV: 157.45 50 100 150 200 2=0 300 350 ms (a) ig. 6. Predictable squaro v.'uve. Subject: 3iiiro. 37 1.1 Tr ~. f occ. 20' 18 12 1C 8 6 . Av.: 133. ST. 50 100 150 200 6. 1. 5 Hz. (res. 51) M: 125ns. I-id: 150ms ' o-P ,^ AV ' : IpO.OmS. ,r 01 occ. orr7w . zo co 16 * -X J 14-' 12 10 8- 6 2 ' C-/-1 xl/- /^/^.x^ ^^ r s\ 50 150 250 7. 1.5 Hz. (res. 65) # cf occ. M: 150ns. lid: 150ms. Av.: 136.11ns. MJV : pu. / i 22 20 18 16 14 12 10' 8 6 4-' 2 "" 1 ; [ 50 100 150 200 ins iS. 6b. 1.7 (resp. M: 12>-.i. . : as, Av. : 1 . . ".c; . ,, .1 Nu/ ^/ 'w/i-H I/ * STDV: I 73.39 8 - 6 2 i I . ! -50 6 56 100 150 200 250 300 ns 11. 2.5 Hz. (resp. 52, abnormal 80) $ of occ. M: 100ms. Md: 100ms. Av.: 126.08ms. STDV: I 77.30 10 8 6' ) 4- ' 2" ; i , t 50 100 150 200 250 300 m: Fi-% 6c. 39 175 150 125 100 Average X- I-Iode Q- i'ledian * D.I 0.3 0.5 0.9 1.1 1.3 1-5 1.7 1.9 2.3 2.5 Hz Standard deviation 0.1 0.3 0.5 0.9 1.1 1.3 1.5 1.7 1.9 2.3 2.5 Hz Vir;. 3d. 1. . :is. (resp. 5) 2. 0.5 Hz. (resp. 16) of occ. or occ. 5 4' 4 3 3 2 2 I ! 1 1 i 1 i 100 200 as ( , I 100 200 ras M: 200ns. Md: 2?0ms. M: 100mc. Md: 100ms. ;-.v: 132.5ns. s: 68.46 Av: 105.5zas. s: 46.8 c 3. 0.5 Hz. (resp. 22) 4. 0.? Hz. (resp. 25) of occ. -/? of occ . 7- 6" 6- 5' 5- 4 4- 3 3- 0+ 2- I in i | i 1 I f i ' i r- 1 I -50 50 100 ns -150 -50 o 50 100 i M: 50ms. Md: 100ns. M: -25ms. Md: -50ms. Av: 51.2ms. s: 76.9 AV: -2ns. s: 74.9 5. 0.? Hz. (resp. 30) ; of occ. M: -25nis. Md: Oas. 10- Av: -22ns. s: 59- r -is. 9- Q 7 6" 5 4- 3 : 2- 1 1; 100 ;:o o 50 100 ns (a) ?iS. 7: Prodictable sour.ro v;ave. Subjsct;: S. 6. 1.1 Hz. (resp. 34) M: Ons. Md: Cms. 10 of occ. s : GO .-.v : 9 8- 7 6 5- 4- 3 2- 1 -100 -50 ( ) 50 100 200 300 ms 7. 1.3 Hz. (resp. 50) 20 18- 16 12 10 8 6- 4 2 of occ . -50 o 50 100 200 M: Oms. Md: Oms. s: 48.5 Av: 18ras, 300 ms 8. 1 10- 9 5 Hz. (resp. 46) of occ. M: 25ms. Md: 25ms. s: 68.8 Av: 14.6:23, c 7 6' 5 ." 5 ! - 1 * i i } I . ! . s\ r- r\ c r\ r\ c.r\ ir\r\ onr\ znn m 42 "I r ' " ( -p. - i r ' : 3. 're sp . 4-6 of occ. ~f .~>"C - ^ # 25 , _ _ 9 24 ^s 22- 21 20 19 18 17 16 15 14 13 - 12 11 10 9 - 8 7 6- 5 - 3 4 2- 1 1 " , i 1 , ' I , 50 100 1= ~C ?0 ms 1 DO 200 300 as M: 75ms. i:i: 100ms. H: 100ms. ] Id: 100ms. 200 150- msec. T- T r>^ ^i ^>/-\ \x_ IMUU.I p^ r* ~ ' ci r\ ' 100 > ^ D -0,1'iea. /r^ x^AVe] cm 50 ^^/^""^ x/ / / \^- X N-^,^ r ^ i ^'3ta3 CS.QQ X X idard \.7 QQ ,> r "^--- :3 0.1 0.3 3-5 \v ^Kl.1 1.3 1. 5 1.75 2 Hz. -50- or Pig. 7c. -;. j.1J, 0.257, 0.468 Hz.' resp.: f . ' . -'id: abnor::'il: 2. \J J. VJO ** V: - ','.^3 20 18 16 12 10 8 6 t 4 2 ' I 50 150 250 350 450 550 650 ms 2. 0.235, 0.488, 0.785 Hz. resp.: 90. M: 150ms. Kd: 150ms. Av.: 167.70ns. # of OGC - STDV: - 52.48. 30' 28 ' 26 ^ 24 .-. 2 - - 20- 18' 16- 14- 12- 10- 8- 6- 4- 2 - \ 50 1CO 150 200 250 300 350 ms Fiq, 8: Unpredictable 3P\;r.rc wave. Subject: G. 44 . , . >, ,23 Hz. of occ. . : ' . : . Av. : 179. 1m DV: - . . 20 16 14 12 ' 10 8' 6 4' 2 ' 100 150 200 250 300 ms 4. 0.73?, 1.23, 1.43 Hz. # of occ. 24i 22 20' 15- 14 12 10" 8-- 6' 4" 2 resp.: K: 175ns. Kd: 150ns Av.: 199.32ms. STDV: I 73.4. c 100 . 150 200 250 300 350 400 450 500 ns 5. 1.23, 1<48, 1.71 Hz. # of occ. resp.: 45. abnormal : 7 . I-I: 200ms. I-Id: Av.: 2Q7.95ms. STDV: - 66.2. 14- 12- 10- 8- r. 4- 2 i ! 1 . 150 200 250 500 550 400 **? 6. 1.43, 1.71, r- 'I l\ t M: . L: 125ms. STDV: -133.5. 4 2 I i , I ns 1C - 5CO OC 700 7. 1.71, 1.95, 2.21 . ,7 of occ. 4i res:..: 21. M: 15Cms. KG: ICO, abnormal: 10^. 125, 150ms. Av.: 159.52ms. STDV: - 65.98 50 1 150 2: 250 300 ns 8. 1.95, 2.21, 2.45 Hz. resp.: 26. M: 125ns. _y ^j> # n ' fOt "V . : IO >--'-" a //- '-'-i- ^ -^ -^ T~ o<~7 ^>^v oxijv : -o/.Oi 4- 3- 2. 1- L 1 56 100 15 200 250 5C 350 as 9. P , <- res?.: 19. M: 175ms. 7- # of occ. ctijxiui; JCLX; ^o^> ^i.v.: o.^ ;..:.. SJDV: 1 48.' . 6' 5- >^ - 2- 1- i r __ 1^Oi --> ,-.,-> OC^O Xr^~' ' '. oC. 4-3 10. 2.4-5, 2.73, 5.0" 3 r 2 1 4- occ, recr>. : 1 " . : 150 250 M: . : : - 300 ms 200 150 100 msec i Average X- ;-:ed- *- Ilodel a- }' \. .../: \ ** -v V , \ Standard . o.evi-.oion 123^4 567 -i of 10 order exrjer. 1: 0.15, 0.235, 0.-':-33, Us. 2; : 0.235, 0.4SG, 0.735, Hz. 3: 0.4-88, 0.785, 1.230, Hz. 4-: 0.735, 1.230, 1.4-80, Hz. 5: 1.230, 1.4-30, 1.710, Hz. 6: 1.4-SC, 1.?10, 1.950, Hz. 7: 1.710, 1.9?0, 2.210, Hz. 3: 1.950, 2.210, 2.4-^0, Hz. 9: 2.210, 2.450, 2.750, Hz. 10: 2.4-50, 2.730, 3.010, Hz. Pis. 8d. 4-7 . . , ccc. 28 26 24 22 20 18 16 14 12 10" 8 6 4 2 : 10. -100 (UN -I ro-v,-. . : . - .7 50 100 150 200 250 500 $50 400 ms 2. 0.235, 0.433, 0.755 Hz. if- of occ. 2? 261 24 22 20 16 12 8 6- resp.: 67. M: 150ras. Md: 1,50ns. abnormal: 3. Av.: 161.15ns. SO?L>Y: - 54.67. 50 100 150 200 250 300 350 400 ms . Unpredictable ccuare ucve. out;' : ^o. 17 -i rr - J. ,. ab. : : . 26 24 22 20 16 14 12- 10 8 6 ^ o f OCC. 100 1. D . A.V. : .3. STBV: - ms 0.755, 1.23' rr OT uCC. r~ s~\ s\ r~ % ^j o 50 Hz. resp. : 77. abnormal: 32. 150 250 450 M: 200ms. Md: 200ns Av.: 2Q2ias. 20 18 16 14 c 12 10 8 6 4- 2; I , i , , , ' , ns 5. i.::33, 1.430, 1.7"<0 Hz. ;' ol occ. 10 resp.: 52. M: 225ns. Md: 2503S. Av.: 211.G7r.s. STDV: i 53.99 4 2 i ! i 100 2-JO 500 400 500 600 ns 49 . 1.710, Hz res . : i . abnor:"-!: 1 100 300 M: 2, ;. lid: ?. r -Cns. Av. : u.u uu^; 3TLV: - ' 14 12 10 i 8 6 4- 2 . ! I i ins 7. 1.710, 1.950, 2.210 Hz. # of occ. 10 res-.: 41. M: 250ms. Md: 250ms, abnormal: 209. Av.: 220.38ns. SOJDV: i 128.58. 8 6 2 ) I I 1 I ! 50 150 250 350 4-50 550 ins, 8. 1.950, 2.210, 2.450 Hz. # Ox OCC, resp.: 13. abnormal: 58. M: 225ms. Md: 200, 250ns . Av.: 234.62ns. STDV: - 66.57. 4 2 i I I I i = 100 200 300 400 ms 9. 2.210, 2.450, 2.730 Hz. # of occ. resp.: 23. M: 200ms. Md: 150ns 6 ' 2 ' 1 i ' ! ! STBV: --124 . ! ! LJ , ' .M.vs. freq. 300 500 ins Average x x "Median - '.ode a - G deviation 50 1. 24 * /N TL ' 0-217 ' ccc. - . . : . . : -2". 1: ' . ". : 1^ . "1ns. : - 7.6v. 22 - 20 13 - 16 12 10 8- 6 2- ; i i ! , , 50 100 150 200 250 300 ns 2. /\ : 0.3--6 PIz. i 3. XL: 0.217 Hz. - 3. # of occ. 8 6 resp.:36. M: 150ms. Md: 150ms. S^DV: - 69.06! -175 -100 100 200 300 ms 3. A : 0. r -55 Hz. : 0.3S'i- Hz. 16 14 12" 10 6' 4. 21 of occ. resp.: 43. H: 150ns. I'd: 150ns. abnormal: 30. Av.: 156.97mc. STDV: - 32.42. 30 100 Fig. 10.: Unpredictable sif-nv.-.l. Subj-co: './ang. : - J\ : . . : 23 . - . - _ . Av . : ' '. : - . , \ 14 12 10 8- 6 occ. 5',} 100 15D 2 ::'.: '-' ' ' 350 4CT 4-^0 $00 ns 5 . A : . 3 n : 0.217 # of occ. 8-- 6 4 p rasp.: 32. ?rr.al: 4. M: 150ns. Hd: 150ms, Av. : ,06ms. STDV: I 62. -3. 100 150 260 250 300 350 ns 6. A : 0-5?G Hz. : 0.217 :-:z. + - .: 60. M: 150ras. Md: l5Cas, abnoriaals 11. Av.: "\y\. 2an. 3TDV: *70.7*. y o I OCC . 12' 10' B- & 2' i I i i ; i __] 100 200 300 400 500 ms Fi-r. Ob. 52 200 150 100 - 50 msec tandard deviation order of experiments. Pig. 10c. 55 ?v;o 1:3 ' .ular -;ave: a * b. ?,:;-: edr;ej Dro ' edge: 1 7-* o Cc.s. c. r. /r 6- /. - t 4 i i- i I 2- . nif 50 100 200 n s . 50 100 1 JO 200 250 rr ri;3. 11: P.o.-p fimc-'cior.. Subject;: '/anj; 54- o O C /~"~ -1 ^ v '.' . ", ' ' ' <~ ' ?f^-r- . 17,fy,~ ,i , ci / -. ^ , - J i '-.> -lo ,;L/.^O. 6- ;/- WJ. UU<^. * V < 2- ? M 1_ j i f 5 ~ 1 00 1 ; , i -i -i 200 250 300 ri o 5. 1 /sec. # 01 occ, o 5' 5 2- 1- I 1 ^ i r r r ! 1C 1^0 '2ho 2^0 500 350 4-00 as " C b. -,, 3:\:s. i-'i : 575^-s M,: 170:::s. 2-f 1 " . . . r ; . f . - r T J 1 1C o' Mo '200 25'0 . 3C50 3^0 40b- 4-50 r.s '# of occ] ,rpo a*~>-r\r* T> "* 'I^Q -^"~i ^ T* ^*'^ r 7 ^r-^ cXllC D i'ii. Z 1 ^ ' / yiilO * 1\ t J'~> / y UiS M d : 175~:3. 5 " j i 1 ' M, , I I . 1 I 1- icto 150 200 250 500 350 4-00 55 6. 0.5 /sec. fypo a. H t : 175-:'. :r : 275ns. M : 175ms. d' # of occ. 2 4 1 ' I 1 1 , . _ - I 100 i::"- 300 ^-00 5 - :.o 600 ms 7. 0.33 /sec. C"pe a. H t : 2; Ons. II: 225ns, of occ. . , t__t__*_. ]_..., . . * I . 14 i J i- i i f ( - L. i | ibo 150 200 250 300 350 350 ms 3. 0.133 /sec. Type a. I-Ir 250r s s. H,,: 600ns. H: 225ms, I ~ O- OCC . 100 200 300 1 500 600 ms 9. 1.G /ssc. Type a. I-L: 175ms. M p : ?00ms. II^: 150ms. i< r l 2 if 11 or occ I ' I L L il --J- --.*- ( l-~l I > i . 100 150 200 250 ms Pic;. 11c. 56 600 - 550 500 450 400 350 -- 300 -' 250 - 200 - 150 msec. 07T53 0.5 0133 0.33 0.7 1 Total response: -1-! CT- T- -. -T^-^ . Drop in:; edge: 1.6 /sec Fir. 11 d. 57 Sinusoid: Latencies (i 20C 10( 0.5 1.0 1.5 2.0 freq. Square wave Latencies (ms) 50Q 200 100 100 .5/1.0 1.5 2.0 -- o - e- or..v. st c;-clo 1 cycle 2 cycle 3 Fir;. 12 c(t) r(t) = target motion c(t) = eye motion 1-1= memory unit D= comparator P(ov;)= predictor g(jw)= plant FiS.13 59 TJ V ( 'H W O OT M OJ c o O i-i O Pi t, 0) c o ! o u Q) (X OO 10 -p g 0> fH O -P * / e . -H w o O o -H nj U-i w rrj i ^ H H M 1"^" 3 O ^ j< 1 P -P o u "" " ~"> d 3 <4H CM r-~ 2 rH W O O rH 2 H| W 4J J) 1 II II W || II O M O N -H C >< - pJn E 3 U> pq RJ c Q S -p W 1 \ > w ^ 6 i i \; to i' < 7* 1 r p-i : Q f| o o *d S M 'd r-l 1 ^^ IN >-i O -P ^ V i O P (3 >i N < n *- / -H j O N 0> rH T3 rH TJ (1) rt Gi 13(1) O 'O y" 1 . J o -i- 1 eO -H 4J /. nJ W T3 -H rH W M (tf p nj j U (0 tt c: t-i CJ O r. o 70 LT\ CM H ft M O H B O 8 hn r: fcD C C H CO O E d o H P O S o O O I n H d c 0) 71 *^-.s c ^ O .0 j; to : 5 ^* O O to "O o o (D O r to .- o IT) I o t.o H U) o c o o n3 0) G) ra o o 3 u o o c w H d Q> -P H O C o fH 45 eo o C o g. OJ 0) c o +> O 3 C O ~ cj -r* to O-H HrS w r., fe) < *-fc* snc *-* -y^r ~. - : V,"vaa j.'**-t_e_t* j . P O * H n *-m G. % rH V I ' ; f 1 % o E v-J V."-'^ M JJ O' s O t I ^-, " . -i-i>_- wr^.j V c- 1 C C fcO CVJ rH cd A ( % o H IS H> C a) 0) a> ^ o, 0) o o OT fl) I! GENERAL LIBRARY U.C. BERKEL H5$ ' M