Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am.

419

Multiplication noise in the human visual system at threshold:

1. Quantum fluctuations and minimum detectable energy

Malvin Carl Teich, Paul R. Prucnal, Giovanni Vannucci.*

Michael E. Breton, t and William J. McGillf
Columbia University, New York, New York 10027

Received June 11, 1981; revised manuscript received November 17, 1981
We have carried out a series of frequency-of-seeing experiments similar to those performed by Hecht, Shlaer, and
Pirenne [J. Gen. Physiol. 25, 819-840 (1942)], using an Ar+ laser operated at 514.5 nm as the source of light. In cer-
tain blocks of trials, our subjects were encouraged to report as seen those trials in which the stimulus might have
been present. It was determined that sensitivity and reliability were traded against each other over a broad range:
for our four subjects, the detection of 147 photons at the cornea with 60% frequency of seeing entailed, on the aver-
age, a 1% false-positive rate (FPR), whereas the detection of 34 photons at the cornea with 60% frequency of seeing
was accompanied by a 33% FPR. A new neural-counting model has been developed in the framework of signal-de-
tection theory. It combines Poisson stimulus fluctuations with additive and multiplicative neural noise, both of
which are known to be present in the visual system at threshold. The resulting probability-of-detection curves,
derived from the Neyman Type-A counting distribution, are in good accord with our experimental frequency-of-
seeing data for sensible values of the model parameters. We deduce that, on the average, our four subjects are able
to detect a single photon at the retina with 60% frequency of seeing, at the expense of a 55%FPR. In Part 2 of this
set of papers [P. R. Prucnal and M. C. Teich, Biol. Cybern. 43,87-96 (1982)], we use the normalizing transform, to-
gether with probit analysis, to provide improved estimates of threshold parameters, whereas in Part 3 [M. C. Teich,
P. R. Prucnal, G. Vannucci, M. E. Breton, and W. J. McGill, submitted to Biol. Cybern.], we consider the effects
of non-Poisson quantum fluctuations.

1. INTRODUCTION to tie them to related measurements. The two-quantum

hypothesis was set forth by van der Velden3 and studied by
Bouman and van der Velden4 and Baumgardt. 5 Blackwell6
The classic experiment of Hecht, Shlaer, and Pirenne' (HSP), developed a comprehensive theory related to Crozier's 7 early
carried out at Columbia University in 1942, needs little in- neural hypothesis.
troduction to an audience of visual scientists. Although the A number of researchers considered the importance of
first measurement of the minimum light energy required to various kinds of additive noise in limiting visual-system per-
elicit a response from the human visual system was made by formance,8' 9 and signal-detection theory (SDT) was applied
Langley some 50 years earlier, 2 HSP conducted their experi- to these problems.10-' 6 Multiple-channel models15' 17
were
ments under optimal conditions for seeing and approached 8
presented by Nachmias and Kocher1 and by Cohn et al. ,19
the problem in two distinct ways. On the one hand, they who incorporated into their SDT models the additional effects
performed a straightforward photometric measurement of the of channel uncertainty and location/temporal uncertainty,
average number of photons required at the cornea for 60% respectively. Some workers in the field' 5"18' 20' 2 ' introduced
frequency of seeing. They concluded that this number lay the use of rating-scale (viz., M-ary detection and estimation' 6 )
between 54 and 148 quanta for their seven subjects, corre- formats in their experiments in place of the yes-no (nonor-
sponding to a range of 5-14 quanta at the retina, after ac- thogonal binary) format used by HSP. A number of review
counting for estimated losses as best as they were known in articles detailing most of these efforts have been written over
1942. On the other hand, they created a simple model for the years. 22 -2 6
retinal response based on the Poisson statistical fluctuations Most closely related to our approach, however, are the
of the small numbers of photons expected to be absorbed quantitative models for threshold visual detection proposed
there. This led them to conclude, from an independent by Barlow 27 in 1956 and by McGill2 8' 29 in 1967. Both are
measurement of the shape of the frequency-of-seeing curve, based on Poisson stimulus fluctuations and SDT, but the
that the number of photons absorbed by the retina lay be- character of the noise is assumed to be different in each. In
tween five and eight. This close agreement has served to Barlow's model, the Poisson noise is considered to arise pe-
support the validity of their model, and the notions introduced ripherally and to be additive, whereas in McGill's model it is
by HSP have pervaded visual theory ever since. It is impor- considered to arise centrally and to be multiplicative. In 1977,
tant to note, however, that the actual number of photons ab- Barlow30 also discussed the importance of including the effects
sorbed by the retina has never been directly determined. of central noise in dealing with threshold detection. The
At about the same time, on the other side of the Atlantic, model that we develop here incorporates both of these sources
similar experiments were reported by van der Velden. 3 Many of noise. It combines Poisson stimulus fluctuations, additive
different hypotheses were developed by various researchers Poisson noise, and multiplicative Poisson noise, all of which
in an attempt to explain the frequency-of-seeing results and are known to be present in the visual system at threshold, in

0030-3941/82/040419-13$01.00 © 1982 Optical Society of America

420 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

an SDT framework. It has been designed to be consistent threshold necessaryto turn the light modulator control on was
with the character of the neural noise that is measured phys- exceeded. By selecting the amplitude and dc bias of the tri-
iologicallyat various stations along the optic pathway. angular wave and the pulse height, various modulation depths
We have conducted a series of frequency-of-seeing experi- (M) and average light intensities could be generated. In the
ments, similar to those performed by HSP, using an Ar+ laser experiments reported in this paper, M was set equal to zero
operated at 514.5 nm as the source of light. In certain blocks so that the acousto-optic device passed 1-msec pulses of un-
of trials, we encouraged our subjects to report as seen those modulated laser light of adjustable average energy. Experi-
trials in which the stimulus might have been present, without ments in which M was not equal to zero are reported in Part
regard to the possibility of reporting a false positive. We 3 of this series of papers. 3 2
thereby obtained sets of experimental data with greatly Part of the light output from the laser was removed by a
varying false-positive rates, enabling us to demonstrate that beam splitter and detected by a silicon photocell (see lower
the model parameters extracted from our data are physio- portion of figure). The photocell output formed the input to
logically sensible. a feedback circuit that stabilized the light output of the laser
In Section 2, we describe the experimental setup and par- to ensure the absence of trial-to-trial and day-to-day varia-
adigm. The results of our experiments are provided in Sec- tions in average light intensity for a given input setting.
tion 3, in which we present frequency-of-seeing curves ob- Another part of the light output from the laser was polarized
tained under conditions of varying reliability of response, and attenuated sufficientlythat photon-counting experiments
which is a critical determinant of the average number of could be performed by using an RCA Type 8575 photomul-
quanta required at the cornea for 60% seeing. Sources of noise tiplier tube, followed by a discriminator, amplifier, and pulse
arising from stimulus fluctuations, and from additive and counter. In this way,the Poisson distribution of photons was
multiplicative noise in the visual system, are described in measured directly.
Section 4. This permits us to frame a plausible model for Absolute photometric calibrations were made using an
threshold visual detection that treats these fluctuations in EG&G radiometer with a silicon photodiode at the front end,
conjunction with a decision strategy based on SDT. The substituted in place of the subject's eye. These photocells are
development presented in Section 4, in turn, allows us to ex- stable, providing an accurate measurement of the mean
tract estimates for various model-dependent physiological number of quanta at the cornea. Three subjects' switches
parameters from our frequency-of-seeing data (e.g., the av- were used to start the trial and to indicate yes and no re-
erage number of photons absorbed at the retina), and we carry sponses. An electronic shutter was opened just before, and
this out in Section 5. The conclusion is presented in Section closed just after, the presentation of a stimulus to minimize
6. the transmission of stray light to the subject. A computer
In Part 2 of this set of papers,3 ' we use the normalizing recorded, on a trial-by-trial basis, the subject's response, the
transform, and probit analysis, to provide improved estimates input voltage to the light modulator, the output of the pho-
of threshold parameters. In Part 3,32we consider quantum tocell, and the output of the pulse counter.
fluctuations for non-Poisson stimuli. A preliminary report
of our experimental results was presented in abstract form in B. Subjects
1979.33 Subjects were four males ranging in age from 23 to 38 years.
Subjects PRP and MEB wore corrective lenses during ex-
2. EXPERIMENT perimental sessions. Subject MEB was an experienced psy-
chophysical observer; the others were not. All subjects ap-
A. Equipment peared to have normal visual response under scotopic condi-
An abbreviated schematic representation of the experimental tions as judged by the results of preliminary testing.
apparatus used to generate stimuli and to record and tabulate
responses is presented in Fig. 1. With the left eye, the subject C. Subject Alignment
viewed a small, dim, red fixation target produced by a Max- Subject alignment in the apparatus was maintained by the use
wellian system (not shown). The 5' disk-shaped light stim- of a dental-impression mouthbite. Initial alignment was
ulus, presented at 17.5° horizontal eccentricity on the tem- achieved by using the followingprocedure. By previous ad-
poral retina, was produced by a laser that was viewed directly justment, the target laser beam was made to pass through the
(see lower portion of figure). The source was a feedback- focal point of the Maxwelliansystem that produced the fixa-
stabilized Spectra-Physics Model 162 Ar+ ion laser, which tion target. The position of this point in space was measured
produced plane-polarized radiation at 514.5 nm, in the blue- and marked by a piece of transparent plastic held in place by
green region of the spectrum, by oscillation on several longi- a clamp mounted on the wall. The plastic marker was used
tudinal modes.3 4 The light was attenuated by using neu- during alignment but removed during experimental sessions.
tral-density filters (not shown) before it entered the subject's The mouthbite was then adjusted to place the center of the
eye. A sheet polarizer was the final element in the optical subject's pupil at approximatelythe focalpoint marked on the
path to ensure that the stimulus sent to the subject was plane plastic. Fine horizontal adjustment was then carried out by
polarized. having the subject look to the left and to the right of fixation
The light intensity was controlled by an acousto-optic and mark the points at which the target beam disappeared.
modulator driven by the system shown in the upper portion Since disappearance was caused by interception of the beam
of Fig, 1. A signal generator produced a continuous triangular by the margin of the iris, equal left and right intervals ensured
wave with a frequency of 2 Hz. When the subject pressed the reasonable centering of the beam in the pupil. At the same
start-trial switch, a rectangular pulse with a duration of 1 msec time, a similar series of vertical adjustments was carried out.
was added to the triangular wave in such a way that the Before each experimental session, the fine-adjustment pro-
Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 421

Fig. 1. Schematic representation of the experimental apparatus. The optical arrangement is illustrated in the lower portion of the figure.

cedure was used as a check on alignment. The need for sub- experiments, the upper limit of a set of intensities was es-
sequent adjustments was infrequent. tablished such that the subject always saw a light flash pre-
sented at this level. The subject was then instructed to report
D. Presentation Format every flash seen in a session, being cautious not to report false
A trial consisted of the presentation of a 1-msec flash of light, positives. A subject was considered familiar with the task
at the laser wavelength of 514.5 nm, or of a blank trial. A flash when he maintained a stable, low-false-positive rate while
was chosen quasi-randomly from one of 10 mean energy levels, maintaining 100%frequency of seeing at an intensity level
separated by 0.115 log unit, yielding a total range of about 1 near his original maximum. Familiarity was typically ob-
log unit. A block of trials contained 5 trials at each mean level tained in two to five sessions.
plus 10blank trials, for a total of 60 trials. Subjects initiated For experimental sessions in which a high false-positive rate
a trial by pressing the start-trial switch. A trial followedthe was required, subjects were simply asked to be more daring.
switch press by an interval of about 1 sec. Subjects were re- They were asked still to report every flash that was seen and
quired to press either the yes or the no switch on each trial. to respond negatively on trials in which nothing was seen.
The interstimulus interval was typically maintained at 3-5 However, on trials in which a stimulus might have been
sec. Time between blocks of trials averaged about 3-4 min present, subjects were encouraged to make positive responses
and was governed by computer constraints. without regard to the possibility of producing a false positive.
In this way, false-positive rates ranging up to 33.3%were ob-
tained. High false-positive rates (FPR's) were observed
E. Experimental Session previously in a number of vision experiments. 1 2
l5,18.2 2
0 1
A session began with 35 min of dark adaptation interrupted
only by two to four blocks of preliminary experimental trials.
These dark-adaptation trials were used to allow the subject 3. EXPERIMENTAL FREQUENCY-OF-SEEING
to become reacquainted with the task and to check for mal- CURVES WITH POISSON LIGHT
functions. From 35 min in the dark until completion of the Frequency-of-seeing data generated by the four subjects
session, blocks of trials, alternating with the 3-4-min rest (PRP, MEB, GV, and MCT) with low and high FPR's are
periods, were presented. An experimental session typically presented as the filled circles in Figs. 2 and 3, respectively.
consisted of six blocks of trials and lasted from 1 to 1.5 h. The ordinate represents frequency of seeing on a linear scale
Subjects knew the proportion of blank trials included in the (unity represents 100%frequency of seeing), and the abscissa
presentation. They were allowed additional rest time if re- represents, on a logarithmic scale, the average number of
quested. photons delivered to the cornea in the 1-msec flash, E, as
photometrically measured with the radiometer. The open
F. Preliminary Training circle at the left in each set of data is the frequency of seeing
All subjects were allowed to gain familiarity with the task in for zerophotons at the cornea and represents the experimental
preliminary sessions. During the training period, all subjects FPR, PF. The vertical bars surrounding each data point are
attempted to achieve a near-zero false-positive rate. To es- the 41 - a brackets, obtained by assuming that the trials at
tablish the range of mean intensities to be used in the main any given mean energy level are identical and independent,
422 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

1.0
-(a) POISSONLIGHT -

0.8 SUBJECT PRP

LOW FPR
0.6 I=t 28-

0
z
0.41 '12
y a =0.5 -
(<d>=30
-

0.2
Cd) -i, "II
Li- I . . l
0
I | I ' I II - 1 ,
C) - (d) POISSONLIGHT
z SUBJECT MCT -
LUJ
LOW FPR
-- ---- -- --- - ------ --- t

i/ 28
IL I a=0.5
/ '= O0.2-
K<d>= 30

- I , !
0 30 50 100 200 300 C) 30 50 100 200 300

AVERAGE NUMBER OF PHOTONS AT THE CORNEA (E)

Fig. 2. Frequency-of-seeing data (filled circles) generated by the four subjects (PRP, MEB, GV, and MCT) with a Poisson light stimulus and
low false-positive-rate conditions. The left-most open circle for each subject represents the experimental false-positive rate PF, and the vertical
bars are the ± 1 - a brackets discussed in the text. The dashed lines represent 60%frequency of seeing. The solid curves and crossesare the-
oretical probability-of-detection curves derived from a single family of summated NTA distributions (a =0.5, 7 = 0.2, (d) = 30), with the best-
fitting threshold value t specified in each quadrant of the figure (see Section 5 of text). Numerical values for the theoretical false-positive
probability PF and the experimental false-positive rate PF (with error brackets), as well as the sum-of-squares goodness-of-fit measure, are
presented in Table 1.

1.0 . . I ,I I I II
(a) POISSONLIGHT (b) POISSONLIGHT
0.8 - SUBJECTPRP - SUBJECT MEB I
HIGH FPR HIGH FPR
0.6 ------- I =22
a=0.5
K -- --. -- -I t 22 -
a =0.5
/ =0.2
- -I
(<d>=30 l <d>= 30

LJJ
0.2
z-1
LUO2
-, -..,| . -
.L- 0l.
.
0 10 C0 30 50 100 200 300
*. (c) POISSON LIGHT

Cy 0.8 -- -- I .

Ld <d> 30 I
11~~ =0.5a

SUBJECT GV
0.2 HIGH FPR

0
0 30 50 100 200 300
AVERAGE NUMBER OF PHOTONS AT THE CORNEA (E)
Fig. 3. Frequency-of-seeing data (filled circles) generated by three subjects (PRP, MEB, and GV) with a Poisson light stimulus and high
false-positive-rate conditions. The left-most open circle for each subject represents the experimental false-positive rate PF, and the vertical
bars are the + 1 - a brackets discussed in the text. The dashed lines represent 60%frequency of seeing. The solid curves and crossesare the-
oretical probability-of-detection curves derived from the same family of summated NTA distributions displayed in Fig. 2 (a = 0.5, 11= 0.2, (d)
= 30), with the best-fitting threshold value t specified in each quadrant of the figure (see Section 5 of text). Numerical values for the theoretical
false-positive probability PF and the experimental false-positive rate PF (with error brackets), as well as the sum-of-squares goodness-of-fit
measure, are presented in Table 1.

with two possible outcomes (each therefore represents a The solid curves and crosses represent theoretical proba-
Bernoulli trial with constant probability of success). The bility-of-detection and false-positive probabilities, respec-
error brackets must be considered as lower limits, however, tively, and will be discussed in detail subsequently. For the
because the data most often represented a combination of two moment, let us consider the solid curves to be empirical
different experimental sessions, and the condition of the functions fitted to the frequency-of-seeingdata. Then, from
subject can change from session to session. the low FPR data in Fig. 2, it is apparent that, for 60% fre-
Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 423

quency of seeing (dashed lines), the average number of pho- Barlow's model2 7 provided a concrete mechanism for
tons at the cornea lies between 127 and 147 for our four dealing with nonzero false-positive reports and, at the same
subjects. This is at the high end of the HSP range. Turning time, demonstrated that detectability and reliability could
now to the high-FPR data in Fig. 3, the average number of be viewed as being traded against each other. In addition to
photons is seen to lie between 42 and 82, which is at the low the parameters t and a, the average number of additive dis-
end of the HSP range. The low-FPR data go hand in hand crete noise events confusable with stimulus events, (d) (L 10),
with a high minimum-detectable average energy, whereas the had to be included as a descriptive parameter ((d) is referred
high-FPR data are accompanied by a low minimum-detect- to the retina). Incorporating such additive noise makes good
able average energy. physiologicalsense because of the existence of the spontane-
It has long been known that sensitivity of detection and ous neural discharge 3 5 (see Appendix A).
2 1 27
reliability of response are traded against each other ' ; from Multiplicative noise events are also important in the visual
our experiments we see that this occurs over a broad range of system at threshold. We argue in Appendix A that the Ney-
both variables. In Section 5 we obtain a quantitative sensi- man Type-A (NTA) pulse-number distribution will satis-
tivity-versus-reliability curve for our four subjects. factorily represent the neural-counting distribution for such
processes. This is the point of view taken by McGill.28; 29
More recently, Lillywhite36 also emphasized the importance
4. MODEL FOR THRESHOLD DETECTION
of multiplicative noise in visual information processing.
Every student of visual science knows that one of the principal We proceed to translate these notions into a quantitative
conclusions drawn by HSP1 was that the Poisson variability model for the processing carried out by the visual system in
of the incident photon flux assumes a critical role in deter- an HSP-type task. It is designed to incorporate a variety of
mining the shape of the psychometric function at threshold. old and new experimental results: Poisson stimulus fluctu-
The model constructed by these authors to explain their ex- ations, high photometric quantum efficiency, central-pro-
perimental results assumes that some number of photons, say, cessing loss, additive noise suggested by the spontaneous
seven or so, is integrated together in the peripheral retina, in retinal discharge, multiplicative noise exemplified by the
a narrow region of space and time, to send a message to higher clustering of neural spikes at the retinal ganglion cell, and
centers in the brain that a stimulus had been detected. This single-threshold processing.37' 38 The predictions of the model
is essentially a noiseless conception of visual detection (the turn out to be quite consistent with our experimental fre-
false-positive probability is zero), in which the retina simply quency-of-seeing data, even when only a single parameter of
mirrors the Poisson fluctuations of the stimulus. The im- the model is permitted to vary.
portant parameter derived from the data in this model is the The block diagram of the model is presented in Fig. 4. We
number of quanta required to be absorbed by the retina to begin with a number of preliminary comments about the light
elicit the sensation of seeing, t (t5-8). The quantum effi- source, which consists of a Poisson photon generator, a shut-
ciency of the visual process, n ( _5-10%), can be inferred from ter, and an optional light modulator. A discussion of the
t and from a knowledge of the average number of photons at modulated radiation, which is represented by the upper light
the cornea. path in Fig. 4, is deferred until Part 3 of this set of papers. 32
Although the version of the HSP model that usually comes The Poisson photon generator, gated by the shutter, produces
to mind ignores all sources of fluctuations save those ascrib- bursts of photons describable by a simple nonstationary, ho-
able to the stimulus, HSP did allow for the qualitative intro- mogeneous Poisson point process (lowerlight path, denoted
duction of a kind of central noise through "biological vari- by PP), which gives rise to the well-knownPoisson counting
ability." After all, a variability in the threshold of the or- distribution
ganism is associated with uncertainty or noise. (It will not
provide a nonzero false-positive probability, however.) p(jjE) = EjeE/j!. (la)
RETINA
OPTIC
IRODS NERVE
CORNEA,

POISSON PP OPTIONAL
` , CNR DSNTA XtYE
PHOTON _` LIGHT ' PROCESSING XTHRESHOLD - YES
GENERATOR MODULATOR'
AND CONIG TEST - *NO
SHUTTER

<j> <s>=ar7E
<n>=a<d>
< x =az(BE+<d >)

| LIGHT SOURCE ,, OCULAR { RETINALI CENTRAL LOSS .I OPTIMAL DECISION .

I LOSS -1NOISE I AND NOISE I

Fig. 4. Block diagram of model for visual system processing at threshold. The lower light path excludes the optional light modulator. For
a - 0, with (s), (n), and (x) fixed, the model becomes identical with that proposed by Barlow2 7 ; for (d) - 0, with a and (s) as parameters,
the model is equivalent to that presented by McGill2 8' 29 ; whereas for (d) - 0 and a - 0, with (s) and (x) fixed, the model reduces to that presented
by HSP.1 A discussion of the upper light path is considered in Part 3 of this set of papers. 3 2
424 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

Here p(jIE) represents the probability of finding j photons efficiency and with the estimated central efficiency (-50%)
in the stimulus duration T at the cornea when the integrated reported by Barlow 30' 42 (see Appendix B). Since a < 1, the
intensity (energy)E is fixed. The quantity E = IAT, where effects of such NTA multiplied-Poisson noise are manifested
I is the light irradiance (intensity) and A is the (uniformly) as a decrease in the average count rate (this is designated as
illuminated area. The mean number of photons at the cornea central loss in Fig. 4) and a concomitant increase in the ob-
(j) = E, as is evident from inspection of Eq. (la). served variance-to-mean ratio (designated as central noise in
The light pulse incident upon the cornea traverses the oc- Fig. 4). We will see subsequently that the character of our
ular medium and the retina, undergoing loss through ab- results does not depend on the precise value of a.
sorption and scattering, and is absorbed by the rods. In our The mean, variance, and variance-to-mean ratio for the
case, the quantum efficiency ij is defined as the fraction of NTA distribution are, respectively,4'
photons incident upon the cornea that survivethe passageand
actually contribute to the activation of the rods; we refer to (z) = aW, (3a)
this as the ocular quantum efficiency. The best current
((AZ)2) = (1 + a)aW, (3b)
photometric estimate for this quantum efficiencyin humans,
30
calculated by Barlow, lies between 0.11 and 0.33. In spite ((A\Z)2)/(Z) = 1 + a. (3c)
of the absorption and scattering in the intervening medium,
the statistics of the photons absorbed by the rods remain Note that the variance is proportional to the mean, so that the
Poisson; the loss simply reduces the mean. At the retina, the noise looks like excess shot noise.4 3 The variance-to-mean
number of photons m registered in time T is therefore de- ratio is alwaysgreater than unity, whereasthis ratio is always
scribed by the Poisson distribution precisely unity for the Poisson distribution. This indicates
that the multiplication of the signal alwaysintroduces excess
p(ml nE) = (nE)me-E/m!. (lb) shot noise. Some years ago, Grimm 4 4 compiled tables of the
cumulative density function for various values of (z) and a.
Clearly, the mean number of photons per flash (m) = qE.
The typical shapes that the distribution takes on are illus-
Equation (lb) provides an appropriate statistical description
trated in Fig. 1 of the paper by Teich.4 ' It is important to note
for the detection of most kinds of light used in vision experi-
that the NTA approaches the Gaussian as a limiting form4 l
ments, including the dc-excited ribbon-filament tungsten
since many models for visual detection postulate a Gaussian
lamp used by HSP and the (unmodulated) multimode Ar+
internal variable.6
laser used in our experiments.39
Finally, the output of the central-processing and -counting
Following the reasoning presented in Appendix A, we now
center is subjected to a likelihood-ratio test, which comprises
use the NTA distribution, a relatively simple counting dis-
an optimal processing strategy in the framework of signal
tribution associated with a clustered point process, to describe
detection theory.10-''21 5' 16 Recently, Prucnal and Teich
the statistical character of the individual events at the output
showed that, under a broad range of conditions, the likeli-
of our central-processing and -counting center in darkness and
hood-ratio test for the classical binary-decision problem re-
at low light levels. Most simply, we may conceive of this
duces to a simple comparison of the number of counts with a
distribution as arising from a Poisson stimulus distribution
single threshold. 38 This is a useful result in the context of
of mean W, p(ml W), associated with the signal or dark light,
visual information processing, as we emphasized previously.3 7
driving an independent Poisson neural-counting distribution
It means that, to achieve optimal processing, the observer need
associated with the retinal and central signal, p(zlm). Since
not tabulate the results of repeated trials to accumulate signal
any number of stimulus events m may occur, and since all
and noise relative-frequency histograms and then carry out
values of m contribute to the neural count, we sum over this
a complex calculation of the likelihood ratio.
variable.
The NTA distribution is written as28'40'41 Four parameters appear in our model: t, -q,(d), and a. In
mathematical terms, the detection problem is formulated as
-(am)ze-am Wme-W follows. At the counting center, the total number of noise
P(zlW) = E p (zIm)p (ml W) = E
m=0 mM Z m! events registered in the central-counting interval Tc is rep-
resented by an NTA discrete random variable n, with prob-
(2a) ability density function PN(nl (d) ), since the dark light at the
p(01 W) = exp[W(ea - 1)], (2b) retina is assumed to be Poisson with mean (d). Thus the
noise mean at the counting center (n) = a(d) (see Fig. 4).
with W > 0. The noise distribution is obtained by setting z For the values of (n) and a that we will encounter, log
= n and W = (d), where (d) is the average Poisson dark-noise [PN(nl (d))] will exhibit no point of inflection, so that sin-
count (referred to the retina) in an integration time. The gle-threshold processing will be optimal. 38 The total number
signal distribution emerges when z = s and W = i7E (see Fig. of signal events at the counting center, in the central-counting
4). The signal-plus-noise distribution is the discrete convo- interval Tc, is similarly represented by a NTA discrete random
lution of the independent signal and noisedistributions. The variable s with probability density function ps (slNE). The
multiplication parameter a (>0) provides a measure of the average number of photons at the cornea is E, so that the av-
average number of neural spikes per effective (light or dark) erage number of photons activating rods in the retina is (im)
photon, assuming that the entire discharge cluster (which will = -E, and the signal mean at the counting center is (s) =
have duration rP - 50-70 msec for light pulses that are of ao(m) = a-qE (see Fig. 4). If the noise and signal random
<32-msec duration; see Appendix A) is included in the cen- variables are additive, independent, and noninterfering, then
tral-counting interval Tc. We have set a = 0.5 to provide whatever their individual statistical character, the total
results in accord with the estimated photometric quantum number of signal-plus-noise events registered at the central-
Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 425

- 10
41 -
0. - - 2 -
a;

23025
2

0. Far=0.5
LL -
04- '(,481

M.1624f
2

> (C)
/ 22 /0 / /-. Cd 3 7=0.5
018-

o..
co0.6 -
00 IO 30
0I0
7 0.36

I0 I IOI0 I300

, t 0=0.5
30 50

0.2
ALL0.2
A08 =26 t ae t
(d 30,and asa2pr3 t 7=
0 30 50 100 200 300 0 30 50 100 200 300
AVERAGE NUMBER OF PHOTONS AT THE CORNEA (E)
Fig. 5. (a) Theoretical probability of detection PD (derived from summated NTA distributions) versus average number of photons at the cornea,
E, withthe thresholdcounttas aparameter. The multiplication parameter a =0.5, the quantum efficiency? = 0.2, and the retinal dark-light
mean (d) =30. (b) a =0.5, j= 0.2, t =26, and (d) as aparameter. (c) a =0.5, t =26, (d)=30, and-nasa parameter. (d) tq= 0.2, t =26,
(d) = 30, and a as a parameter.

counting center, in the interval T>, is represented by a discrete respectively (see Fig. 4). The theoretical probability-of-
random variable x, with probability density function detection curves (PD versus E) are, in accordance with Eq. (6),
PSN(XI (Kd),NE) given by obtained from summated NTA distributions. The receiver-
operating characteristic (ROC) is obtained from Eqs. (5) and
PSN(XI (cd),qE) = PN(nj ((d))* ps (slnE). (4) (6).
In Fig. 5, we present computer-generated families of such
The symbol * represents discrete convolution.
theoretical probability-of-detection curves versus the average
The theoretical false-positiveprobability PF represents the
number of photons at the cornea, E. In each quadrant of the
probability that the noise count n is greater than, or equal to, figure, one of the parameters that enters our model is varied
the threshold t, i.e.,16 parametrically while the remaining parameters are kept
constant. Thus in Fig. 5(a) we fix a = 0.5, tj = 0.2, and (d )
PF=EPN(nj (d )) (6) = 30 (we have chosen these values for reasons explained in
n=t
Appendix B). In this set of curves, the threshold count t is
The theoretical probability of detection PD represents the varied parametrically; in Section 5 we will make use of Fig.
probability that the signal-plus-noisecount x is greater than, 5(a) in fitting the experimental frequency-of-seeing curves
or equal to, the threshold t, i.e.,16 reported earlier. The crossesat the left-hand boundary of the
figure (zero photons at the cornea) represent the false-positive
PD(E) = E PSN(XI(Kd),nE). (6) probabilities PF for each of the 10 curves.
x=t In Fig. 5(b), parameters are fixed at a = 0.5, 7i= 0.2, and t
26, and (d) is varied parametrically. Comparing Figs. 5(a)
In the model considered by Barlow,2 7 both the noise PN
-
(n)
and 5(b), it is clear that increasing (d) has an effect on PD
and the signal ps(s) are Poisson. Since the convolution of two
similar to that of decreasing t, i.e., the curves become in-
Poisson distributions remains Poisson, the signal-plus-noise
creasingly shallow. In Fig. 5(c), ij is varied, with a = 0.5, t =
distribution in that case is
26, and (d ) = 30. The PDcurves appear primarily to trans-
PSN(X) = (x)xe-(x)/x!, (7) late to the left with increasing qj;the false-positiveprobability
PF depends only on a, t, and (d) and is independent of X [see
and the theoretical probability-of-detection curves are, in Eq. (5)]. Finally, in Fig. 5(d), a is varied with i = 0.2, t = 26,
accordance with Eq. (6), obtained from summated Poisson and again (cd) = 30; the curves for increasing a and increasing
distributions. The convolution of two NTA distributions, Xqbehave somewhat similarly, as expected, althouth PF does
with identical multiplication parameters a, can be similarly depend on a. The dependence of the probability-of-detection
shown to remain NTA with parameter a. 40,41 Thus the sig- curves, and the associated false-positive probabilities, on each
nal-plus-noise distribution for the model considered here is of the parameters in our model is thus explicitly demonstrated
described by Eq. (2) with z = x and W = (OE+ (d) ). From in Fig. 5. ROC curves could be presented in the same way. In
Eq. (3), the mean and variance of the signal-plus-noise count Section 5 we use this model to fit our experimental fre-
are (x) = a(nE + (d)) and ((AX)2 ) = (1 + a)a(oqE+ (d)), quency-of-seeing curves for Poisson light.
426 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

5. EXTRACTION OF MODEL PARAMETERS the sum of squares in Table 1 provides a measure of the
FROM FREQUENCY-OF-SEEING DATA goodness of fit of the theory to the experimental data.
Our quick success with this elementary approach simply
There are many approaches that can be employed in at- illustrates that there is at least one sensible set of average
tempting to fit to the frequency-of-seeingdata the theoretical parameters (a, in, (d)) that, when plugged into our model,
probability-of-detection curves generated by our model. The provides a theory that is reasonably consistent with the data
simplest approach,perhaps, is to fix all parameters at constant from all the subjects. Although we do not believe that these
values, save the threshold count t, which we permit to vary parameters are truly the same for all subjects, this assumption
parametrically. There then results a family of one-parameter has provided us with a point of departure without getting us
probability-of-detection curves, such as that shown in Fig. into trouble (as it likely would if our data were more accu-
5(a), with each curve representing a different value of t. rate).
Using this method, we were able to fit, reasonably satisfac- We next proceeded to fit each of the seven sets of data in
torily, all seven sets of frequency-of-seeing data presented in Figs. 2 and 3 individually by means of a slightly different
Figs. 2 and 3 (for both low and high FPR's). The particular one-parameter procedure in which a and 11were fixed at 0.5
family of curves illustrated in Fig. 5(a) (a = 0.5, X = 0.2, (d) and 0.2, respectively, while (d) and t were simultaneously
= 30) was used in this procedure. The solid curves and crosses adjusted both to match the theoretical and experimental
in Figs. 2 and 3 all belong to this theoretical family, with false-positive rates (PF and PF) and to minimize the sum-
best-fitting threshold values ranging between t = 18 and t = of-squares measure of fit. The theoretical probability-of-
28, as specified in the figures and in Table 1. Also presented detection curves in this case are derived from different families
in Table 1 are the theoretical false-positive probability PF and of summated NTA distributions. The results are illustrated
the experimental FPR PF (with ±1 - a error brackets) for in Table 2, where it can be seen that the best-fitting values of
each curve. These are as consistent as can be expected, given the threshold obtained by this method range between t = 18
the magnitude of the error brackets. The column containing and t = 32. These values are quite similar to those obtained

Table 1. One-Parameter (t) Collective Fit to All Seven Sets of (Poisson-Light) Frequency-of-Seeing Data with
Theoretical Probability-of-Detection Curves Derived from a Single Family of Summated Neyman Type-A
Distributionsa
Theoretical Experimental
Best-Fitting False-Positive False-Positive
Subject/ Threshold Probability Rate Sum of
Condition Countt PF (%) PF (%) Squares
PRP (low FPR) 28 1.0 1.0 4 0.9 0.0154
MEB (low FPR) 26 2.2 2.2 i 1.3 0.0451
GV (low FPR) 26 2.2 0.8 + 0.8 0.0546
MCT (low PPR) 28 1.0 1.5 + 1.1 0.0122
PRP (high FPR) 22 9.2 10.8 + 2.8 0.0331
MEB (high FPR) 22 9.2 14.0 + 3.2 0.0175
GV (high FPR) 18 28.0 33.3 + 4.3 0.0435

I The threshold countt that provides the best fit is indicated in the second column,when the remaining parameters are fixedat a = 0.5,i = 0.2,and (d) =. 30. These
theoreticalprobability-of-detectioncurvesare shownas the solid curvesin Figs.2 and 3. Alsotabulated, in the third and fourth columns,respectively,are the theoretical
false-positive probability PF and the experimental false-positiverate PF (with error brackets) as wellas the sum-of-squares measure of fit of theory to experiment
(fifth column). The theoretical false-positive probabilities are indicated by crosseson Figs. 2 and 3. The number of blank trials b = 120 for each set of data.

Table 2. Individual Fits to Each of the Seven Sets of (Poisson-Light) Frequency-of-Seeing Dataa
Theoretical Experimental
Best-Fitting Best-Fitting False-Positive False-Positive
Subject/ Dark-Light Mean Threshold Probability Rate Sum of
Condition Count (d) Count t PF (%) PF (%) Squares
PRP (low FPR) 32 29 1.1 1.0 + 0.9 0.0176
MEB (low FPR) 32 27 2.5 2.2 + 1.3 0.0498
GV (low FPR) 16 18 1.0 0.8 I 0.8 0.0198
MCT (low FPR) 38 32 1.6 1.5 ± 1.1 0.0196
PRP (high FPR) 36 25 11.0 10.8 + 2.8 0.0406
MEB (high FPR) 48 31 14.0 14.0 i 3.2 0.0349
GV (high FPR) 56 31 34.0 33.3 i 4.3 0.0407

a The values of the dark-light mean count (d) and the threshold t that providethe best fit for a givensubject/conditionare indicated in the second and third columns,
when the remaining parameters are fixed at a = 0.5 and n = 0.2. The theoretical false-positiveprobability PF was constrained (in the fitting procedure) to be ap-
proximatelyequal to the experimentalfalse-positiverate PF,as is apparent from the fourth and fifth columns. This method thereforealsorepresents a one-parameter
fit. Comparisonof the sum-of-squares measure of fit of theory to experiment (sixth column) with the equivalent column in Table 1 (fifth column) demonstrates
that the fits are about equally good. These theoretical curves are derived from different familiesof summated Neyman Type-A distributions and are not shown
in Figs. 2 and 3. The number of blank trials b = 120for each set of data.
Teichet al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 427

range. With an average of five photons at the cornea, for

example, 60% frequency of seeing is accompanied by a 55%
false-positiverate. But with an average of 140photons at the
20 cornea, 60% frequency of seeing is accompanied by a false-
positive rate that is only about 1%. We must keep in mind,
40
however, that our experimental data extended only to a 33%
FPR. The validity of the extrapolation beyond this point is
based on the assumption that the set of probability-of-de-
I- 60 tection curves presented in Fig. 5(a) is representative of fre-
quency-of-seeing data for t < 18.
z 80- Furthermore, insofar as we accept the 20% model-depen-
I-' dent value for the ocular quantum efficiency 11(admittedly
uJ this is a rough estimate), we can translate our minimum cor-
100
CD neal energies into minimum retinal energies. The right or-
Cf) dinate of Fig. 6 illustrates this. Improved estimates of the
wL 120 model parameters would permit us to construct a personal
Qt F
U) sensitivity-reliability curve for each subject.
2 14 We have also been able to fit the experimental frequency-
of-seeing data presented in Figs. 2 and 3 with theoretical
probability-of-detection curves derived from Barlow's
16
model,27 in which both the signal and noise are Poisson, as well
as with theoretical PD curves derived from a variation of this
FALSE-POSITIVERATE(%/ model, in which the signal is NTA and the noise is Poisson.
INCREASING RELIABILITY-b Using a three-parameter fitting procedure to minimize the
Fig. 6. Minimum-detectable average energy for 60% frequency of
sum of squares, we obtained excellent fits to the data in both
seeing of Poisson light (left ordinate, number of photons at the cornea; cases, but the model parameters associated with the best-
right ordinate, number of photons at the retina) versus false-positive fitting curves are highly variable and appear to show no reg-
rate (%) for four subjects. Sensitivity and reliability are traded ular pattern. The associated theoretical false-positive
against each other over a broad range. With an average of one photon probabilities, furthermore, are almost always smaller than the
at the retina, for example, 60% frequency of seeing is accompanied
by a 55% false-positive rate. But with an average of 28 photons at the experimental FPR's. We have interpreted this as indicating
retina, 60% frequency of seeing is accompanied by a false-positive rate that the effective noise in the visual system near threshold
that is only about 1%. The curve was derived from the family of should properly be represented by a counting distribution
summated NTA distributions that was used to fit collectively all seven with a variance greater than that of the Poisson.45 The NTA
sets of (Poisson-light) frequency-of-seeing data [a = 0.5, '7 = 0.2, (d)
= 30; see Figs. 2, 3, and 5(a) and Table 1]. It must be kept in mind
is just such a distribution.
that our experimental data extend only to a 33%FPR, however(see It is evident from the foregoing that theoretical probabil-
text). A personal curve for each subject could be obtained if longer ity-of-detection curves derived from summated NTA distri-
experimental sessions could be conducted without fatiguing the butions do a good job of fitting experimental frequency-of-
subject. seeing data and provide model parameters of reasonable
values under a broad range of false-positive rates.4 6 But it
by the first procedure. The sum of squares for the two pro- is also evident that the frequency-of-seeing paradigm does not
cedures (fifth column of Table 1 and sixth column of Table appear to permit the kind of precision that will allow the ex-
2) are also comparable, indicating that nothing is seriously perimenter to discriminate sensitively among related alter-
awry. We could, of course, attempt to provide a two- and even native models. There are many ways in which probability-
a three-parameter fit to each data set separately in an effort of-detection curves with virtually identical shapes can be
to extract model parameters for individual subjects under generated, and one cannot conclude from the shape of these
specified conditions. But it appears to us that this level of curves alone that the nature of the process has been deter-
effort is not warranted by the limited accuracy of our data, and mined. Nor will classical frequency-of-seeing data permit the
we are quite satisfied with these simple, but consistent, one- extraction of more than approximate values of the relevant
parameter fits. In Part 2 of this set of papers, 3! we use the physiological parameters, even if it is known with certainty
normalizing transform and probit analysis to provide an im- that a given model is the proper one. Variants of this proce-
proved estimation procedure. dure, in which data are collected at a small number of carefully
Having demonstrated that the shapes and behavior of the chosen average energies, can be useful in this connection,
probability-of-detection curves presented in Fig. 5(a) are however. 3 1
sensible, we now cast the results in a somewhat different form,
to elucidate the trade-off between sensitivity and reliability
6. CONCLUSION
of response. The horizontal dashed line at the 60% proba-
bility-of-detection level, together with the associated crosses Following the approach used by HSP in 1942, we collected
on the left ordinate of Fig. 5(a), immediately allows us to calibrated frequency-of-seeing curves near the threshold of
provide a quantitative measure of the minimum-detectable vision. Our experiments differed from theirs in that high
average number of photons at the cornea versus the false- false-report rates were encouraged in certain blocks of trials.
positive rate for our four subjects, as shown in Fig. 6. Sensi- A quantitative relationship between the average number of
tivity and reliability are traded against each other over a broad photons required at the cornea, for 60%frequency of seeing,
428 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

and the FPR demonstrates that sensitivity and reliability are Finally we note that, aside from the range of physiological
traded against each other over a broad range. data with which our model is consistent, it consists of only a
The statistical properties of various sources of noise at single channel. As such, it is inherently simpler than multi-
threshold (stimulus, retinal, and central) were distilled from ple-channel models.'7-' 9 In short, the effects of neural noise
a number of physical, neurophysiological, and psychophysical can be subtly made to join the effects of Poisson stimulus
experiments. The overall noise has been categorized into two fluctuations, leaving many important results intact. The role
types: additive Poisson and multiplicative Poisson. A of non-Poisson stimulus fluctuations can also be established
neural-counting model for threshold detection was developed. in the framework of such a model, as is considered in Part 3
It incorporates a variety of old and new experimental and of this set of papers.3 2
theoretical results: Poisson stimulus fluctuations, high
photometric quantum efficiency, central-processing loss,
additive noise suggested by the spontaneous retinal discharge, Appendix A. Neurophysiological Evidence for
multiplicative noise exemplified by the clustering of neural Additive and Multiplicative Noise in the Visual
spikes at the retinal ganglion cell, and single-threshold pro- System
cessing. The resulting probability-of-detection curves are In 1971, Barlow et al. 5 5 conducted an important series of ex-
derived from summated NTA distributions. Within the periments on the neural discharge in the retinal ganglion cell
limitations of the paradigm, they are in accord with our fre- of the cat. For a number of on-center cells, they demon-
quency-of-seeing data for sensible values of the model pa- strated that the neural pulse-number distribution (PND),
rameters (dark-noise count, quantum and central efficiencies, both in darkness and in response to brief but dim flashes of
and threshold count). light, exhibited a variance in excess of the mean. Thus the
The internal neural-counting distribution is represented spike discharge could not be understood in terms of a Poisson
in our model by the NTA distribution, both in the dark and process, even if additive Poisson noise were included. They
in response to a stimulus. This counting distribution, and presented a possible anatomical explanation for these obser-
indeed the entire class of counting distributions arising from vations in terms of multiple pathways in the retina from the
multiplied and cascaded Poisson processes, 4l143A4749exhibits rod to the ganglion cell, both for photons and for dark light.
a count variance proportional to the count mean. Thus their 6 4 65
Indeed, Baylor and Yau and their collaborators ' have
fluctuation properties are manifested as excess shot noise43; convincingly shown that the response of vertebrate rod outer
they evidently bear a close relationship to the Poisson. This segments to single photons and to dark photons6 6' 67 takes the
similarity appears in the limits of the distributions as well: form of smooth current pulses (of about 1-pA magnitude) that
the NTA is skewed at low levels, and it approaches the reflect the underlying Poisson excitations. But whatever the
Gaussian at high levels.4' The body of psychophysical and mechanism, the observation of a neural-count variance in
neurophysiological measurements that appear to require these excess of the count mean reveals that the neural spikes form
particular characteristics of the Poisson can, therefore, be a clustered point process and that the fluctuation character-
satisfied by the NTA distribution. In fact, a bit more latitude istics of the incident Poisson photons are not faithfully relayed
is available in the latter case, because the NTA has two free by the retina. Support for this observation is provided in
parameters rather than one. It is interesting that the outcome more recent data, at higher adaptation levels but still not far
of Sakitt's threshold-rating experiments 2 1' 50
can be inter- above threshold, 68
as well as in the study of Levick and
preted as suggesting that the internal decision random vari- 69
Zacks. The latter work demonstrated that, near threshold,
able is discrete and Gaussian-like, with a variance greater than the ganglion-cell spike response exhibits a time course with
the mean.50 The NTA distribution fits this description a minimum duration of 50-70 msec, no matter how brief the
well. flash.
Because of the properties of this distribution, it immedi- Barlow et al. showed that the mean and variance of the
ately followsthat our model gives rise to different increment NTA counting distribution,4 0 '41although they did not refer
and decrement thresholds, 51 that it can be coupled with probit to it by name, adequately accounted for their experimental
analysis to provide improved threshold-parameter estima- observations of these quantities. They were careful to dem-
tion,3 1 and that it leads to the deVries-Rose intensity-dis- onstrate, however, that a number of different mathematical
crimination law52'53 at light levelsthat are low (but abovethe constructs would fit the data. We have recently shown 6 l that
additive noise level).28'29,36'54 We know that the retinal the maintained-discharge interspike-interval histograms for
ganglion cell discharge measured by Barlow et al. 5 5 (for an one of their (on-center brisk-sustained) units could be mod-
on-center, brisk-sustained unit stimulated by short, small light eled as a shot-noise-driven doubly stochastic Poisson point
stimuli) exhibits counting behavior describable by the NTA process (SNDP), modified somewhat by relative refractori-
distribution (see Appendix A). Interestingly, under the same ness. For a counting time T that is long in comparison with
conditions, the intensity-discrimination behavior of such a the linear-filter response time rp in the SNDP model, the
cell is describable by the deVries-Rose law.56 ' 57 At higher expected pulse-number distribution turns out to be precisely
levels of adaptation, the effects of saturation 58 and refracto- the NTA.'1 43, 47 This is true both for the stationary case
riness59 - 61 come into play. The NTA distribution is quite (maintained discharge) and for the nonstationary case, in
robust, however, and, under rather broad conditions, the re- which the system is excited by a pulse of light, provided only
fractoriness-modified counting distribution will remain NTA that the duration of the pulse Tr is short (r <<7p + T).4 8 In
(the refractoriness will simply alter the variance-to-mean the experiments of Barlow et al., 5 5 Tr = 10 msec, p-, 30
ratio).62 Ultimately, however, the count variance will de- msec, and T = 200 msec, so that this condition is well satis-
crease below the count mean, and Weber's law will pre- fied.
vail.63 The NTA distribution has been considered before in the
Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 429

context of visual neurophysiology.7 0 It results from a mul- In whatever way we choose to view the information transfer
41 4 3
tiplicative cascade of two Poisson distributions l but enjoys in the visual system, it is not likely that the NTA counting
a generality that reaches beyond that description. Even in distribution provides a perfect representation for our neural
the presence of refractoriness or dead time, it often provides signal at the hypothetical counting center. Rather, it is a
an appropriate description for the pulse-number distribu- well-understood two-parameter distribution that captures the
tion.6 2 Indeed, we have been able to fit the PND data of essence of the effects that we wish to represent: stimulus
Barlow et al.5 5
for different luminance levels by using a con- fluctuations (and additive dark-light fluctuations) and mul-
volution of two NTA distributions, one representing the tiplicative neural noise.
dark-light discharge and the other representing the response
to a flash of light of specified energy. 7 ' In short, additive
Appendix B. Choice of Model Parameters
Poisson noise, together with multiplicative Poisson noise,
provides a good description for the discharge behavior in (at We briefly discuss our choice of model parameters. Barlow30
72 has estimated that the photometric quantum efficiency-j lies
least some ) retinal ganglion cells at the threshold of vi-
sion. between 11 and 33% for the human. If the more recent results
The axons of the retinal ganglion cells comprising the optic of Baylor et al.65 for the "fraction exciting" in the toad are
nerve travel to central locations in the midbrain (superior applicable, then this range contracts to 11-17%. We used 7)
colliculus) and thalamus [lateral geniculate nuclei (LGN)]. = 20% in our calculations, but the choice is not critical.
The signals continue on to the primary visual (striate) cortex. Barlow4 2 also considers a central efficiency of 50% as a
Fibers project from the visual cortex to associated cortical reasonable value to be expected. He sees this3 0 as the most
areas and back to the superior colliculus and to the lateral likely way of reconciling the 5.5% quantum efficiency from
geniculate nuclei. The spontaneous interspike-interval his- psychophysics' with the 11%lower limit of the photometric
tograms recorded at the LGN often appear to exhibit an excess quantum efficiency. Based on this argument, we used a =
of moderately short intervals relative to the exponential dis- 0.5.
tribution.7 3 ' 74 This indicates spike clustering at the LGN, as For the average dark count, we chose (d) = 30, referred to
6
observed in the ganglion cell and described by the SNDP. ' the retina. This was obtained as follows. We used 0ster-
It is tempting to conjecture that the combination of additive berg's7 9 figure of 1900 rods/0.0069 mm 2 at 17.5° eccentricity
and multiplicative Poisson noise also provides a suitable de- on the temporal retina. The receptive field size in the pe-
scription for the statistical behavior of the discharge at LGN ripheral retina is not well known under scotopic conditions.
cells in darkness and at low light levels. However, extrapolating Wilson and Bergen's80 data to the
In the current state of our knowledge, however, it is hopeless conditions of our experiment yields a representative value of
2 6
to attempt to detail, and to follow microscopically, the sta- 10 (0.088 mm ). We also used the estimate of Baylor et al.. 7
tistical character of the signal along its individual pathways. of 0.008 isomerizations/sec/rod for the human, and a temporal
Instead, let us consider for a moment a h-stage cascade of integration time of 0.2 sec. Combining all these estimates
Poisson point processes, each buffered by a linear filter. We yields 38 isomerizations/summation area/summation time,
have shown4 9 that this produces a clustered output point close to our value (d) = 30. An estimate not too different
process, for which the count variance remains proportional from this can also be inferred from the retinal-ganglion-cell
to the count mean.75 This is true for both the stationary and dark discharge. 6 1
nonstationary cases. We may wish to consider the informa- We note that our probit estimates of all these parameters 31
tion flux reaching the counting center as being carried on a lie in the same range.
parallel set of such channels. Qinlar76 has shown that, under
specified conditions, the superposition of a set of such com- ACKNOWLEDGMENTS
ponent clustered point processes itself converges to a clustered
point process. It seems that, once present, multiplication The authors are grateful to the National Science Foundation
noise arising from clustering remains. This is true both for and to the National Institutes of Health for support of this
the dark discharge and for the response to light. work.
We therefore posit that the statistical distribution of neural * Present address, Bell Laboratories, Holmdel, New Jersey
events, on which a decision is based for the detection of a flash 07733.
of light at threshold, is associated with a clustered (multiplied) t Present address, Department of Visual Physiology, Wills
point process. We use the NTA distribution to capture the Eye Hospital, Philadelphia, Pennsylvania 19107.
essential statistical character of this clustering. t Present address, Department of Psychology, University
The mechanism can be viewed in a simplified way. of California at San Diego,La Jolla, California 92093.
McGill2 8'29 argued that the smearing together of many neural
paths at a hypothetical counting center in the chain to the
visual cortex produces a Poisson-like central noise process REFERENCES
under a broad range of conditions. There is ample mathe- 1. S. Hecht, S. Shlaer, and M. H. Pirenne, "Energy, quanta, and
76
matical support for such a convergence -78; it basically re- vision," J. Gen. Physiol. 25, 819-840 (1942).
quires independence and uniform sparseness for the indi- 2. S. P. Langley, "Energy and vision," Philos. Mag. Ser. 5, 27, 1-23
vidual channels. The Poisson flow of photons from the light (1889).
3. H. A. van der Velden, "Over het aantal lichtquanta, dat nodig is
source then modulates this Poisson central noise process, voor een lichtprikkel bij het menselijk oog," Physica 11, 179-189
leading directly to the NTA counting distribution. This is (1944);"The number of quanta necessary for the perception of
1 52 53
seen to be a hybrid approach; it is in part stimulus based, ' ' light of the human eye," Ophthalmologica 111, 321-331 (1946)
6 7
in part neural based. ' [translation].
430 J. Opt. Soc. Am./Vol. 72, No. 4/April 1982 Teich et al.

4. M. A. Bouman and H. A. van der Velden, "The two-quanta ex- tems," in Concepts of Communications: Interpersonal, In-
planation of the dependences of the threshold values and visual trapersonal, and Mathematical, E. F. Beckenbach and C. B.
acuity on the visual angle and the time of observation," J. Opt. Tompkins, eds. (Wiley, New York, 1971), Chap. 9, pp. 257-
Soc. Am. 37, 908-919 (1947). 281.
5. E. Baumgardt, "The quantic and statistical basis of visual exci- 30. H. B. Barlow, "Retinal and central factors in human vision limited
tation," J. Gen. Physiol. 31, 269-290 (1950). by noise," in Photoreception in Vertebrates, H. B. Barlow and
6. H. R. Blackwell, "Neural theories of simple visual discrimina- P. Fatt, eds. (Academic, New York, 1977), Chap. 19, pp. 337-
tions," J. Opt. Soc. Am. 53, 129-160 (1963). 358.
7. W. J. Crozier, "On the sensory discrimination of intensities," Proc. 31. P.R. Prucnal and M. C. Teich, "Multiplication noise in the human
Nat. Acad. Sci. U.S.A. 22, 412-416 (1936). visual system at threshold: 2. Probit estimation of parameters,"
8. S. Hecht, "Energy and vision." in Science In Progress, A Baitsell, Biol. Cybern. 43, 87-96 (1982).
ed. (Yale U. Press, New Haven, Conn., 1945), pp. 75-95, 309- 32. M. C. Teich, P. R. Prucnal, G. Vannucci, M. E. Breton, and W.
310. J. McGill, "Multiplication noise in the human visual system at'
9. E. J. Denton and M. H. Pirenne, "The absolute sensitivity and threshold: 3. The role of non-Poisson quantum fluctuations,"
functional stability of the human eye," J. Physiol. 123, 417-442 Biol. Cybern. (submitted for publication).
(1954). 33. M. C. Teich, P. R. Prucnal, G. Vannucci, M. E. Breton and W. J.
10. W. W. Peterson, T. G. Birdsall, and W. C. Fox, "The theory of McGill, "Role of quantum fluctuations and the Neyman Type-A
signal detectability," IRE Trans. Prof. Group Inf. Theory distribution in human vision," J. Opt. Soc. Am. 69, 1469(A)
PGIT-4, 171-212 (1954). (1979).
11. W. P. Tanner, Jr., and J. A. Swets, "The human use of informa- 34. M. E. Breton, M. C. Teich, and L. Matin, "Intensity fluctuations
tion-I. Signal detection for the case of the signal known exactly," produced by multimode lasers in combination with dielectric
IRE Trans. Prof. Group Inf. Theory PGIT-4, 213-221 (1954). beamsplitters," Behavioral Res. Methods Instrum. 9, 324-325
12. W. P. Tanner, Jr., and J. A. Swets, "A decision-making theory of (1977).
visual detection," Psychol. Rev. 61, 401-409 (1954). 35. S. W. Kuffler, R. FitzHugh, and H. B. Barlow, "Maintained ac-
13. M. Treisman, "Noise and Weber's law: the discrimination of tivity in the cat's retina in light and darkness," J. Gen. Physiol.
brightness and other dimensions," Psychol. Rev. 71, 314-330 40, 683-702 (1957).
(1964). i 36. P. G. Lillywhite, "Multiplicative intrinsic noise and the limits to
14. M. Treisman, "A statistical decision model for sensory discrim- visual performance," Vision Res. 21, 291-296 (1981).
ination which predicts Weber's law and other sensory laws: some 37. M. C. Teich, P. R. Pruenal, and G. Vannucci, "Optimum photon
results of a computer simulation," Percept. Psychophys. 1, detection with a simple counting processor," Opt. Lett. 1,208-210
203-230 (1966). (1977).
15. D. M. Green and J. A. Swets, Signal Detection Theory and Psy- 38. P. R. Prucnal and M. C. Teich, "Single-threshold detection of a
chophysics (Wiley, New York, 1966) (reprinted by Krieger, random signal in noise with multiple independent observations:
Huntington, N.Y., 1974). 1. Discrete case with application to optical communications,"
16. H. L. van Trees, Detection, Estimation and Modulation Appl. Opt. 17,3576-3583 (1978).
Theory-Part I (Wiley, New York, 1968). 39. The Poisson nature of the detected photons follows directly from
17. G. S. Brindley, "The relation of frequency of detection to intensity the small value of the degeneracy parameter of the light, 6 =
of a stimulus for a system of many independent detectors each (m)TcIT, where T, is the coherence time of the light [see L.
of which is stimulated by a m-quantum coincidence," J. Physiol. Mandel, Proc. Phys. Soc. 74, 233-242 (1959); B. E. A. Saleh,
169, 412-415 (1963). Photoelectron Statistics (Springer-Verlag, Berlin, 1978)]. For
18. J. Nachmias and E. C. Kocher, "Detection and discrimination of the source used by HSP, '(m) 10, T 10-3 sec, and r c- 10-13
luminance increments," J. Opt. Soc. Am. 60, 382-389 (1970). sec, whereas for our multimode Ar+ laser source -r c 10-9 sec
19. T. E. Cohn, L. N. Thibos, and R. N. Kleinstein, "Detection of a [see G. Vannucci and M. C. Teich, Appl. Opt. 19, 548-553 (1980)].
luminance increment," J. Opt. Soc. Am. 64, 1321-1327 (1974); T. In both cases S <<1. We shall see in Part 3 of this series3 2 that
E. Cohn and D. J. Lasley, "Detectability of a luminance incre- certain rather common sources of visible radiation, such as tele-
ment: Effect of spatial uncertainty," J. Opt. Soc. Am. 64, vision and oscilloscope images produced by cathodoluminescence,
1715-1719 (1974). as well as image-intensified light [see A. van Meeteren, Vision Res.
20. J. Nachmias and R. M. Steinman, "Study of absolute visual de- 18, 257-267 (1978)], in general, generate non-Poisson photon
tection by the rating-scale method," J. Opt. Soc. Am. 53, 1206- statistics [see Refs. 43 and 47]. Their use in vision experiments
1213 (1963). therefore requires exceptional care.
21. B. Sakitt, "Counting every quantum," J. Physiol. 223, 131-150 40. J. Neyman, "On a new class of 'contagious' distributions, appli-
(1972). cable in entomology and bacteriology," Ann. Math. Stat. 10,35-57
22. M. A. Bouman, "History and present status of quantum theory (1939).
in vision," in Sensory Communication, W. Rosenblith, ed. (MIT 41. M. C. Teich, "Role of the doubly stochastic Neyman Type-A and
Press, Cambridge, Mass., 1961), Chap. 21, pp. 377-401. Thomas counting distributions in photon detection," Appl. Opt.
23. M. A. Bouman and J. J. Koenderink, "Psychophysical basis of 20, 2457-2467 (1981).
coincidence mechanisms in the human visual system," in Reviews 42. H. B. Barlow, "The absolute efficiency of perceptual decisions,"
of Physiology: Biochemistry and Experimental Pharmacology Phil. Trans. R. Soc. Lond. Ser. B 290,71-82 (1980); A. F. Burgess,
(Springer-Verlag, Berlin, 1972), Vol. 65, pp. 125-172. R. F. Wagner, R. J. Jennings, and H. B. Barlow "Efficiency of
24. E. Baumgardt, "Threshold quantal problems," in Handbook of human visual signal discrimination," Science 214, 93-94
Sensory Physiology: Visual Psychophysics, D. Jameson and L. (1981).
M. Hurvich, eds. (Springer-Verlag, Berlin, 1972), Vol. VII/4, 43. M. C. Teich and B. E. A. Saleh, "Fluctuation properties of mul-
Chap. 2, pp. 29-55. tiplied-Poisson light: Measurement of the photon-counting
25. J. Nachmias, "Signal detection theory and its application to distribution for radioluminescence radiation from glass," Phys.
problems in vision," in Handbook of Sensory Physiology: Visual Rev. A 24, 1651-1654 (1981).
Psychophysics, D. Jameson and L. M. Hurvich, eds. (Springer- 44. H. Grimm, "Tafeln der Neyman-Verteilung Typ-A," Biom. Z.
Verlag, Berlin, 1972), Vol. VII/4, Chap. 3, pp. 56-77. 6, 10-23 (1964).
26. G. Buchsbaum, "The retina as a two-dimensional detector array 45. M. C. Teich, P. R. Prucnal, G. Vannucci, M. E. Breton, and W.
in the context of color vision theories and signal detection theory," J. McGill, "Non-Poisson nature of the effective noise in the visual
Proc. IEEE 69, 772-786 (1981). system near threshold," J. Opt. Soc. Am. 68, 1454(A) (1978).
27. H. B. Barlow, "Retinal noise and absolute threshold," J. Opt. Soc. 46. Several of our stipulations on the format of detection are, no
Am. 46, 634-639 (1956). doubt, unnecessarily restrictive. For example, the assumptions
28. W. J. McGill, "Neural counting mechanisms and energy detection of a single counting focus and a fixed integration time could likely
in audition," J. Math. Psychol. 4, 351-376 (1967). be relaxed without seriously affecting the outcome.
29. W. J. McGill, "Poisson counting and detection in sensory sys- 47. B. E. A. Saleh and M. C. Teich, "Multiplied-Poisson noise in
Teich et al. Vol. 72, No. 4/April 1982/J. Opt. Soc. Am. 431

pulse, particle, and photon detection," Proc. IEEE, 70 (to be 67. D. A. Baylor, G. Matthews, and K.-W. Yau, "Two components
published, 1982). of electrical dark noise in toad retinal rod outer segments," J.
48. B. E. A. Saleh and M. C. Teich, "Statistical properties of a non- Physiol. 309, 591-621 (1980).
stationary Neyman-Scott cluster process," IEEE Trans. Inf. 68. H. B. Barlow, T. E. Cohn (School of Optometry; University of
Theory (submitted for publication). California, Berkeley; Calif. 94720), W. R. Levick, and L. N. Thi-
49. K. Matsuo, B. E. A. Saleh, and M. C. Teich, "Cascaded Poisson bos, "Low light level performance of cat retinal ganglion cells"
processes," J. Stat. Phys. (submitted for publication). (personal communication).
50. B. Sakitt, "Canonical ratings," Percept. Psychophys. 6,478-488 69. W. R. Levick and J. L. Zacks, "Responses of cat retinal ganglion
(1974). cells to brief flashes of light," J. Physiol. 206, 677-700 (1970).
51. T. E. Cohn, "A new hypothesis to explain why the increment 70. S. Yeandle and J. B. Spiegler, "Light evoked and spontaneous
threshold exceeds the decrement threshold," Vision Res. 14, discrete waves in the ventral nerve photoreceptor of Limulus,"
1277-1279 (1974). J. Gen. Physiol. 61, 552-571 (1973).
52. A. Rose, "The relative sensitivities of television pick-up tubes, 71. We are grateful to H. B. Barlow and W. R. Levick for providing
photographic film, and the human eye," Proc. IRE 30, 293-300 us with experimental data.
(1942). 72. W. R. Levick, "Form and function of cat retinal ganglion cells,"
53. HI. de Vries, "The quantum character of light and its bearing Nature 254, 659-662 (1975).
upon the threshold of vision, the differential sensitivity and acuity 73. P. 0. Bishop, W. R. Levick, and W. 0. Williams, "Statistical
of the eye," Physica 10, 553-564 (1943). analysis of the dark discharge of lateral geniculate neurones," J.
54. P. R. Prucnal and M. C. Teich, "An increment threshold law for Physiol. 170, 598-612 (1964).
stimuli of arbitrary statistics," J. Math. Psychol. 21, 168-177 74. H. Nakahama, N. Ishii, M. Yamamoto, and H. Saito, "Stochastic
(1980). properties of spontaneous impulse activity in central single
55. H. B. Barlow, W. R. Levick, and M. Yooh, "Responses to single neurons," Tohoku J. Exp. Med. 104, 373-409 (1971).
quanta of light in retinal ganglion cells of the cat," Vision Res. 1, 75. If we assume that the counting time is very large in comparison
Suppl. 3, 87-101 (1971). with the decay times T, for all k stages, and that the multiplica-
56. H. B. Barlow and W. R. Levick, "Threshold setting by the sur- tion parameter of each stage is given by aj, the variance-to-mean
round of cat retinal ganglion cells," J. Physiol. 259, 737-757 ratio is 4 9
(1976).
57. P. Lennie, "Scotopic increment thresholds in retinal ganglion
cells," Vision Res. 19, 425-430 (1979).
R = ((An) uiejjai++P.
k- I (kfi a
2)/(n)= i+zfi

58. Y. Y. Zeevi and S. S. Mangoubi, "Noise suppression in photore-

ceptors and its relevance to incremental intensity thresholds," This is the origin of the quantity p used in Part 2 of this series of
J. Opt. Soc. Am. 68, 1772-1776 (1978). papers.31 If the multiplication parameters of all stages are
59. M. C. Teich, L. Matin, and B. I. Cantor, "Refractoriness in the identical, and equal to a, then
maintained discharge of the cat's retinal ganglion cell," J. Opt. R = 1 + a[(i - ak-1)/(1 - a)]
Soc. Am. 68, 386-402 (1978). and
60. M. C. Teich and P. Diament, "Relative refractoriness in visual
information processing," Biol. Cybern. 38, 187-191 (1980).
p = a[(l - ak1)/(1 - a)].
61. M. C. Teich and B. E. A. Saleh, "Interevent-time statistics for For k = 1 and h = 2, we recover the usual expressions for the
shot-noise-driven self-exciting point processes in photon detec- Poisson and Neyman Type-A distributions, respectively. Ran-
tion," J. Opt. Soc. Am. 71, 771-776 (1981). dom deletion from an SNDP also maintains the cluster property
62. B. E. A. Saleh, J. Tavolacci, and M. C. Teich, "Discrimination of of this process, although with a reduction in the multiplication
shot-noise-driven Poisson processes by external dead time: parameter a. 47
application to radioluminescencefrom glass,"IEEE J. Quantum 76. E. Cinlar, "Superposition of point processes," in Stochastic Point
Electron. QE-17, 2341-2350 (1981). Processes: Statistical Analysis, Theory and Applications, P.
63. H. A. van der Velden, "Quanteuse verschijnselen bij het zien," A. W. Lewis, ed. (Wiley-Interscience, New York, 1972), pp.
Ned. Tijdschr. Natuurkd. 15, 147-151 (1949). 549-606.
64. K.-W. Yau, T. D. Lamb, and D. A. Baylor, "Light-induced fluc- 77. D. R. Cox and W. L. Smith, "The superposition of several strictly
tuations in membrane current of single toad rod outer segments," periodic sequences of events," Biometrika 40, 1-11 (1953).
Nature 269,78-80 (1977). 78. D. R. Cox and W. L. Smith, "On the superposition of renewal
65. D. A. Baylor, T. D. Lamb, and K.-W. Yau, "Responses of retinal processes," Biometrika 41, 91-99 (1954).
rods to single photons," J. Physiol. 288, 613-634 (1979). 79. G. 0sterberg, "Topography of the layer of rods and cones in the
66. K.-W. Yau, G. Matthews, and D. A. Baylor, "Thermal activation human retina," Acta Ophthalmol. Suppl. 6,1-103 (1935).
of the visual transduction mechanism in retinal rods," Nature 279, 80. H. R. Wilson and J. R. Bergen, "A four mechanism model for
785-786 (1979). threshold spatial vision,," Vision Res. 19, 19-32 (1979).