BACKGROUND/SIGNIFICANCE

1. Specific Aims

Visual information typically arrives as a complex continuous stream. This stream must be parsed, and from it relevant features must be extracted. These features are used for such tasks as localizing an object in space, identifying a letter, directing gaze, and recognizing a face. Although these are just a few examples, these tasks are typical in that they rely on the initial spatial and temporal sensory information provided by the visual system. I propose to study how initial spatial and temporal information are jointly used in a variety of tasks involving locating, fixating and identifying an object in space. The findings will help constrain models of these tasks, and help illuminate the neurophysiological architecture of the visual system. A second goal is identifying how information processing mechanisms act on the spatial and temporal information to produce visual awareness.

The theoretical basis for these explorations is a model that I have been developing over the past four years with Dr. Geoffrey Loftus, called the Linear Systems Theory (LST) model (Busey, 1994; Busey & Loftus, 1994; Loftus, Busey & Senders, 1993). This model, described in detail below, represents a synthesis of three major domains: a) low-level temporal vision (Sperling and Sondhi, 1968; Watson, 1986), b) information processing theory (Townsend, 1981; Rumelhart, 1970), and c) visible persistence and iconic decay (Sperling, 1963; Coltheart, 1980). The LST model provides a good starting point because it has been able to account quantitatively for data in a wide variety of domains, it has a theoretical structure that is both interpretable and falsifiable, and it codifies basic research findings from low-level vision research into a theory that can account for the temporal and spatial processing of complex naturalistic objects and everyday tasks.

It may seem odd to place as strong a research emphasis on temporal processing as spatial processing, given that objects are defined by their spatial distribution on the retina. However, even visual tasks as simple as letter and object recognition tasks are not carried out as instantaneous operations. Sensory information accumulates over time, only then providing inputs to later information extraction mechanisms. These dynamic properties of vision are critical in real life situations in which visual information is almost always moving across the retina or remaining stable for just a few hundred milliseconds between saccades. An understanding of the interactions of temporal and spatial mechanisms is important not only for modeling visual tasks, but also for developing an understand of how disruptions in these mechanisms contribute to visual problems such as developmental dyslexia.

These projects are organized into four Specific Aims:

1. In processing alphanumeric characters, do localization and identification rely on the same temporal frequencies, and how does this change when high-contrast letters are used?

In my current work I have found differences in the temporal frequency information used during letter identification and localization tasks, suggesting that the early visual pathways carrying high and low temporal frequency information may remain segregated through the primary visual cortex to later visual processing centers. The proposed research extends these findings to lighting and contrast conditions common to natural reading. This is of considerable interest since reading involves both localizing letters prior to a saccade, and identification of characters thereafter. Pilot data suggests that the temporal frequency information used during letter identification changes with increasing contrast, demonstrating the need to extend low-contrast findings to high-contrast stimuli.

2. What is the role of pathways carrying high temporal frequencies in saccadic suppression, dyslexia and directing gaze?

The second project examines the role of eye movements in letter identification and localization tasks. These experiments address the interactions of parallel visual pathways during visual saccades, and provide information about the functional neuroanatomy of the visual system as well as visual disorders such as dyslexia. Questions about the role of high temporal frequencies in directing gaze are also addressed.

3. Do temporal integration and segregation tasks rely on the same temporal frequency information?

The third project examines the processes underlying integration and segregation tasks such as combining two parts of a stimulus over time, or detecting the presence of a temporal gap in a visual stimulus. While integration and segregation may be viewed as complimentary tasks, the visual system may selectively attend to the temporal frequencies that provide the best information for each task.

4. How do different spatial and temporal frequency channels interact when perceiving naturalistic objects?

The fourth project looks at the time-course of processing of various spatial frequency bands, including an examination of the proposition put for by Hughes, Nozawa & Kitterle (submitted) and Navon (1977) that low spatial frequencies inhibit the processing of higher spatial frequencies.

The theoretical aim of this project is to characterize the temporal and spatial information used when processing complex stimuli in complex tasks. Model development will start with the LST model and its extensions. Although the LST model is used to guide experimentation, it must be stressed that the data produced are also used to address general questions suggested by existing literature, and are not limited to model development.

2. Background/Significance

The present research assumes that predictions for complex stimuli and situations require both a precise model for low-level visual mechanisms and a correct mapping of these through higher level processes. By conjoining the linear filter models of the temporal vision literature with the information-processing models of the perception and cognition literature, the LST model can be used as a tool to characterize the initial representations of a visual stimulus, and describe how information-extraction components extract information from this initial stimulus representation. An exploration of the interactions between the two stages is then possible.

This holistic model-based approach has some clear benefits: 1. A number of leading models of information processing and letter identification (e.g. Townsend, 1981; Fisher, 1982) do not consider the initial stimulus representation (e.g. on the retina or in the striate cortex) from which features are extracted to enable recognition. 2. Some prominent models of spatial vision ignore the buildup of information over time, instead assuming that this representation simply exists, and features are extracted from it (e.g. Wilson & Regan, 1984). 3. By conjoining the initial stimulus representation with later information processing components, the LST model provides the theoretical structure to address current debates in the spatial vision literature, such as whether global information contained in low spatial frequencies is processed before higher spatial frequencies (e.g. Navon, 1977). 4. A model of the spatial and temporal information used in a given task can be used to test theories of the functional neuroanatomy of the visual system, addressing mechanisms such as saccadic suppression, the cortical targets of early visual pathways, and the temporal disruptions that result in specific reading disorder.

3. The LST model and Preliminary Studies

To motivate the LST model I begin with a relatively simple result, called the 'Gap' effect, that cannot be accounted for by standard models of iconic memory but can be handled by LST. A number of more complex predictions of the LST model are then described, including the model's ability to account for temporal integration experiments and as a result produce estimates of the temporal frequency information underlying a given task.

3.1 Accounting for Temporal Integration: The "Gap" effect

In a temporal integration experiment, the observer views a stimulus, such as a row of four digits, presented in one of two basic conditions, schematized in Figure 1. In both conditions the two stimuli are shown at low contrast (e.g. 2.5%) for the same overall duration (80 ms). In the 'Gap' condition a 250 ms blank screen is inserted in the middle of the display. The 'No-Gap' condition is identical, with the exception that the gap size is zero; accordingly there is just one 80-ms presentation. Note that in the 'Gap' condition, the same stimulus appears twice, each time in the same location. The actual experiments vary the duration of the gap and add a condition with an opposite-contrast pulse. In this condition a light-gray stimulus is followed by a dark-gray stimulus, both consisting of the same letter shown on the same gray background. As a result of the two identical stimulus presentations on each trial, this experiment is called a "Two-Pulse" experiment.

Figure 1. Two conditions of the 'Gap' experiment, along with the presumed icons that are engendered by each presentation. The Iconic Decay model makes the prediction that Gap performance will exceed No Gap performance, because the two icons produce a stronger internal representation in memory. This prediction is incorrect, as illustrated in Figure 2.

The standard view of iconic decay and visible persistence (e.g. Coltheart, 1980) holds that the visual representation of a stimulus persists beyond the offset of the physical stimulus. This continuation of the stimulus is termed 'visible persistence' or 'iconic decay,' and each presentation is assumed to engender an 'icon' that decays exponentially (e.g. Di Lollo, 1984). Accordingly, this view makes the prediction that performance in the 'Gap' condition would be superior to performance in the 'No-Gap' condition, because each presentation engenders an icon and the 'Gap' condition has two icons to the 'No-Gap' condition's one icon.

Figure 2. Data from a Two-Pulse experiment, in which the gap between two pulses of a stimulus is varied. The same-contrast conditions consist of two presentations of the same letter in the same location at the same contrast. The opposite-contrast conditions consist of a brief positive-contrast pulse followed by a variable gap and a negative-contrast pulse. Contrary to the predictions of the Iconic Decay model, performance is best in the No Gap condition and gets progressively worse as the gap size increases. For this subject, the opposite-contrast data rise above the same-contrast data at 30-45 ms ISI. These two data patterns are predicted by the LST model, as discussed below.

This prediction, however, is incorrect: performance in the 'Gap' condition is substantially worse than performance in the 'No-Gap' condition. Figure 2 shows the data from an experiment from Busey (1994) in which two 30 ms presentations of the same digits in the same locations were separated by a variable gap. For same-contrast pulses, performance is best in the zero-ms gap condition, and decreases with increasing gap size. For opposite-contrast pulses, performance is worst at the zero-ms gap condition and then initially rises above same-contrast performance before declining to the same performance level as the same-contrast stimulus at long gap sizes.

3.2 Description of the LST model

Figure 3 shows the components of the LST model that can be used to a) account for the temporal summation that occurs in the Two-Pulse data in Figure 2, and b) determine the temporal characteristics of the sensory representation used in a given visual task (a more complete version of the theory can be found in Busey and Loftus (1994), which is included as an appendix to this proposal). The onset and offset of the letters are represented as changes in contrast over time, defined as a function f(t), which is then temporally low-pass filtered by the model into a sensory representation, a(t), that is a temporally blurred representation of the physical stimulus. Information is then extracted from the sensory response at a rate r(t).

3.2.1 From the Physical Stimulus to the Sensory Response

The physical stimulus, f(t) is represented as changes in contrast over time (Figure 3, top panel). The initial sensory representation engendered by this physical stimulus is generated by convolving the physical stimulus representation with an impulse response function, g(t).

Eq. 1

The result a(t), called the sensory response function, is shown the middle panel of Figure 3. The filtering operation defined by Eq. 1 is linear, so that the sensory representation contains the same overall energy as the physical stimulus, but spread out in time. The exact form of g(t) is described below, but the effect is that a(t) is a temporal blurring or filtering of the physical stimulus f(t). The shape of the impulse response function, g(t), dictates the type of temporal blurring. In keeping with previous linear filter models (e.g. Watson, 1979, 1986) I have chosen this function to be the difference of a two gamma functions, each with a different time constant,

g(t) = - s[] Eq. 2

where t represents the time-constant of the gamma function and provides an estimate of the temporal response properties of the mechanisms mediating a given task. The sensory response function component of the LST theory (a(t)) is based on work from Andrew Watson (Watson, 1986), George Sperling (Sperling and Sondhi, 1968) and others working in the temporal domain of perception.

The first term in Eq. 2 represents an excitatory component, and this term alone is used to produce the solid function shown in the middle panel of Figure 3. The second term of Eq. 2 represents an inhibitory component of the response, which tends to sharpen the response and allows it to respond to higher temporal frequencies; both terms together produce the dashed curve in the middle panel of Figure 3. The parameter r represents the ratio of the time-constant of the inhibitory component of the response to the excitatory component of the response, and s represents the magnitude of the temporal inhibition component. The parameter n represents the number of stages in each process, and is usually fixed at an integer between 5-9, although the shape of the impulse-response function is relatively unchanged by the precise value chosen.

Detection data for stimuli such as high spatial frequencies and color are often modeled by setting s to 0, which gives a monotonic impulse response function g(t) as shown as a solid curve in the left panel of Figure 4 (Note that the a(t) function in Figure 3 looks like the impulse response function, g(t), in Figure 4, because the physical input leading to the Figure 3 curve is a brief rectangular pulse, similar to an impulse). An alternative way of representing the same information is by taking the Fourier transform of the impulse response function g(t), which results in a temporal contrast sensitivity function (TCSF). The TCSF plot show the sensitivity of a system to different temporal frequencies. The TCSF corresponding to the solid line in the left panel of Figure 4 is given by the solid line in the right panel. These curves represent a purely sustained response, and give a monotonically-decreasing TCSF, as shown in Figure 4, right panel.

Detection data for stimuli containing mainly low spatial frequencies, or stimuli presented on bright backgrounds, often are modeled by s> 0. In this case the impulse response inhibits processing after an initial excitatory response, which results in an inhibitory lobe in the impulse response function g(t) (dashed line in Figure 4, left panel) and a characteristic TCSF with a decrease in sensitivity at low temporal frequencies (dashed curve in Figure 4, right panel).

Figure 4. The temporal frequencies underlying a task may be characterized by an impulse response function (left panel) that characterizes the time course of the perceptual response engendered by a stimulus, or by the temporal contrast sensitivity function (right panel), that characterizes the fidelity by which the pathways subserving a given task pass different temporal frequencies. The TCSF plots are the Fourier transform of the impulse response functions into frequency space. High spatial frequency stimuli tend to elicit monophasic impulse response functions, which have no falloff at low temporal frequencies in the TCSF plot. Stimuli containing low spatial frequencies tend to elicit biphasic impulse response functions that contain an inhibitory lobe. This temporal inhibition acts to sharpen the response of the visual pathway, allowing it to respond to faster changes in the visual scene. However, this inhibition causes in a falloff at low temporal frequencies in the TCSF plot, which results from tendency for the biphasic impulse response function to inhibit itself when processing slow temporal changes. Parameters used: Monophasic: {t = 4.38, r = 0} Biphasic: {t = 3.58, r = 2.0, s = 0.39}

3.2.2 Information Extraction from the Sensory Response Function

It has long been recognized that information is carried both by the positive going part of a(t) (the upper lobe of the dashed line in the middle panel of Figure 3) and by the negative going part of a(t) (the below-zero part of the dashed line in the middle panel of Figure 3). Note that the negative going portion can result either from temporal inhibition or by a negative contrast pulse as in the two-pulse study described above. Representing information in both negative and positive a(t) functions is usually handled by taking the absolute value of a(t), |a(t)|, in a process called rectification (e.g. Watson, 1986).

The critical assumption of the LST model is that information is extracted not from |a(t)| but from that part of |a(t)|, termed , that lies outside a sensory threshold, Q. To be precise,

Eq. 3

A fundamental consequence of this formulation is that if the sensory response a(t) is not outside the positive or negative sensory threshold, the stimulus will not be visible to the observer. Although there is evidence against such a high-threshold formulation, the psychometric function relating contrast to performance is quite steep, and thus the sensory threshold serves as an approximation to the true mechanism.

Information is extracted from according to assumptions based on the information processing literature (e.g. Rumelhart, 1970; Townsend, 1981). Information of a given sort is assumed to be extracted from the sensory response function at a given rate, r(t), which is proportional to both the height of the above-threshold sensory response and the amount of already-acquired information:

Eq. 4

where represents the degree to which a(t) lies above the sensory threshold and I(t) represents the amount of already-acquired information at time t. The rate of information extraction is given by a model parameter, c_s. The parameter c_s determines the acquisition rate for the different types of information required to satisfy the demands of different tasks. Information is acquired until time t is termed I(t), and is simply the integral of r(t) over time. The total amount of acquired information is termed I(_). The form of r(t) for two examples of a(t) is illustrated in the bottom panel of Figure 3. I(_) is the area under each r(t) function.

3.2.3 An application of LST to the "gap" result

To apply this model to the digit recall task as well as related tasks such as detection, it is assumed that the proportion correctly-recalled-digits, p, is equal to the amount of extracted information: p = I(_). Thus the total amount of acquired information, expressed as a proportion of total task-relevant information, is assumed to be equal to the proportion of correctly recalled digits. This provides a quantitative prediction for performance in the digit-recall and related detection and identification tasks. Different tasks such as a segregation task may require a different decision mechanism. For example, a temporal gap may be detected if the height of the sensory response curve drops a specified amount.

Figure 5. The "Gap" effect accounted for by the LST model. The two conditions generate the same overall area under the sensory response functions. However, the 'No-Gap' condition generates much more above-sensory threshold area, thus making the prediction that 'No-Gap' performance will be superior to the 'Gap' performance. This prediction is opposite the prediction made by the iconic decay model, and conforms to the actual data which shows a substantial advantage to the No-Gap condition. The gray lines in the middle panel correspond to the individual contributions of the two pulses.

One implication of this formulation of the LST model is that the area of is monotonically related to performance. The implications for the "Gap" experiment are illustrated by Figure 5. A consequence of the linear nature of the convolution process is that all three a(t) curves have the same overall area. However, once the sensory threshold has been applied we see that the 'Gap' condition generates much less above-threshold area in than the 'No-Gap' condition, thus accurately predicting the actual results that 'Gap' performance is much worse than 'No-Gap' performance. Figure 5 also demonstrates a strength of the LST model, in that it can account for the temporal summation caused by overlaps in the individual responses from each of two pulses. Iconic decay theory does not have a mechanism for summing the icons to produce a combined response. Thus Figure 5 demonstrates that as the gap size increases, there is less overlap of the individual responses, which provides less above-threshold area and thus less identification accuracy.

In addition to predicting the decreasing performance for increasing gap size with same-contrast stimuli, the LST model also accounts for the finding that, at intermediate gap sizes (e.g. 45-60 ms), performance in the opposite-contrast condition is better than performance in the same-contrast condition. I account for this finding by assuming that the response to the first pulse contains an inhibitory lobe which overlaps with the excitatory lobe from the second pulse, as demonstrated by Figure 6. This results in the inhibitory lobe subtracting from the excitatory lobe, reducing the above-threshold area for the same-contrast condition and increasing the above-threshold area for the opposite contrast condition. At longer or shorter gap sizes, the two lobes would not overlap and therefore the gap size that produces the best opposite-contrast performance determines the width and shape of the underlying sensory response functions.

Figure 6. Model predictions for same-contrast condition (left panel) and an opposite-contrast condition (right panel), demonstrating how opposite-contrast performance can rise above same-contrast performance in the Figure 2 data. In the same-contrast condition, the inhibitory lobe from the first pulse inhibits the facilitory lobe from the second pulse. The phase of these peaks helps constrain the width of the pulses and thus provides an estimate of the temporal frequencies used in the task. The same holds for the opposite-contrast pulse, with the exception that the inhibitory lobe from the first pulse facilitates the negative-going facilitory lobe from the negative-contrast pulse. This provides more above-threshold area and thus accurately accounts for the finding in Figure 2 that opposite-contrast performance is superior to same-contrast performance at 30-45 ms ISI. These demonstrate how the model not only accounts for temporal integration, but can be used to determine the temporal parameters that underlie the processing of a given task. Note that here I plot an unrectified for purposes of exposition.
Model Parameter	Description	Interpretation
t s r	time-constant amount of temporal inhibition inhibition time-constant ratio	Together these define the temporal frequencies underlying a task by determining the shape of the impulse-response function.
q	sensory threshold	Determines the amount of information lost to early perceptual mechanisms such as noise
cs	information-extraction rate	Defines the rate at which information is acquired from a stimulus.
Table 1. LST model parameters and their interpretations. Different tasks selectively manipulate each parameter.

3.2.4 Assessment of temporal frequencies underlying a given task

When the LST model is applied to two-pulse data as described in the preceding paragraph and in Figure 6, the parameter values provide information about the temporal frequencies underlying the given task through bounds on the model parameters t, s and r, as shown in Table 1. This approach will underlie a good part of the present research.

3.3 Different Tasks Rely on Different Temporal Frequencies

A major theme of this proposal is to apply the LST model to different tasks that rely on the same stimulus, and estimate the temporal parameters associated with each task. When these differ significantly for different tasks, we have evidence that different pathways or visual mechanisms may act on the same stimulus. As a first step towards applying low-level vision research techniques to more complex tasks such as reading letters and words, I have examined the temporal frequencies underlying letter identification and localization tasks. The two tasks appear to rely on different sources of information, and I describe both the methodology and the extensions to the LST model that are required to account for these differences.

The left panel of Figure 7 shows an example of Two Pulse data collected in my laboratory at Indiana University, in which participants performed both character identification and localization tasks on each trial. Contrast was adjusted according to a threshold-finding procedure to find a contrast threshold that provided 82% correct identification or localization performance. The pattern of data changes with the task, and modeling via LST reveals that the Localization task relies on higher temporal frequencies then the identification task, as shown in the Right panel of Figure 7.

Figure 7. Left Panel: Data from Two-Pulse localization and identification experiment. On each trial observers indicated where a character was presented (either right or left of fixation, 7° in the periphery) and whether the character was a '2' or a '5'. The Localization data show an overall pattern that is different than the Character Identification data. The curves represent the fit of the LST model, assuming that the two tasks rely on different temporal frequencies. Right Panel: Parameter estimation reveals that the observers are relying on higher temporal frequencies in the Localization task than in the Identification task, as shown as an impulse response function with a sharper peak and a short t time constant of 2.62 ms.
Figure 8. Extension of the LST model to account for the Figure 7 and Figure 11 data. The Localization task relies on higher temporal frequencies, which gives an impulse-response function with a smaller t (Figure 7, right panel) and a smaller integration interval. We also must assume that the information extraction rates cLoc and cID differ to account for the different task demands.

Figure 8 shows the modeling steps that were used to deduce that Localization and Character Identification tasks in Figure 7 rely on different temporal frequencies. Different tasks required different t, r and s parameters, which results in different impulse response functions, different sensory response functions and thus different above-threshold areas. The two tasks also require different amounts of information to complete, and thus we also must include different rates of information acquisition, c_Loc and c_ID. This demonstrates that a complete account of data across tasks must include consideration of the information-processing components of the tasks.

3.4 Parameter Estimation

Many of the conclusions that are possible with the LST model require estimation of parameters, including t, s and r that determine the temporal frequencies underlying a given task. In many cases, different stimuli or tasks will produce parameter values that are obviously different. However, in some cases it will be necessary to determine whether parameter values are significantly different. These situations are addressed via hierarchical model testing and statistical hypothesis testing.

Figure 9. If two parameters trade off, confidence regions can be constructed by systematically varying the parameter values until a statistically-worse model fit results. The dot represents the parameter values giving the best model fit, and the ellipse represent the region of parameters that provide fits that are still close enough to be statistically non-significantly different from the best fit.

Hierarchical model testing involves fitting a general version of LST with all parameters free, and comparing the results to restricted versions with one or more parameters specified. Likelihood ratio or ² tests allow one to determine whether the extra parameters significantly improve the fit. For example, one could fit localization and identification data with different parameters, or with common temporal parameters see if the fit changes significantly. There are other cases where one may be less interested in comparing models than in obtaining reliable quantitative estimates of parameter values. In such cases one may create confidence intervals around parameter values by systematically varying the parameter values until the model fit becomes significantly worse by a metric such as the log likelihood fit. Often interpretational problems result from parameters that trade off, such as the t and c_s parameters in the LST model. In such cases, a multi-dimensional confidence region can be created around a group of parameters, as shown in Figure 9 for two parameters.

One criticism often leveled at linear-filter type models such as LST is that they are parameterized so richly as to be unfalsifiable. This should not be a problem in the present applications, since the number of free parameters should not exceed 6 or 7 (depending on the to-be-modeled task), and the amount and type of data should allow tight constraints in these parameters. In addition, evidence from psychophysics, anatomy and physiology suggests that the formulation of the LST model represents the minimum amount of complexity required to model the visual behavior. Note also that the impulse-response function parameters can be constrained through pilot work. Eqs. 1 and 2 allow various stimulus wave forms to be tested empirically, through use of ramping and flickering stimuli. Finally, simple versions of the model have been disconfirmed, thus pointing the way towards more complex formulations (e.g. Busey & Loftus, under revision). Thus the LST model achieves an appropriate balance between explication and complexity.

3.5 Previous tests of the LST model

One of the predictions of the LST model is a very simple and strong one: any two conditions yielding the same above-threshold area in the sensory response function a(t) must yield the same performance. This prediction has been upheld in many experiments and paradigms, summarized in Table 2. Of course, the LST model also predicted many other characteristics of the data from these studies.

Domain	What is manipulated	What is measured
Digit Recall	Stimulus temporal distribution (e.g. gap vs. no-gap)	The amount of information extracted as measured by proportion of correctly-recalled digits. Busey & Loftus, 1994.
Binocular Information Combination	Viewing eye (e.g. monoptic vs. dichoptic vs. binocular)	The amount of information extracted as measured by proportion of correctly-recalled digits. Busey & Loftus, under revision.
Visual Masking	Delay between onset of a backward mask	The amount of masking at different delays. Loftus, Busey & Senders, 1993.
Visible Persistence	Delay of a marker 'click'	The amount of phenomenological persistence engendered by the stimulus following stimulus offset. Wolford, 1992.
Temporal Integration	Frame-1 duration, ISI, Frame-2 duration in a Di Lollo Missing Dot integration task	The amount of persistence and temporal integration as indexed by temporal integration performance. Loftus & Irwin, submitted.
Picture memory	Stimulus temporal distribution (e.g. gap vs. no-gap)	The information extracted as measured by later old/new recognition test. Loftus & McLean, in preparation.
Character Detection and Identification	Stimulus temporal distribution (e.g. gap vs. no-gap) with the addition of negative-contrast second pulses	The temporal frequencies underlying character identification and detection tasks as measured by a two-pulse task. Busey, 1994.
Table 2. Applications of the LST model to different experimental domains.

3.6 Summary of the LST model

The ability to model both the creation of an initial sensory representation as well as subsequent information-processing components positions the LST model at the interface between sensation, perception and cognition. When the sensory information is relevant, the model can be used as a tool to determine the nature of the information used in a given task. However, more complex tasks such as reading or object recognition require a more complete representation of the information processing components, and the LST model provides a foundation upon which such components can be added.

4. Research Design and Methods

The major theme of this work is accounting for performance in complex perceptual tasks using, in part, a model based on fundamental visual properties. In many cases the LST model can be used as a tool to characterize the temporal and spatial information used in a task. However, many of the experiments proposed below enable conclusions that are independent of the particular model I have adopted, and directly address current issues in the vision and perception literatures. Thus validating and extending the LST model is only one goal of the proposed research, although the model is used to organize the research program in a systematic way.

Figure 10 provides an overview of the organization of the research program. Two pathways leading through the lateral geniculate nucleus (LGN) of the thalamus may differ in the temporal frequencies carried by each pathway. Evidence for this dichotomy may be found in different tasks that rely on different temporal frequencies. The proposed experiments follow a similar theme. A set of stimuli are created through a temporal manipulation, such as varying the size of a temporal gap or the frequency of a flickering stimulus. Different tasks are performed on each of stimulus, to evaluate whether different tasks rely on different sources of temporal or spatial information derived from the same stimulus. In some cases the stimuli are altered to try to select one visual pathway over another. Each proposed experiment begins with a summary of previous findings and unanswered questions, followed by a description of the research design.

4.1 The Temporal and Spatial Frequencies Underlying Letter Identification and Localization Tasks

Tasks involving visual displays of letters from which information must be extracted (like reading), include two major processes: typically the letters must be localized in space and fixated (to maximize the rate and quality of extracted information), and then the information from the letters must be extracted. These two processes may rely, in part, on different sources of spatial and temporal information. For example, localizing a letter or word in space might use one set of spatial or temporal visual channels, while identifying that letter might require a different set of spatial or temporal channels. If this turns out to be the case (and evidence collected in my laboratory and briefly described below suggests it is the case), we could then begin to understand how deficits in different visual mechanisms might affect either of these two processes. An important first step is to develop a methodology to identify the temporal and spatial frequencies underlying localization and identification tasks.

A major unanswered question relating to the localization and identification tasks is whether separate parallel visual pathways mediate flicker and pattern perception. A seminal work by Kulikowski and Tolhurst (1973) demonstrated that the temporal frequencies used in a task varied according to whether the observer was instructed to detect "pattern" or "flicker." This suggests that two distinct pathways exist: one pathway mediates flicker detection, while the other mediates spatial structure. Functionally, Tolhurst (1973) and Kulikowski and Tolhurst (1973) separated the mechanisms supporting form identification into a sustained mechanism and flicker detecting into a transient mechanism. Tolhurst (1973) linked form-analyzer and flicker detection pathways with the X- and Y-cell pathways of the cat visual system. The attributes of the potentially homologous Parvo (P) and Magno (M) parallel pathways in primates appear to support the form/flicker distinction: P cells support lower temporal frequencies and higher spatial frequencies, while M cells carry higher temporal frequencies and lower spatial frequencies (Merigan & Maunsell, 1990; Merigan, Katz & Maunsell, 1991). However, Lennie (1980) found no major differences in the spatial sensitivities of X- and Y- cells, and a host of other researchers have demonstrated that the spatial and temporal sensitivities of the M and P pathways overlap to some extent. Lennie (1988) concluded that while the role of the P pathway is clear in making fine spatial discriminations (Lynch, Silveira, Perry and Merigan, 1992), the role of the M pathway is less clear. He concludes that one possible role for the M cells is localizing objects in space and directing gaze.

In addition to the early M and P parallel pathways, research from neuroscience suggests that different aspects of a stimulus may be processed along different cortico-cortico pathways leading out of the primary visual cortex. A set of pathways in cortical areas described by Ungerleider and Mishkin (1982) and confirmed in humans via PET scanning by McIntosh, Grady, Ungerleider, Haxby, Rapoport, & Horwitz (1994) are thought to carry different forms of visual information. In monkeys, a dorsal pathway extends from V1 into the inferior parietal area. Human patients with lesions of the parietal lobe pathway show neglect for regions of the visual field, and thus this parietal lobe pathway has been linked to the processing of the locations of objects in visual space. A second cortical pathway projects ventrally from V1 into the temporal lobe. Patients with lesions in this temporal lobe pathway show impaired form perception, and thus this pathway has been linked to object identification. The location and object identification tasks used by McIntosh et al. (1994) selectively activated the parietal and temporal pathways in the PET data.

Livingstone and Hubel (1988) tentatively proposed linkages directly from the M-pathway to the parietal pathway and from the P-pathway to the temporal pathway. This is perhaps a reasonable proposal: in other sensory system such at the tactile pathways, targets to somatosensory cortical areas remain segregated in the thalamus (Kaas & Pons, 1988). However, this suggestion has proven somewhat controversial. For example, the temporal pathway receives approximately equal number of inputs from both the M and P pathways, and the fate of the M and P pathways is quite complex within the primary visual cortex. Baizer, Ungerleider and Desimone (1991) provide evidence for M- and P-pathway "roots" to the parietal and temporal pathways, but suggest that it is unlikely that either the parietal or temporal cortex would become completely deafferented following lesions of the M- and P-pathways. Merigan and Maunsell (1993) concur that the relationship between subcortical and cortical pathways is neither one-to-one nor simple.

If the early pathways maintain segregated input to the cortical pathways, or at least selectively influence the two pathways, we might expect that the two tasks would exhibit differences in the temporal frequencies that are used in localization and identification tasks, which reflects the relative contributions of the M- and P-pathways. This prediction was confirmed in the simultaneous localization/identification experiment described in Figure 7 in Section 3.3 above.

Figure 11. Data from Temporal Contrast Sensitivity Function (TCSF) experiments with letter identification and localization tasks. On each trial participants viewed a '2' or a '5' appearing left or right of fixation. This stimulus was flickered at different rates, and the participant adjusted the contrast of the stimulus until the letter appeared just detectable or just identifiable. The final threshold contrast is inverted and plotted against the temporal flicker frequency. The localization data show evidence of temporal inhibition, suggesting a reliance on higher temporal frequencies. The identification data show a sustained response, with no sensitivity falloff at the low temporal frequencies. This is consistent with a reliance on lower temporal frequencies. The LST model provides estimates of the temporal frequencies used in each task, as defined by the impulse response functions associated with localization and identification tasks (inset).

In addition to the Two-Pulse method described in Figure 7 for measuring the temporal frequencies underlying a given task, a more direct method involves temporally modulating the stimulus according to a sine-wave at different temporal frequencies. The observer adjusts the contrast of the stimulus until it is just barely localizable or just barely identifiable, and the inverse of this contrast threshold is plotted as a function of the temporal frequency at which the stimulus was modulated. Figure 11 provides an example of these data, collected in my laboratory using letters appearing at different temporal frequencies in locations either to the right or left of the fixation point.

Figure 7 and Figure 11 provide evidence that letter identification relies on different temporal frequencies than letter localization. These differences may be accounted for based on task considerations: letter identification might require higher spatial frequencies, which are linked to lower temporal frequencies (Robson, 1966). This cannot be the entire story, however. Data collected in my lab demonstrate that when the letters are severely low-pass filtered to restrict the available range of spatial frequencies, the letter identification task still shows evidence of relying on slower temporal frequencies than the localization task (Busey, in preparation). In addition, Kulikowski and Tolhurst (1973) used low- and high spatial frequency gratings in their dissociation of the temporal frequencies associated with pattern and flicker detection, which restricted the range of available spatial frequencies.

The data from Figure 7 and Figure 11 are consistent with the view that the pre-cortical M- and P-pathways retain some segregation through the primary visual cortex and provide differential stimulation to the cortical parietal and temporal pathways extending out of area V1. Thus the 'where' pathway is served best by high temporal frequencies that are easily detected, while the 'what' pathway makes the best use of the lower temporal frequencies that provide a perceptibly stable stimulus suitable for feature identification. While these data are clear for threshold-contrast stimuli, whether such a distinction exists for supra-threshold contrasts is still an open question.

The goal of the current project is to measure a set of curves similar to those shown in Figure 11 for the localization and identification of high contrast letters. This methodology would enable us to address questions such as the role of the M-pathway in localizing and fixating peripheral targets, as well as the causal factors of a high-temporal frequency deficit in reading disorders. However, we cannot simply increase the luminance of the letters, because participants would identify the letters 100% of the time, wiping out differences between conditions. Instead, I propose an extension of previous flicker studies that allow high-contrast stimuli using a reaction-time based approach. This design will characterize the temporal frequencies underlying a given task, and thus can be used to derive the relative contributions of the M- and P-pathways to the localization and identification of high-contrast letters. While this project does not initially use the LST model, the data will be modeled by the LST theory in Experiment 5.

4.1.1 Experiment 1:A Reaction Time Methodology for High-Contrast Letters

Figure 12. Threshold estimates from hypothetical reaction time data that provide relative sensitivities to different stimuli (e.g. different temporal frequencies). A psychometric function relating contrast to reaction time is fit with a power function for each of many temporal frequencies (only two are shown here). An arbitrary criterion of +100 ms is defined, and the contrast associated with a 100 ms increase in reaction time from asymptotic performance gives the contrast threshold for that condition. This contrast is then inverted to give relative sensitivity for that condition on a TCSF plot, similar to Figure 11. The threshold contrast is 9.5% for the 4 Hz condition (dark arrow) and 26% for the 16 Hz condition (gray arrow). The specific choice of the +100 ms criterion is standard within RT-Threshold applications (e.g. Palmer, 1996). Specific conclusions do not depend on the choice of criterion, since I will analyze the resulting TCSF plot to determine the temporal frequencies underlying a given task. Absolute reaction time is irrelevant for this analysis.

Reaction times have a long tradition in Psychology and Vision research, but typically require a great many assumptions about the rate of information accumulation and the initiation of a motor response. I propose a method that will alleviate many of these concerns and provide a direct estimate of the contributions of different spatial and temporal frequencies to letter perception.

The data of Figure 7 and Figure 11 provide converging information: both allow estimates of the temporal frequencies underlying the processing of a given stimulus. However, the TCSF data in Figure 11 require fewer assumptions about the information-extraction rate, because temporal frequency is manipulated and so estimates of the sensitivity of the processing pathways to each frequency is directly measured and need not be inferred through parameter estimation of the impulse response function parameters. A similar methodology for obtaining thresholds in a reaction-time paradigm is sketched out in Figure 12.

The basic methodology is similar to the threshold definition within accuracy, but instead of measuring a psychometric function relating contrast to accuracy, I measure a psychometric function relating contrast to reaction time. Rather than fit this psychometric function with a weibull curve, I use a shifted power function as shown in Figure 12. As with accuracy, I adopt an arbitrary, roughly midrange, performance criterion, but instead of choosing a performance criterion below perfect performance (e.g. 75% correct identification), I choose a reaction time increment of a fixed criterion (e.g. 100 ms) and choose the contrast that provides an increase of 100 ms from the asymptotic reaction time defined by the power function fit. This process is repeated for all conditions in which a threshold is desired (e.g. all temporal frequencies used in a TCSF experiment). The estimated threshold contrast for each temporal frequency is inverted to give sensitivity, and plotted in a curve exactly like Figure 11. This RT-based TCSF is assumed to characterize the fidelity by which the pathway subsuming some task processes that particular temporal frequency, and extends the TCSF methodology to above-threshold stimuli.

Pilot Data

The RT-Threshold technique has been successfully used in a number of domains, from simple detection experiments (Mansfield, 1973) to links to flicker-fusion data (Roufs, 1974) and more recently visual search paradigms (Palmer, 1996). To validate this methodology with the flickering stimuli such as those used to measure the TCSF of Figure 11, I conducted a pilot study to both verify the applicability of the RT threshold technique to flickering stimuli as well as provide converging evidence that Localization and Identification tasks rely on different temporal frequencies. These data, shown in Figure 13, demonstrate that for Localization, the relation between flicker rate and reaction time does not depend on stimulus contrast. However, for Identification, flicker rate and reaction time interact across stimulus contrast. At dim contrasts (where integrating the stimulus response over time might be important for sensitivity) the participant was faster in the 2 Hz condition than in the 8 Hz condition. This pattern reverses for larger contrasts, suggesting that the range of temporal frequencies used in the Identification shifts with increasing contrast. The possibility of such a shift makes using high-contrast stimuli very important if we are to generalize low-contrast data to high-contrast tasks such as reading.

Figure 13. Pilot data with power function fits demonstrating the RT-Threshold technique for flickering stimuli. Although only two flicker frequencies are used (2 Hz and 8 Hz), the data demonstrate that the relation between react time and flicker rate changes with increasing contrast: For Localization, mean reaction time does not interact with flicker frequency across contrasts. However, for Identification, the 2 Hz data have a shallower slope than the 8 Hz data. The data confirm the low-contrast accuracy TCSF's such as that shown in Figure 11: power functions with different asymptotes fit to the identification data curves reveal that the 8 Hz data will have a higher contrast threshold than the 2 Hz data. This means that Identification performance is less sensitive overall to the 8 Hz condition than the 2 Hz condition, which replicates typical Identification TCSF's that show a monotonic decrease in sensitivity from 2 Hz to 8 Hz. The Localization thresholds are identical to each other, also replicating typical TCSF's that show Localization sensitivity remaining constant or increasing from 2 Hz to 8 Hz. Note that the overall mean reaction times are irrelevant to these analyses, since a particular flicker rate might deliver information at a faster rate once the trial begins.

The pilot data also confirm the accuracy data shown in Figure 11. Applying the RT threshold technique to the Localization data in Figure 13 with a criterion of 200 ms gives a threshold contrast for both frequencies of about 0.13, although any criterion will give virtually identical thresholds for the two localization frequencies. The longer reaction times for the Identification data requires a longer criterion of 400 ms, which gives a contrast threshold of 0.5 for the 8 Hz data and 0.21 for the 2 Hz data. While the absolute threshold contrasts are not directly comparable to the Figure 11 TCSF data, the overall pattern is, and when such a comparison is made I find the same pattern of data as in Figure 11: Identification sensitivity decreases with increasing temporal frequency, while Localization sensitivity increases or remains constant up to 8 Hz. Since sensitivity is the inverse of contrast threshold, the same pattern holds in the Figure 13 data, thus tentatively confirming for high contrasts the finding that Localization and Identification tasks rely on different temporal frequencies. This validates the use of the RT Threshold technique for use with high-contrast, flickering stimuli.

Even without the RT-Threshold technique, the Figure 13 data demonstrate that the relationship between flicker frequency and reaction time changes with increasing contrast. This suggests that one should not assume that the findings from near-threshold stimuli will generalize to high-contrast stimuli without explicitly testing such an assumption. Experiment 1 is designed to be such a test.

Experimental Design

The specific experimental paradigm will involve a letter appearing left or right of fixation, flickering at one of 8 temporal frequencies for 500 ms. During localization blocks, participants will make speeded responses indicating the location of the letter. During identification blocks, participants will indicate the identity of the letter (either a '2' or a '5') using stimulus-compatible response mappings. The contrast of the letter will be systematically varied in order to produce an RT psychometric function and estimates of RT thresholds such as those in Figure 12. These threshold contrasts are measured for a range of temporal frequencies for both localization ('where was the letter') and identification ('what was the letter') tasks. These contrast thresholds are then plotted in a graph similar to Figure 11, from which estimates of the temporal frequencies underlying letter localization and identification tasks can easily be derived. If we find evidence that the two tasks rely on different temporal frequencies we will then have support for the segregation of the M- and P- pathways through the primary visual cortex to the parietal and temporal lobe pathways. This experiment will be repeated using low-pass, band-pass and high-pass spatially filtered letters to examine the degree to which the differences depend upon the tasks relying on different spatial frequencies.

Evidence that localization is driven by the M-pathway and identification by the P-pathway will come from two related experiments using the RT-threshold technique to derive TCSF's. In these experiments, I will change the stimulus rather than the task. Participants will make speeded left/right localization judgments to two different types of stimuli. One stimulus consists of a low-frequency luminance grating and is designed to specifically stimulate the M-pathway. The other stimulus consists of a high-frequency isoluminant grating specifically designed to stimulate the P-pathway. These will be flickered at different rates according to the RT-Threshold TCSF methodology described above. The resulting reaction time data will be used to produce two TCSF's that can be compared to the localization and identification TCSF to determine if indeed the localization and identification tasks are selectively influenced by the M- and P-pathways.

Establishing this methodology for high-contrast stimuli extends previous research to more naturalistic stimuli, and provides a foundation for future development of the LST model. Once the temporal frequency information is determined, the sensory front-end of the model can be fixed and development of the information-extraction back-end of the model can proceed. Many of the difficulties suffered by reading-disabled persons may be attributable to the information acquisition stages, rather than the early information registration stages (Rayner & Pollatsek, 1989), and thus the interface between sensory representation and information acquisition is likely to be a fruitful research topic for understanding reading disorders, as addressed in the next topic.

4.2 The Temporal Dynamics of Eyemovements

The second step in testing the proposition that the M-pathway localizes objects in space and directs gaze (Lennie, 1988) involves identifying the temporal frequencies underlying eyemovements. A further extension of this research examines the role of the M-pathway in saccadic suppression. This research has direct bearing on reading disorders such as Developmental Dyslexia, and is also relevant to our understanding of normal reading patterns.

4.2.1 Experiment 2: Does the M pathway trigger saccades?

The proposition that the M-pathway directs saccades is tested using stimuli specifically designed to activate the M- and P-pathways. As in Experiment 1, reaction times will be the dependent measure, although in Experiment 2 a reaction time will be determined by the onset of a saccade, rather than the key presses used in Experiment 1.

Experiment 2 uses different stimuli to segregate the M- and P-pathways. The Magno stimulus consists of low-spatial frequency gratings flickering at 12 Hz, while the Parvo stimulus consists of a high-spatial frequency isoluminant grating flickering at 2 Hz. These stimuli correspond to the spatio-temporal characteristics of the M- and P-pathways derived from macaque lesion data (Merigan & Maunsell, 1990; Merigan, Katz & Maunsell, 1991). The isoluminant stimuli used in this and subsequent experiments are red-green sine-wave gratings in which the luminance of the red and green colors have been matched. The M-pathway has a much-reduced sensitivity to such isoluminant stimuli (Livingstone & Hubel, 1988), thus effectively isolating the P-pathway.

On each trial, two oblique-oriented gratings will appear on the left and right sides of a fixation point. The participant must saccade to one orientation while ignoring the other (e.g. saccade to the left-oblique stimulus, ignoring the right-oblique stimulus). Figure 14 schematizes the design of this experiment. Some trials will consist of a mixture of magno and parvo stimuli, while others will contain just magno or just parvo stimuli. If the M-pathway drives saccades, then we would expect fast reaction times when saccading to a Magno stimulus in the presence of a Parvo stimulus, but slower reaction times when saccading to a Parvo stimulus when a Magno stimulus is to be ignored. The contrast of the Magno stimulus will be reduced, to equalize the overall reaction times to the Magno and Parvo stimuli.

Figure 14. Experiment 2 stimuli. On each trial the participant makes a saccade to the left-oblique grating. If the saccade mechanisms are driven by the M-pathway, then we would expect interference on trials in which the participant must ignore the Magno stimulus and saccade to the Parvo stimulus. This interference results from having to inhibit an otherwise fast response resulting from a direct connection between the M-pathway and the saccade mechanisms.

On each trial, the participant will see two stimuli, one on each side of the fixation point. They must make a saccade to the stimulus oriented to the left, and the onset of the saccade, as measured with eye movement recording equipment, will determine the response time (as opposed to the overt keypress from Experiment 1). If the M-pathway is somehow more directly connected to the saccade mechanisms, then this fast response will interfere with on trials in which the task is to saccade to the parvo stimulus. Such interference is determined by examining the reaction time on trials in which saccades are made to the parvo stimulus when the magno stimulus is the to-be-ignored stimulus. These reaction times are compared to the reaction times on trials in which saccades are made to the parvo stimuli when a second parvo stimulus is the to-be-ignored stimulus.

Converging evidence for the proposition that the M-pathway drives both localization and saccades comes from an extension of the RT threshold technique of Experiment 1. In this extension, the same stimuli and design from Experiment 1 will be used, with the exception that participants will make a saccade to the location of a letter, and the onset of the saccade will determine the reaction time. The resulting TCSF's can be compared to the localization TCSF, and if the M-pathway drives both localization and saccades we expect a great deal of overlap between the temporal frequencies used in each task. Further converging evidence will come from comparing this TCSF to appropriately-scaled TCSFs from parvo-lesioned macaques from Merigan and Maunsell (1993). The parvo-lesioned macaques show a clear band-pass pattern with a peak sensitivity of around 10 Hz, while magno-lesioned macaques show a low-pass function with a peak frequency of around 2 Hz. If the human data appear low-pass, then we have evidence against Lennie's proposition that the magno pathway mediates saccades.

4.2.2 Experiment 3: Applications to Transient-On-Sustained Inhibition and Developmental Dyslexia

The differences between identification and localization tasks seen in the two-pulse and TCSF data are particularly relevant to specific reading disorder. Several researchers have identified a deficiency in high temporal frequencies as a potential factor in dyslexia (Lovegrove, Garzia & Nicholson, 1990; Evans, Drasdo & Richards, 1994; Williams & LeCluyse, 1990). Less clear, however, is what the role these high temporal frequencies presumably supplied by the M-pathway play in reading or eyemovements. Several authors have suggested that while a transient deficit may exist, it is not causal for dyslexia (Evans et al. 1994) or that dyslexia results from faulty cognitive processing of the text, rather than low-level sensory deficits (Morris & Rayner, 1996).

Several theories have been proposed for the behavior of the M-pathway during saccades. Recent evidence from Burr, Morrone & Ross (1994) and Burr & Morrone (submitted) suggests that the magno pathway is suppressed during saccades. However, Burr & Morrone report an increase in the use of high temporal frequencies during the saccade, rather than a decrease that would be expected if the magno pathway was suppressed during the saccade. Contrary to this suppression hypothesis, Breitmeyer and Gantz (1976) have proposed that the M-pathway becomes activated during the saccade by low spatial frequency and high temporal frequency input caused by the rapid eye motion. This pathway then inhibits the processing of the P-pathway, reducing persistence from the preceding fixation and 'clearing the slate' in time for the next fixation. A deficit in the M-pathway would result in a combination of images from the current and previous eye fixations in reading disabled persons. However, Burr & Morrone found no evidence of suppression in isoluminant color stimuli designed to activate the parvo pathway, casting doubt on the parvo suppression hypothesis.

This evidence is somewhat contradictory, and the vast majority of research has been done using near-threshold stimuli. It is quite reasonable to ask whether findings from low-contrast vision have any application at all to natural tasks such as reading with high contrast letters, and what role the M-pathway plays when processing high-contrast stimuli. Experiment 3A extends the saccadic suppression findings of Burr et. al. to high-contrast letters in order to explore the implications of saccadic suppression for letter stimuli. Experiments 3B and 3C addresses the transient-on-sustained inhibition hypothesis using supra-threshold and near-threshold stimuli, to examine the possible role of inter-pathway suppression for dyslexia. These data will help reveal the role of high temporal frequencies in normal saccades to letter stimuli. Once an accurate model has been verified for normal viewers, the implications of specific visual deficits for dyslexia can be explored in future studies.

Experiment 3A methods

Experiment 3A adopts the RT-Threshold methodology developed in Section 4.1.1 to derive the TCSF of the visual pathways responsible for processing letters immediately after a saccade. The observer will make a saccade to a peripheral location to start a trial, and computer-controlled eyemovement recording equipment will present a letter flickering at a given contrast and temporal frequency in the location of saccade termination. The letters appear upon saccade termination, and will continue to flicker for 500 ms, although much of the information driving the response will come from the initial portion of the stimulus presentation. The contrast of the supra-threshold stimuli will be varied systematically to create an RT psychometric function such as that shown in Figure 12. The time-limited nature of the experiment ensures that the effects of saccadic suppression, thought to last as long as 170 ms (Burr & Morrone, submitted), will still influence the processing of the flickering letter and provide relative estimates of the amount of suppression at different temporal frequencies.

The data will consist of choice reaction times to letters presented in one of 8 temporal frequencies shown at one of 5 contrast levels. The eye movement recording equipment will initiate a trial within 4 ms of saccade termination, and the participant will respond according to the identity of the letter.

A variant of the task will use sine wave gratings rather than letters in order to isolate the M- and P-pathways. Participants will make speeded orientation judgments to either low-spatial-frequency luminance gratings or high-spatial-frequency isoluminant gratings. The major difference between this experiment and previous experiments is that here the onset of the stimulus occurs after saccade termination, while in previous experiments the onset of the stimulus caused the saccade.

The analysis of the data follows that of previous experiments, in that reaction time TCSFs will be derived for the saccade conditions and compared to TCSFs from foviated stimuli. If the magno-pathway suppression findings from Burr and Morrone extend to high-contrast, higher-spatial frequency stimuli such as letters, then we would expect an increasing reliance on higher frequencies during the saccade, as revealed by a shift in the saccade TCSF. We would expect sensitivity to increase at higher temporal frequencies in the saccade TCSF, compared to a foviated TCSF.

The pathways subsreving letter identification are much more sustained than that subserving location (Busey, in preparation), and thus M-pathway suppression either may not affect processing of letter stimuli or may cause a deceleration of the saccade TCSF. Comparisons between the different TCSFs from the Magno and Parvo stimuli will reveal the extent of suppression in the two pathways.

Experiment 3B Methods

While the Experiment 3A addressed the implications of M-pathway suppression during saccades, it did not address the transient-on-sustained inhibition proposed by Breitmeyer & Ganz (1976). Such suppression can be manipulated by extending the previously-described experiments to include a horizontal or vertical low-spatial frequency grating presented during the saccade. The vertically-oriented grating will provide much more activation of the M-pathway, and if the resulting saccade TCSF with the vertical grating becomes more transient relative to the TCSF with the horizontal grating then I will have evidence to support the transient-on-sustained inhibition hypothesis.

Experiment 3C Methods

Figure 15. Experiment 3C trial events. A saccade is initiated by the subject, and eye movement recording equipment causes a horizontal or vertical low-spatial frequency grating to appear during the saccade. This grating provides stronger activation of the M-pathway if it is vertically oriented. A high-spatial frequency red-green isoluminant grating is briefly (100 ms) displayed at near-threshold contrasts at the end of the saccade, which is designed to selectively activate the P-pathway. Evidence for transient-on-sustained inhibition comes from the finding of poorer sensitivity to the parvo stimulus when the magno stimulus is vertically oriented compared with sensitivity to the parvo stimulus when the magno stimulus is horizontally oriented.

Further converging evidence for or against transient-on-sustained inhibition will come from a planned collaboration with Dr. Arthur Bradley at the Indiana School of Optometry. While previous experiments have used supra-threshold stimuli and reaction time measures, the current experiment uses near-threshold stimuli in a two-alternative forced choice detection task. Using the eye movement recorder, we will briefly present a high-spatial frequency, low temporal frequency isoluminant stimulus (parvo stimulus) at the termination of a saccade, and vary the presence or absence of a low-spatial frequency grating (magno stimulus) presented during the saccade. The low-spatial frequency luminance grating is designed to selectively stimulate the M-pathway, while the high-spatial frequency low-temporal frequency isoluminant stimulus is designed to selectively isolate the P-pathway. Participants will indicate whether the parvo stimulus was present or not at the end of the saccade. Figure 15 shows the sequence of events on a given trial.

A horizontal grating will only weakly stimulate the M-pathway during the saccade, while a vertical low-spatial-frequency grating will provide much more activation. By changing the orientation of the low-spatial frequency grating during the saccade from horizontal to vertical, we can manipulate the response of the M-pathway and thus measure the effects of transient-on-sustained inhibition on the sensitivity of to the saccaded target. If we find that sensitivity to the parvo stimulus is reduced when the magno stimulus provides more activation (vertical orientation) then we would have evidence for transient-on-sustained inhibition. Further extensions include a parvo presented at fixation prior to the saccade, to test Brietmeyer's 'clearing-the-slate' hypothesis directly. If an irrelevant pre-saccadic parvo stimulus inhibits processing of the post-saccadic parvo stimulus only when the M-pathway is not activated, then we would have evidence for Brietmeyer's hypothesis that an activated M-pathway inhibits the P-pathway.

4.3 Experiment 4: Do integration and segregation tasks rely on the same temporal frequencies?

In perceiving the everyday world, an observer must make sense of a constant stream of visual information, interrupted by eye blinks, eye movements and changing stimuli. Making sense of this information requires both providing perceptual continuity over time, as well as detecting changes in the environment. These are two competing design requirements: perceptual continuity requires integrating temporally contiguous events into a single unified percept, while detecting change requires segregating temporally contiguous stimuli into separate objects. Researchers have identified no simple rule that the visual system might use to determine when to integrate and when to segregate this visual stream. One reason for this failure to understand the mechanisms behind integration and segregation is the lack of a common reference among experiments. Integration tasks that make up the bulk of the previously-described experiments are rarely combined with segregation experiments, and thus even the most straightforward questions remain unanswered. Chief among these is whether integration and segregation rely on the same temporal frequencies. Clearly segregation benefits from higher temporal frequencies, while integration might rely more on lower temporal frequencies to provide perceptual continuity. However, the question whether the visual system actually adopts such a dichotomy is unclear. To answer these questions, I propose a factorial combination of integration and segregation tasks with carefully chosen stimuli and paradigms, to address two issues: a) do integration and segregation tasks rely on the same initial sensory representations (and thus the same temporal frequencies) and b) what decision criteria are applied in each task to translate some aspect of the internal representation into a response?

Figure 16. Missing Dot Task. Twenty-five dots in a 5x5 matrix are split into two displays of 12 dots each. The two displays are shown sequentially with a short intervening gap, making this a temporal integration task. The participants task is to report the missing dot, which is easy when the duration of the second display is short. Longer durations produce worse performance, which has been challenging to model via linear filter models and demonstrates the inverse-duration effect for this task.

Most tasks looking at integration and segregation use two brief displays, separated by a variable gap. An important issue in terms of temporal integration and segregation is whether the stimuli in two displays overlap. The letters in the Gap and Two-Pulse experiments always appeared in the same location. However, a related task that addresses temporal integration directly is the Missing Dot task (Dixon & Di Lollo, 1984), in which dots in a 5 by 5 array are split into two displays, each containing 12 dots that do not overlap spatially. When the two displays are presented for brief exposures, separated by a short blank screen, the 24 filled locations define a single missing dot location which the participant must report. This task is trivial if the gap between the two displays is short, but becomes quite difficult with gaps of longer than 70-100 ms. I will also vary the duration of the second display, since this results in an intriguing finding of an inverse duration effect. Figure 16 demonstrates that as the duration of the second display is increased, integration performance decreases. The inverse-duration effect provides a strong test of linear filter models, including the LST model. These models currently require improbable model extensions to account for this effect, and thus one goal of the current proposal is to derive more physiologically-sound mechanisms that account for this finding.

To summarize, Integration and Segregation depend on both spatial and temporal contiguity, and thus spatial contiguity may have important implications for whether two pulses are perceived as an integrated whole. As a result, I propose to systematically investigate integration and segregation using both overlapping and non-overlapping stimuli.

4.3.1 Experimental Design

The proposed experiments are all extensions of previous experiments conducted in my lab. The design involves a systematic manipulation of presentation variables such as stimulus duration and interstimulus interval, which will directly address the question of whether the integration and segregation tasks rely on the same initial sensory representation. Once this question has been addressed, modeling via LST will determine how decision mechanisms are applied to the initial sensory representation(s). The manipulated stimulus variables are chosen on the bases of their theoretical interest to visual scientists, not merely as tests of the LST model.

Table 3 illustrates the design of this experiment. Although 4 different tasks will be used, common to all tasks will be the perceptual variables that will be manipulated. All tasks involve presenting two displays separated by a short gap. In all experiments the size of the gap will vary from 0 ms to 100 ms.

	Task
Stimuli	Integration	Segregation
Overlapping Stimuli (Letters)	Letter Identification	Gap Detection
Non-Overlapping Stimuli (5x5 Dot Matrix)	Missing Dot Task	Gap Detection
Table 3. Experimental Paradigm used to determine whether Integration and Segregation tasks rely on the same initial sensory representation (and therefore the same temporal frequencies). The stimuli will vary from task to task (Overlapping Stimuli tasks will use letters, while Non-Overlapping Stimuli tasks will use dots from a 5x5 matrix). Different tasks also require different decisions to be made by the participant: A letter identification

Experimental Design

The spatially-contiguous task will consist of a single letter shown in the fovea for two brief, 30 ms, pulses, separated by a gap that varies from 0 to 100 ms. The participant will respond either with the identity of the letter (integration task) or whether a gap was present (segregation task). Counterbalancing measures call for 50% of the trials to be the zero-gap condition to make the segregation task a balanced forced-choice design.

The spatially-separate task will consist of two stimuli, each composed of 12 dots from a 5x5 grid. The integration task will require a participant to identify the location of a missing dot. The segregation task will require a gap detection response as in the spatially-contiguous task. Each pulse will last 30ms, and the gap will again vary from 0 ms to 100 ms with the same counterbalancing as above. In each case, performance will be determined by the proportion of correct responses.

Analysis of these experiments has two stages. First, as described in Section 4.3.2 below, the question of whether Integration and Segregation tasks rely on the same initial sensory information can be addressed using Bamber State-Trace Analysis (Bamber, 1979). The second stage involves an analysis of the data from different tasks with the LST model, in order to determine the precise decision rule used in each task.

4.3.2 State-Trace Analysis

The basic question of whether Integration and Segregation tasks rely on the same initial sensory information can be addressed using Bamber's State Trace Analysis (Bamber, 1979). The top panel of Figure 17 represents the hypothesis that different tasks rely on the same underlying sensory representation. The bottom panel of Figure 17 represents the hypothesis that the two tasks rely on different underlying sensory representations. Note that these two hypothesis are exclusive.

Two assumptions underlie the Single Source model: first, the two independent variables jointly influence the formation of an unidimensional internal representation, in this case the initial sensory representation. Second, this representation is the sole determinant of performance in the integration and segregation tasks. In other words, performance in the integration and segregation tasks are monotonic functions of SR1: m_I(SR1) and m_S(SR1).

The prediction of the Single Source model is straightforward: the rank order correlation over experimental conditions between Integration and Segregation performance is 1.0. The reasoning is based on the monotonicity assumption. Define P_I(i) and P_I(j) as two conditions from the Integration task and P_S(i) and P_S(j) as two conditions from the segregation task. A finding that P_I(i) > P_I(j) implies that SR1(i) > SR1(j), where SR1(i) and SR1(j) are the internal representations issuing from conditions i and j. This in turn implies that P_S(i) > P_S(j). This relationship defines the over-condition monotonic relationship between P_I and P_S.

A rejection of the Single-Source model implies that different tasks rely on different aspects of the initial sensory representation and/or different temporal frequencies. For example, while integration might benefit from a more visible stimulus, the segregation task might benefit from a representation that had lots of change. In terms of temporal frequencies, this implies that the integration task relies on lower temporal frequencies, while the segregation task relies on higher temporal frequencies.

4.3.3 Modeling with LST

Figure 18. Response curves generated by two pulses of a letter presented in the same location. Participants perform both integration (Letter Identification) and segregation (Gap Detection) tasks. Panel A: Under a Single-Source model, Integration and Segregation rely on the same decision criterion (e.g. Gap Height). Large gaps might decrease the perceptual continuity of the stimulus and decrease integration performance while increasing segregation performance. Panel B: Under one version of a Two-Source model, Integration and Segregation rely on different criteria (e.g. Gap Height and Above-Threshold Area). Panel C: Under another version of a Two-Source model, Integration and Segregation rely on different temporal frequencies as well as different criterion. An intermediate Two-Source model could be constructed in which Integration and Segregation rely on different temporal frequencies but the same criterion (e.g. Gap Height).

Independent of the conclusions derived from State-Trace Analysis, the LST model can be used to determine, in a quantitative fashion, how different aspects of the underlying sensory response function are used to determine performance in different tasks. Figure 18 shows three possibilities. If a single-source model is confirmed, the LST model can be used to determine what aspect of the sensory response function is used by both integration and segregation. For example, Panel A of Figure 18 shows how under a single-source model, integration and segregation are both mediated by the height of the gap, which decreases perceptual continuity therefore decreases integration performance while increasing segregation performance. Alternatively, Panel B of Figure 18 demonstrates how different response characteristic can be applied to different tasks that rely on the same underlying sensory response function. Panel C of Figure 18 demonstrates how the two tasks rely on different temporal frequencies in addition to different response characteristics. The situations sketched in Panels B and C of Figure 18 are both examples of a dual-source model, although only the second model involves the assumption that the two tasks rely on different temporal frequencies.

The question of whether the two tasks rely on different decision criterion and different temporal frequencies will be addressed with both spatially-contiguous stimuli such as repeated letters as well as spatially non-contiguous stimuli such as a missing dot task. For example, previous modeling of letter identification tasks has relied on the amount of above-threshold area to predict performance (Busey & Loftus, 1994; Busey, 1994), while modeling of Missing Dot tasks has relied on the correlation between the two sensory representations engendered by the two displays. However, linear filter models have rarely been applied to segregation tasks, and then only qualitatively (Dixon and DiLollo, 1986; Wolford, 1992). Thus one goal of the current project is to extend the LST model to include a set of mechanisms that is more physiologically plausible and that can account for integration, segregation and inverse-duration effects.

4.4 Experiment 5: Extensions of the LST model to Reaction Times

The reaction time threshold analyses required for the projects proposed in Sections 4.1 and 4.2 did not require a full model of the information extraction processes underlying response time. As a result, conclusions could be drawn that were not dependent upon the structure of the information processing components of the LST model. However, such an approach has limitations, in that it is often difficult to account for more complex tasks without attempting to account for all processing stages. Thus I propose a project aimed to extend the theoretical structure of the LST model to account for the processing mechanisms that underlie speeded letter and word identification.

Figure 19. Reaction time model formulation following that of Ratcliff (1988) and Heath (1984). A random walk simulates the information-processing stages, and the termination point determines the reaction time. The drift rate of the random walk is determined by the height of the information extraction rate function, r(t), at each point following stimulus onset. A bright stimulus will give a taller r(t) function and thus a larger drift rate and a faster mean reaction time.

This project benefits from the large numbers of reaction times collected for the projects proposed in Sections 4.1 and 4.2, and thus this project mainly involves theoretical extensions to the information processing components of the LST model. Rather than starting the model building from scratch, I propose adopting the theoretical structure of Ratcliff's diffusion model (Ratcliff, 1988) as a back-end to the linear-filter front-end of the model. Discussions with Ratcliff have suggested that such an approach could prove very fruitful in terms of integrating the traditional reaction-time-based information processing literature with the low-level vision literature. The exact formulation of the model will have to wait for simulations to be conducted, but as a starting point I will adopt the formulation sketched in Figure 19.

The processes underlying reaction time can be thought of as a process of accruing evidence for a response in the presence of noise. The rate of information acquisition is related to the information-acquisition rate r(t) from the LST model, and a response is made once the process crosses a decision boundary. The reaction time is the time necessary to cross the boundary, and errors result from crossing the wrong boundary due to limited information and the presence of noise. Other candidate models include a version in with the integral of r(t), I(t), determines the drift rate of the process towards the boundary.

The data analysis and model fitting will follow that established in previous work (Busey & Loftus 1994, Busey, 1994). A major advantage of the linear nature of the sensory front end is that one can rigorously test the sensory front-end of this model using a vast variety of stimulus presentations, such as repeated stimuli, ramping on and ramping off, flickering stimuli, and the standard contrast and duration manipulations. Once verified, this should give a fairly good picture of how the information-extraction mechanisms work to acquire information from the initial sensory representation.

4.5 Spatio-temporal interactions in the processing of naturalistic images

The LST model can be used to address the spatio-temporal interactions between different spatial frequency channels that occur during the processing of naturalistic images. These images contain a wide range of spatial frequencies, and Navon (1977) first proposed the notion that naturalistic scene perception proceeds first with an analysis of low spatial frequencies and then an analysis of higher spatial frequencies. Recent corroborating evidence from object recognition and neuroscience literatures suggests that large spatial frequencies are given processing precedence over fine spatial frequencies (Lamb & Robertson, 1989) or have faster transmission times (Schiller & Malpeli, 1978). This low spatial frequency information contains global information about an object which is then used to guide extraction of more local information, presumably contained in the higher spatial frequencies.

Such temporal interactions between different spatial frequency bands is the topic of the current project. Qualitative modeling of the data is possible using a variant of General Recognition Theory (Ashby & Townsend 1986). Once interactions are identified, quantitative modeling via LST is made possible through an extension described in Section 4.5.2.

4.5.1 Experiment 6: Temporal interactions between spatial frequency bands

To identify the interactions between spatial frequency channels, I will use spatial frequency filtered stimuli such as those shown in Figure 20. Wilson & Regan (1984) have proposed that up to 6 spatial frequency channels exist in the visual system, and these channels can be selectively stimulated by filtering a given image such as a face into widely-separated spatial frequency bands.

Figure 20. Low, medium and high spatial frequency bands of a single face. In an example stimulus of the proposed experiment, the onsets of the medium and high bands are varied relative to the onset of the low band in a speeded identification task. Independence analysis can be used to examine if one spatial frequency band interacts with another, and at what delays these interactions occur.

Experimental Design

In this experiment, stimuli such as those shown in Figure 20 are embedded in a brief movie consisting of visual noise of various spatial bandwidths. The participant must identify which stimuli appeared in the noise movie. In order to examine interactions between these stimuli, two or even three spatial frequency bands may be combined and presented simultaneously. The participant must make three independent decisions on each trial concerning the presence or absence of each of the spatial frequency bands. Some trials will consist of a blank stimulus, which allows me to determine the sensitivity to each individual spatial frequency band relative to a null stimulus.

The data from this experiment are analyzed according to an extension of signal detection theory into multiple dimensions, as embodied by Ashby and Townsend's GRT model (Ashby and Townsend, 1986). This model separates interactions into perceptual and decisional interactions, which can be used to determine whether interactions between spatial frequency bands occur early on at the perceptual level, or later in the decision stage.

Further extensions of this paradigm directly address the putative global-to-local processing strategy. On trials in which more than one spatial frequency band is presented, the offsets of each spatial frequency band will be varied. For example, participant might be quite good at identifying the presence of a high-spatial frequency stimulus if it is preceded by a low-spatial frequency stimulus. However, if the order is reversed and the low spatial frequency stimulus follows the high spatial frequency stimulus, identification performance of the high spatial frequency stimulus may suffer.

The specific design for this extension calls for two lag durations (30 and 15 ms delay between onsets). These are combined with the 6 stimulus orders for conditions with all three frequency bands and 6 stimulus orders for conditions with two frequency bands to give 28 total conditions (12 multiple-band conditions at each lag plus 3 single-band conditions and one blank condition). The number of single-stimulus conditions must be increased to remove potential biases that occur from the increase in the number of multiple-band stimuli in this extension.

Figure 21. Extensions to the LST model that account for the temporal dynamics of interactions between spatial frequency channels. These interactions are identified using the stimuli in Figure 20 presented at variable temporal onsets. Inhibitory and facilitory inter-channel interactions are modeled prior to the decision mechanisms that perform sensory non-linearity and information extraction processes. This inhibition or facilitation between channels may account for observed interactions between stimuli at different temporal onsets. Facilitory and inhibitory connections are shown as filled and open circles, respectively.

4.5.2 Accounting for temporal interactions via the LST model

The mechanisms underlying interactions between spatial frequency bands are address through modeling via the LST model. Recent work by Hughes, Nozawa & Kitterle (submitted) has suggested that low spatial frequency channels receive 'prior entry' to the form and object recognition processing mechanisms, perhaps through a greater contrast gain in the pathways containing lower spatial frequencies or dynamic inhibitory interactions between spatial frequency channels.

Modeling of the Experiment 6 data addresses both of these suggestions. The contrast gain-control proposal is addressed by the relative scaling of the sensory response functions for different spatial frequency bands. Evidence that the low spatial frequency mechanism is more efficient at extracting information than the other two mechanisms would support this notion. Modeling inter-channel inhibition between sensory response functions for different channels examines the proposal that global precedence proceeds via dynamic inhibition of the medium and high spatial frequencies. A finding of inhibition between channels would support this proposition.

The number of free parameters in this model can increase up to 18 parameters depending on model assumptions, making model disconfirmation somewhat more difficult. However, the proposed experimental extension increases the number of conditions to 28 as discussed above, and thus total number of conditions easily exceeds the number of free parameters. In addition, the data from this experiment can also be qualitatively analyzed using standard hierarchical log-linear analyses and Townsend's GRT model (Ashby & Townsend, 1986).

4.6 Timeline

The high-contrast letter identification experiments, the integration/segregation experiments and the spatial frequency interaction experiment(Experiments 1, 4 and 6) represent the most direct extensions of my current research. Thus I will pursue these in the first 2 years of the grant period.

The eyemovement recording equipment will take some time to master, although Dr. Dave Irwin at the University of Illinois has agreed to assist with advice and software routines. Thus I anticipate that I will begin the eyemovement experiments (Experiments 2 and 3) in the second or third year of the grant period.

The modeling required for analysis of the reaction time data (Experiment 5) will move into high gear in the third year.

The fourth and fifth years will be devoted to extending the LST model to account for more high-level cognitive activities such as word reading and object identification. Follow-up studies spun off from the previous research will, no doubt, also require my attention in the fourth and fifth years.

4.7 Gender and Minority Inclusion

Participants are volunteers from the undergraduate and graduate populations at Indiana University. Participants are paid for their participation. Every effort is made to recruit and include ethnic minorities in the studies. No subgroups are excluded from the proposed experiments.

HUMAN SUBJECTS

The proposed research has current human subject certification from Indiana University. The design does not involve invasive procedures or deception. Subjects will be recruited from the Indiana University community, based on their willingness to participate in an extended research project and possessing normal or corrected-to-normal vision. Every attempt will be made to recruit women and minorities as participants. Such a policy is currently in place in the PI's laboratory, with the result that no experiment has been conducted to date without representation from various minority groups. The small number of participants in each experiment prohibits a completely diverse representation, but the design of each experiment calls for an active recruitment of participants from a variety of backgrounds. At no time will a possible participant be excluded based on age, gender, or minority status.

The data collection for all experiments involve keypresses on a computer keypad. The visual stimuli are of low luminance and thus pose little threat to the visual systems of the participants. Eye strain will be ameliorated with frequent rest periods between data collection periods.

These data will be collected for research purposes only, and anonymity will be preserved for all participants. Although the information collected is not sensitive, all names and data will be kept confidential. Only the consent forms will contain any identifying information, and these will be kept in a locked storage room in the PI's laboratory, accessible only to the PI.

The experiments involve no risk to subjects, and adequate procedures will be taken to ensure informed consent and confidentiality. The potential benefits include contributions to the field's knowledge of the functional aspects of the human visual system. The participants will gain exposure to the methods and techniques employed during psychophysical experimentation. The benefits outweigh the minimal risk involved.

LITERATURE CITED

Ashby, F. &. Townsend, J. (1986). Varieties of Perceptual Independence. Psychological Review, 93, 154-179.

Baizer, J. S., Ungerleider, L. G. & Desimone, R. (1991). Organization of visual inputs to the inferior temporal and posterior parietal cortex in macaques. The Journal of Neuroscience, 11(1), 168-190.

Bamber, D. (1979). State-Trace Analysis: A Method of Testing Simple Theories of Causation. Journal of Mathematical Psychology, 19, 137-181.

Breitmeyer, B. &. Ganz, L. (1976). Implications of sustained and transient channels for theories of visual pattern masking, saccadic suppression and information processing. Psychological Review, 83, 1-83.

Burr, D. &. Morrone, M. (Submitted). Temporal impulse response functions for luminance and colour during saccades. Submitted to Vision Research.

Burr, D., Morrone, M. & Ross, J. (1994). Selective suppression of the magnocellular visual pathway during saccadic eye movements. Nature, 371, 511-513.

Busey, T.A. (in preparation). The temporal frequencies underlying localization and letter identification.

Busey, T.A. & Loftus, G.R. (1994). Sensory and cognitive components of visual information acquisition. Psychological Review , 101, 446-469.

Busey, T.A. & Loftus, G.R. (in preparation). Binocular Information acquisition and visual memory.

Busey, T.A. (1994). Temporal inhibition in two-pulse character detection and identification tasks. Ph.D. Dissertation, University of Washington, Seattle, WA.

Coltheart, M. (1980). Iconic memory and visible persistence. Perception & Psychophysics, 27(3), 183-228.

De Yoe, E.A. & Van Essen, D.C. (1988). Concurrent processing streams in monkey visual cortex. Trends in Neuroscience, 11, 219-226.

Dixon, P. &. Di Lollo., V. (1984). Beyond visible persistence: An alternative account of temporal integration and segregation in visual processing. Cognitive Psychology, 26, 33-63.

Evans, B., Drasdo, N. & Richards, I. (1993). An investigation of some sensory and refractive visual factors in dyslexia. Vision Research, 34, 1913-1926.

Fisher, D. (1982). Limited-capacity models of automatic detection: Capacity and scanning in visual search. Psychological Review, 89, 662-692.

Heath, R. (1984). Response time and temporal order judgment in vision. Australian Journal of Psychology, 36, 21-34.

Hughes, H., Nozawa, G. & Kitterle, F. (Submitted). Variations on a theme of parallel processing: Global precedence, spatial frequency channels and the statistics of natural images. ,

Kaas, J. H. &. Pons, T.P. (1988). The somatosensory system of primates. Comparative Primate Biology, Volume 4: Neurosciences(pp. 421-468). Alan R. Liss, Inc.

Kulikowski, J.J., & Tolhurst, D.J. (1973). Psychophysical evidence for sustained and transient mechanisms in human vision. Journal of Physiology, 232, 149-163.

Lamb, M. &. R., L. (1989). Do response time advantage and interference reflect the order of processing of global- and local- level information. Perception and Psychophysics, 46, 254-258.

Lennie, P. (1980). Perceptual signs of parallel pathways. Phil. Trans. R. Soc. Lond. B, 290(23-37), 23-36.

Lennie, P. (1988). Roles of M and P Pathways. In R. &. L. Shapley D. (Ed.), Contrast Sensitivity (pp. 201-213). Cambridge, MA: MIT Press.

Livingstone, M., & Hubel, D. (1988). Segregation of form, color, movement and depth: Anatomy, physiology and perception. Science, 240, 740-749.

Loftus, G. R., Busey, T.A. & Senders, J. (1993). Providing a sensory basis for models of visual information acquisition. Perception & Psychophysics, 54, 535-554.

Lovegrove, W., Garzia, R. & Nicholson, S. (1990). Experimental evidence for a transient system deficit in specific reading disability. Journal of the American Optometric Association, 16, 137-146.

Lynch, J. S., L., Perry, V. & Merigan, W. (1992). Visual effects of damage to P ganglion cells in macaques. Visual Neuroscience, 8, 575-583.

Mansfield, R. (1973). Latency functions in human vision. Vision Research, 13, 2219-2234.

McIntosh, A. R., Grady, C.L., Ungerleider, L.G., Haxby, J.V., Rapoport, S.I., & Horwitz, B. (1994). Network analysis of cortical visual pathways mapped with PET. The Journal of Neuroscience, 14, 655-666.

Merigan, H., & Maunsell, J. (1990). Macaque vision after magnocellular later geniculate lesions. Visual Neuroscience, 5, 347-352.

Merigan, W., Katz, L., & Maunsell, J. (1991). The effects of parvocellular lateral geniculate lesions on the acuity and contrast sensitivity of macaque monkeys. Journal of Neuroscience, 11, 994-1001.

Merigan, W. H., & Maunsell, J. H. R. (1993). How parallel are the primate visual pathways? Annual Review of Neuroscience, 16, 369-402.

Morris, R., & Rayner, K. (1991). Eye movements in skilled reading: Implications for developmental dyslexia. In J. Stein (Ed.), Vision and Visual Dyslexia (pp. 233-242). Boca Raton: CRC Press.

Navon, D. (1977). Forest before trees: The precedence of global features in visual perception. Cognitive Psychology, 9, 353-383.

Palmer, J. (1996). Attentional effects in visual search: Relating search accuracy and search time. In R. Wright (Ed.), Visual Attention London: Oxford University Press.

Pelli, D. G. (1981). Uncertainty in visual detection and identification. Ph.D. dissertation (Cambridge University, Cambridge, 1981).

Ratcliff, R. (1988). Continuous versus discrete information processing: Modeling accumulation of partial information. Psychological Review, 95, 238-255.

Rayner, K. &. Pollatsek, A. (1989). The Psychology of Reading . Englewood Cliffs, NJ: Prentice-Hall.

Robson, J. G. (1966). Spatial and temporal contrast sensitivity functions of the visual system. Journal of the Optical Society of America, 56, 1141-1142.

Roufs, J. (1974). Dynamic properties of vision-V. Vision Research, 14, 853-869.

Rumelhart, D.E. (1970). A multicomponent theory of the perception of briefly exposed visual displays. Journal of Mathematical Psychology, 7, 191-218.

Schiller, P. &. Malpeli, J. (1977). Properties and tectal projections of monkey retinal ganglion cells. Journal of Neurophysiology, 40, 428-445.

Solomon, J. &. Pelli, D. (1994). The visual filter mediating letter identification. Nature, 369, 395-397.

Sperling, G. (1963). A model for visual memory tasks. Human Factors, 5, 19-31.

Sperling, G., & Sondhi, M.M. (1968). Model for visual luminance discrimination and flicker detection. Journal of the Optical Society of America, 58, 1133-1145.

Tolhurst, D. (1973). Separate channels for the analysis of the shape and the movement of a moving visual stimulus. Journal of Physiology, London, 285, 275-298.

Townsend, J. T. (1981). Some characteristics of visual whole report behavior. Acta Psychologica, 47, 149-173.

Ungerleider, L. G., & Mishkin, M. (1982). Two cortical visual systems. In D. J. Ingle R. J. W. Mansfield, & M. S. Goodale (Ed.), The Analysis of Visual Behavior (pp. 549-586). Cambridge, Mass: MIT Press.

Watson, A. (1979). Probability summation over time. Vision Research, 19, 515-522.

Watson, A. B. (1986). Temporal sensitivity. In K. R. Boff, L. Kaufman, and J.P. Thomas (Eds.), Handbook of Perception and Human Performance (Vol I) New York: Wiley.

Watson, A. B., & Nachmias, J. (1977). Patterns of temporal interaction in the detection of gratings. Vision Research, 17, 893-902.

Williams, M. &. Lecluyse, K. (1990). Perceptual consequences of a temporal processing deficit in reading disabled children. Journal of the American Optometric Association, 61, 111-121.

Wilson, H. &. Regan, D. (1984). Spatial-frequency adaptation and grating discrimination: Predictions of a line-element model. Journal of the Optical Society of America A, 1, 1091-1096.

Wolford, G. (1992). A model of visible persistence based on linear systems. Unpublished manuscript.

LIST OF REFEREES FOR FIRST AWARD

Dr. Geoffrey R. Loftus

Department of Psychology

University of Washington

Seattle, WA 98195-1525

(206) 543-8874

gloftus@indiana.edu

Dr. Richard Shiffrin

Department of Psychology

Indiana University

Bloomington, IN 47405

(812) 855-4972

shiffrin@indiana.edu

APPENDIX A- ADDITIONAL REFERENCE ARTICLE

Contents:

Busey, T.A. & Loftus, G.R. (1994). Sensory and cognitive components of visual information acquisition. Psychological Review , 101, 446-469.

APPENDIX B- LETTER OF SUPPORT FROM DR. ARTHUR BRADLEY