Compound Stimulus Processing

Animals both in nature and in the laboratory often have to respond to multidimensional stimuli. For instance, animals in nature typically forage through complex environments where food items must be recognized against various backgrounds. The successful detection of target food items depends on the ability to filter out non-food background items. Furthermore, the animal may need to distinguish between toxic and safe foods that are highly similar in appearance. Therefore, although a stimulus may have a number of properties, such as its color, size, shape, and spatial position, only one of these properties may be relevant for the problem the organism needs to solve. An animal's ability to respond to the relevant property is the outcome of attention. If it can isolate the relevant property and respond only to it, the animal is said to attend to the relevant dimension. On the other hand, if it responds to other properties of the stimulus as much as it does to the critical property, it shows detrimental effects of divided attention and an inability to devote sufficient attention to the relevant dimension. The degree to which animals can successfully divide attentional resources is limited by a number of factors, including the sensory capabilities of the species (see Dukas, 1998 for a discussion of constraints on information processing).

The process of attention played a key role in an early debate about the nature of discrimination learning in animals (see review by Riley & Leith, 1976). In discrimination learning, an animal is presented with two T-Maze

or more stimuli and must learn to respond in different ways to these stimuli. For example, a rat might have to repeatedly choose between black and white goal boxes in a T-maze, with food placed only in the white goal box. Since the positions of the black and white goal boxes change from left to right randomly from one trial to the next, the rat must learn to attend to the brightness of the boxes to find the food. When rats are trained on this problem, it is often observed that they initially turn repeatedly to the right or to the left, a behavior called a position habit. While in a position habit, the rat fails to discriminate between black and white above a chance level of accuracy. The rat eventually breaks out of the position habit and progressively shows a preference for the food-baited white box over the non-baited black box.

Debate about the processes involved as rats learned to solve the T-maze focused on the role of attention in learning. Some theorists proposed that rats gradually learned to solve the maze through experiences with reinforced responses over a number of trials. For instance, Spence (1936) proposed that repeated responses to a stimulus followed by reinforcement (entering the white box and obtaining food in the example above) led to a tendency to approach that stimulus (excitation), and repeated responses to a stimulus that were nonreinforced (entering the black box and finding no food) led to a tendency to avoid that stimulus (inhibition). These tendencies were developed over experience with the two choices. Furthermore, all stimuli impinging on an animal's receptors at the time of response and reinforcement or nonreinforcement acquired excitation and inhibition, which might include stimuli other than box brightness or location. The finding that rats initially displayed position habits in visual discrimination problems was explained as an effect of pre-experimental differences in the tendency to turn right or left. Over repeated training trials, the bias was gradually overcome by the processes of excitation and inhibition that developed to the reinforced and non-reinforced boxes. Thus, what appeared to be a sudden shift in attention from position to brightness could be explained as a change in approach and avoidance tendencies within a continuous learning process.

Theories such as Spence�s that proposed a gradual learning process are referred to as continuity theories. Other theorists, such as Lashley (1929) and Krechevsky (1938), argued that the rat attends to only one dimension of the discrimination problem at a time instead of gradually learning how to solve the task. Thus, the period of a rat's position habit is one in which the animal attends only to the dimension of spatial position (left or right) and not to brightness of the goal box (black or white). The rat will only learn about the correctness of white over black when it switches its attention from the position dimension to the brightness dimension. As Krechevsky (1932) suggested, the rat tried different hypotheses to solve the task and only learned the correct response when it hit upon the correct hypothesis. These theorists considered learning to be noncontinuous and insightful- a distinct shift in attention from one dimension to another. Such non-continuity theories therefore assumed that as long as a rat was persisting in a position habit, it was learning nothing about relative values of black and white.

The numerous studies carried out to test these theories (e.g., Bitterman & Coate, 1950; Ehrenfreud, 1948; Spence, 1945) led to the conclusion that animal attention was not as narrow as originally conceived by Lashley and Krechevsky�s non-continuity theories. Rats appeared to learn about the differential reinforcement of black and white while responding only to position. On the other hand, animals did not have unlimited ability to attend to all stimuli impinging on their receptors, as the continuity position had proposed.

In an attempt to account for animals� ability to focus attention on only relevant stimulus dimensions while learning in a gradual way, Sutherland and Mackintosh (1971) proposed a theory of discrimination learning that defined neural mechanisms that respond to stimuli along different dimensions as analyzers. Their theory proposed that discrimination learning involves two processes, first strengthening the analyzer for the relevant dimension of the discrimination, and, second, building approach and avoidance responses to different values or outputs of the analyzer. In the T-maze example used above, the theory assumes that the brightness analyzer is gradually being strengthened by consistent reinforcement of white and nonreinforcement of black. At the same time, the left-right spatial position analyzer is being weakened by nondifferential reinforcement of its outputs. The theory clearly argues for differential attention to stimulus dimensions; because the strengths of all analyzers sum to a constant value, as one analyzer is strengthened, others must be weakened. On the other hand, an animal may attend to more than one dimension at the same time, because two or more analyzers may have sufficiently high strengths to both control behavior.

Attention in Information Processing

By the 1960s, a new interest in attention as a factor in information processing systems had developed. Although this new approach was concerned with questions about the breadth of attention, as had continuity and noncontinuity theories of discrimination learning, its focus was otherwise quite different from these earlier theoretical concerns. Its emphasis was on attentional processes in humans and not animals. Furthermore, it stressed innate limits on attention and not the role of attention in the learning process. Analogies with computers as information processing systems that had to deal with multiple sources of information simultaneously were drawn, and limits on attention were conceived of as processing bottlenecks or overloads.

Human dichotic listening experiments provided a particularly dramatic demonstration of limited attention. A person was instructed to listen to a message coming in one ear over a headphone and to shadow or repeat that message. If information was played into the other ear at the same time, subjects acquired almost no information about the message presented to the unattended ear (Cherry, 1953). Broadbent�s (1958) filter theory was a particularly important account of these findings. He proposed that a subject attending to a message in one ear selectively blocked or filtered out the information coming in the other ear. Furthermore, these early attentional processes were based on the physical properties of the stimulus and not its semantic content. Later research showed, however, that a person did pick up some information coming in the unattended ear based on semantic content. Thus, a person shadowing a message in one ear would still recognize his own name played in the other ear. Such discoveries led to the subsequent modification of filter theory. Treisman (1969), for example, argued that the unattended channel acts as an attenuator that weakens the unattended message, so that only the most salient information will reach consciousness.

In other experiments involving more than one sensory modality, subjects had to attend to one, two, or four dimensions of auditory and visual stimuli at the same time. In the case of four dimensions, a person could be asked to judge the pitch or intensity of a tone and the horizontal or vertical position of a dot on a TV monitor. People's ability to detect either the visual or auditory stimuli declined as the number of dimensions processed increased (Lindsay, Taylor, & Forbes, 1968). It was suggested that the number of dimensional channels a person can process simultaneously is limited. Thus, detection breaks down as the number of channels to be processed increases.

Soon after researchers began studying attention in human information processing, an interest developed in investigating attention as a factor in the information processing systems of pigeons. A number of different types of experiments have been performed. The hierarchical framework shown in Figure 1attempts to organize these attentional studies on the basis of the combinations of sensory input used. At the lowest level of the hierarchy are intramodal studies, in which a bird must attend to information coming in a single sensory modality. A compound might consist of two visual elements within a compound, such as a color and a shape, or two auditory stimuli of different frequencies presented together. At the next level in the hierarchy, the intermodal level, compound test stimuli may be composed of elements from different modalities. For example, an intermodal compound might consist of a color and tone presented together. At the highest

level of the hierarchy are studies that use the dimensions of space, time, and number (See Figure 1). These dimensions can be thought of as more abstract than the dimensions at the intermodal and intramodal levels because their signals can be carried by stimuli from any sensory modality (Roberts, 1998). For instance, a spatial location can be indicated by a visual stimulus presented at a specific location or by a tone emanating from a specific location. Likewise, counting and timing can be carried out with stimuli from either visual or auditory dimensions. These abstract dimensions may be combined with one another or with one of the dimensions lower in the hierarchy to form compound stimuli. The following section reviews findings of such experiments, as well as the theoretical explanations proposed to account for the pigeons� performance.

Intramodal Studies

Most laboratory studies with pigeons have tested attention to two visual elements using a delayed matching-to-sample procedure. In delayed matching-to-sample, pigeons are trained to peck the one of two comparison stimuli that is identical to a previously presented sample stimulus. Pecking the correct comparison stimulus is followed by food. Suppose a pigeon is placed in an operant chamber (Skinner box) that contains a row of three circular keys along one wall. Any one of these keys can be illuminated with a color or a pattern, as determined by a computer program. For instance, the center key might be illuminated with red light as the sample. After the sample is darkened (usually after a certain amount of

sample exposure time or after a certain number of pecks on the sample by the pigeon), two comparison stimuli are illuminated on the side keys. One comparison might be colored red and the other green. A peck to the red comparison (the comparison identical to the sample) would result in the delivery of a food reward to the pigeon. A peck to the green comparison would not be reinforced and would be followed by darkness until the next trial began (see Grant & Kelly (2001) for an animated demonstration of delayed matching-to-sample).

In studies of attention, pigeons typically are trained on two different element matching-to-sample problems and then are presented with samples that are a combination of the two samples used in the training problems. For example, after training on one delayed matching-to-sample problem with red and blue keys as samples and comparisons and another delayed matching-to-sample problem with vertical and horizontal lines as samples and comparisons, possible compound samples could be vertical lines on a blue background, horizontal lines on a red background, and so on. Memory for either element of the compound (i.e. line orientation or color) can be tested by presenting comparisons from one of the

two dimensions (See Figure 2). For instance, a sample of

horizontal lines on a blue background could be followed by blue and red comparisons (a color dimension test) or by vertical lines and horizontal lines on a noncolored background as comparisons (a line orientation dimension test). The correct choice would be either the blue comparison stimulus or the horizontal lines stimulus. (To experience an interactive demonstration of compound matching to sample click here)

Using similar line orientation and color stimuli, Maki and Leith (1973) conducted such an intramodal test of compound stimulus processing. They found that pigeons� matching accuracy was lower when the sample was a compound than when it was one of the two elements (color or lines) used in training. This decrement on compound sample trials was attributed to a division of attention. The information processing capacity of pigeons was assumed to be limited; Maki and Leith proposed that pigeons� attention was divided between the color and line orientation elements of the compound. This division of attention resulted in a weaker memory for each element and therefore less accurate matching. A decrease in performance on compound sample trials relative to element sample trials became known as the divided attention effect. Could this detrimental effect of divided attention on element matching performance be overcome? Leith and Maki (1975) and Brown, Cook, Lamb, and Riley (1984) demonstrated that sustained testing with one element of a compound increased performance on the frequently tested element while performance on the non-tested element suffered. Thus, pigeons could attend to one element of a compound, if attention was biased toward that element, but otherwise performed worse on compound tests relative to element tests.

Maki and Leith�s (1973) divided attention hypothesis soon was challenged by others who suggested that the decrement on compound sample trials could be explained on the basis of the pigeons� perception of the stimuli rather than the amount of attention allocated to each element. Proponents of a generalization decrement account of the divided attention effect argued that it was the difference in appearance of the compound sample and the element comparisons that caused pigeons to perform relatively poorly on compound tests. Recall that in delayed matching-to-sample training problems, the samples were identical to the comparisons, while the compound test samples were not identical to the element comparisons that followed them.

Support for this view came from studies which varied the length of time the sample was presented (Roberts & Grant, 1978; Santi, Grossi, & Gibson, 1982). Recall that the divided attention hypothesis suggested that subjects divided their attention between the two elements of a compound. It should follow that increasing the amount of time the sample is illuminated should allow a subject more time to process each element of the compound, and matching performance should increase as the duration of the sample increases. Contrary to this prediction, Roberts and Grant and Santi et al. found that compound sample matching accuracy did not improve more than element sample matching improved at longer sample durations. These results were difficult for a divided attention account of the compound sample matching decrement to explain and were used to support a generalization decrement account.

While results from studies in which the duration of the sample was manipulated weighed against the divided attention hypothesis, they did not directly support a generalization decrement account. Grant and MacDonald (1986) suggested a third hypothesis to explain the compound sample matching decrement. Like the generalization decrement hypothesis, their interpretation was based on the perceptual differences between compound stimuli and element stimuli. Instead of the difference of interest being that between the compound sample and the element comparisons (as was the focus of the generalization decrement hypothesis), they focused on the difference between the element samples used in training and the compound samples used in testing.

Grant and MacDonald�s account was based on the formation of prospective codes during training. For example, in a matching-to-sample procedure, a red key sample may come to elicit a prospective instruction or code to peck a red comparison (see Grant & Kelly, this

volume, for an animated illustration of prospective coding). If a horizontal line is then superimposed on the red key on test trials to form a compound, the �peck red� code is only weakly elicited, and fewer correct choices are made to a red comparison key. As a consequence, performance on compound trials suffers. Additionally, a prospective coding account predicts that differences in element and compound sample accuracy should

not be affected by sample duration: the same code should be elicited no matter how long the sample is illuminated. In support of this account, Grant and MacDonald found that pigeons trained with compound samples and element comparisons showed a decrement in performance when tested with element samples. Figure 3 illustrates the training procedure used by Grant & MacDonald, and Figure 4 shows the percentage of correct responses on element and compound tests over blocks of test sessions. They accounted for this finding by arguing that

the correct code was not elicited by element samples, even though an element sample should not require the division of attention predicted by the divided attention hypothesis. They also found that pigeons showed the same decrement in performance even when the test sample remained illuminated while the comparisons were presented (a simultaneous matching-to-sample procedure). Allowing the sample to remain illuminated should give pigeons plenty of time to process both dimensions of the compound. According to the divided attention hypothesis, the result should be no decrement in performance on compound test trials (an argument similar to that tested by researchers testing the generalization decrement hypothesis by lengthening the sample presentation time). If the sample simply elicits a code to peck a comparison, however, a simultaneous procedure should not eliminate the performance decrement. Grant and MacDonald found that performance still declined on compound test trials. The coding decrement hypothesis accounted for results that contradicted the divided attention hypothesis.

In a more recent investigation of divided attention, Langley and Riley (1993) have provided evidence for the original divided attention hypothesis while controlling for generalization decrement and coding dec

rement explanations. They used a symbolic matching-to-sample procedure rather than the identity matching

procedure used in previous studies. The design of their experiment can be seen in Figure 5. (To experience an interactive demonstration of Langley and Riley's element and compound tasks - click here). In one task, blue or green key samples were followed by lit upper and lower halves of comparison keys. In the other task, a triangle or circle sample was followed by vertical and horizontal lines as comparisons. Since samples and comparisons were taken from different dimensions, the comparisons should be no more perceptually different from compound samples than from element samples, thereby ruling out a stimulus generalization decrement explanation. In order to rule out the possibility that pigeons were forming prospective codes to element samples that would later fail to be elicited by compound samples, subjects were trained with element and compound samples from the beginning. With these procedural modifications, subjects still showed poorer matching accuracy on compound stimulus trials than on trials with both types of elements. It seems, therefore, that pigeons may indeed be dividing their attention between the two elements of a compound stimulus. This shared attention hypothesis and others suggested to account for the compound sample test trial decrement are summarized in Table 1.

Table 1. Three hypotheses proposed to account for pigeons' accuracy deficits on compound sample matching tests
Shared Attention	Maki & Leith (1973), Langley & Riley (1993)	Attention is divided between the two elements of a compound test sample, resulting in a weakening of memory for each dimension and therefore less accurate matching.
Generalization Decrement	Roberts & Grant (1978), Santi et al (1982)	On compound test trials, the difference in appearance between the compound sample and the element comparisons results in less accurate matching relative to element sample trials.
Coding Decrement	Grant & MacDonald (1986)	Compound samples do not elicit the same prospective code as do element samples resulting in less accurate matching on compound sample trials relative to element sample trials.

Recently, Roberts (1998) has offered an interpretation of the shared attention hypothesis that may account for seemingly contradictory findings, especially those from experiments in which the duration of the sample was lengthened. He suggested that switching back and forth between processing one dimension and another dimension may results in decay of the memory for the non-processed dimension at any given point during a sample presentation. If repeated switches in processing continued for the entire duration of the sample, the deficit in accuracy on compound trials relative to element trials should persist at longer sample durations. Also, presenting the sample and comparison stimuli simultaneously should not necessarily produce higher accuracy with compound samples, since the animal would still switch back and forth between the two elements of the compound, resulting in lower accuracy on each one.

Intermodal Studies

An intermodal compound stimulus is made up of two elements from different sensory modalities. For example, Kraemer and Roberts (1985) conducted an intermodal study by combining stimuli from the visual and auditory modalities. A group of pigeons was trained on two matching-to-sample procedures with sample stimuli that consisted of the color of a key (red or green) in one task and the frequency of a tone (high or low) in the other. Notice that this experiment again used a symbolic matching-to-sample procedure in which the samples and comparisons were not identical. Thus, pigeons had to choose between vertical and horizontal line comparison stimuli following red and green sample stimuli and between yellow and blue comparison stimuli following high and low tones. When the tone and color samples were presented in compound and only one dimension was tested, pigeons� accuracy did not decline on color tests, but performance was significantly worse on tone frequency tests than performance on tone alone control tests. Unlike previous studies that used two visual samples to make up a compound, Kraemer and Roberts found that one dimension (color) seemed to dominate processing of a tone/color compound. Interestingly, Foree and LoLordo (1973) and Shapiro, Jacobs, and LoLordo (1980) found that tone stimuli acquired more control over responding than did light when the reinforcer was shock avoidance. These findings illustrate that the stimulus that controls responding can differ depending on the reinforcer.

Studies with Abstract Dimensions

As illustrated in the hierarchical framework, the dimensions of space, time, and number may also be used as elements in compound stimuli. Kraemer, Mazmanian, and Roberts (1987) tested pigeons with compound stimuli and used the abstract dimension of space as one element. Pigeons were trained with left and right keys illuminated with white light as samples and a triangle pattern and three dots in a diagonal pattern as comparisons. Pigeons matched based on the left-right location of the sample; thus, the triangle would be the correct comparison choice after the sample appeared on the right key, but the diagonal dot pattern would be correct after the sample appeared on the left key. In another task, the birds were trained to match red and green comparisons to blue or yellow colored samples on the center key. Here, they matched based on the color of the sample. After learning both tasks, pigeons were tested with compound samples consisting of blue or yellow illumination on the left or right key. Pigeons showed no deficit in matching accuracy on test trials in which the dimension tested was either color or space. Spatial location and color appear to have been processed simultaneously, with no decrement on either dimension.

Sutton and Roberts (1998) investigated attention to compound stimuli composed of time and visual pattern elements (Experiments 1 and 3a) and time and spatial location elements (Experiments 2 and 3b). They found that pigeons could process the duration of a stimulus and the orientation of a line (vertical or horizontal) presented on that stimulus simultaneously. The subjects required explicit duration training with the line stimulus, however, in order to accomplish this simultaneous processing. That is, they were trained on a duration matching problem with a white center key illuminated for 2 or 10 seconds as the sample and with red and green comparison keys. The same pigeons also were trained on a line orientation matching problem, with a vertical or horizontal line on the center key as the sample and with choice between yellow and blue comparison stimuli. On compound tests, the vertical or horizontal line was presented for 2 or 10 seconds, and either the line orientation dimension was tested by presenting yellow and blue comparisons, or the time duration dimension was tested by presenting red and green comparison stimuli. The pigeons were unable to accurately match the duration of these compounds.

It was felt that pigeons might need temporal training with the specific visual stimuli used in the compounds. Thus, these pigeons then were given further training to match the duration of a vertical or horizontal line stimulus. Eventually they were tested on time/line orientation compounds in sessions where they were unaware of the dimension tested on a given trial until the comparisons were illuminated. Now, pigeons matched the time element of the samples as accurately as they matched time in sessions that contained only time tests. The same pattern of results was found with birds tested on space/time compounds; birds needed explicit duration training with left and right

stimuli in order to process both dimensions simultaneously. Results of these tests are presented in Figure 6. Table 2 summarizes the test procedures used by Sutton and Roberts (1998). The results of tests with both compounds suggest that attention needs to be directed to the time dimension of a stimulus in order to match based on duration, but that simultaneous processing is easily shown once that training has occurred.

Table 2. Sutton & Roberts (1998) test procedure (Experiments 3a and 3b)
	Line orientation/duration test	Spatial location/duration test
Sample	Vertical or Horizontal for 2 or 10 sec	Left or Right white stimulus for 2 or 10 sec
Comparisons	Duration test (red and green) or Line orientation test (blue and yellow)	Duration test (red and green) or Spatial location test (blue and yellow)

IV. Other Studies That Combine Abstract Dimensions Within The Hierarchical Framework

A few other studies with pigeons have combined abstract dimensions while not directly focusing on attention to each dimension. The intention often was to determine simply whether an animal attended to both dimensions presented or how it might average information from different dimensional cues presented simultaneously. For instance, Cheng, Spetch, and Miceli (1996) used a touch screen apparatus to present a compound stimulus using the dimensions of time and space. On training trials, a rectangle moved across the screen at a constant rate of 1 cm per second. Subjects were rewarded for the first peck to the target after 10 seconds had passed. Since the rectangle was moving at a constant rate, not only could the pigeons use how much time had passed to determine when to respond, but they could also attend to the location of the rectangle, since it was always in the same location just prior to reinforcement. Cheng et al. were interested in which cue, if not both, the pigeons used to determine when to peck. On test trials, the rectangle moved across the screen faster or slower than in training. Specific predictions about the peak of the pigeons� responding were made based on whether they were using time or location. That is, if the subjects were using the duration of the stimulus presentation, the response peak on slow rate tests should have been before the target had passed its location of reinforcement from training. On fast rate tests, the peak should occur after the rectangle had reached the same location. Alternatively, if the pigeons were using the location of the stimulus to determine when to peck, their responding should peak late (in terms of time) on slow trials and early on fast trials. The results showed that the pigeons tended to average the two dimensions but showed a slight bias toward the time cue. Cheng et al. concluded that the pigeons had learned to use both the time and space cues in training to determine when to respond.

Time and number were investigated in compound by Roberts and Mitchell (1994) with pigeons. Pigeons were trained with flashes of a houselight that flashed at the rate of 1/second. Pecking on one key was reinforced after two flashes in 2 seconds and pecking on another key was reinforced after eight flashes in 8 seconds. Thus, time and number cues were confounded, in that the low number and short time cue signaled reward for one response and the large number and long time cue signaled reward for the other response. To unconfound the time and number dimensions, tests were performed in which one dimension was held constant, while the other dimension was varied. To measure control by time, tests consisted of four flashes that lasted for 2, 3, 4, 5, 6, 7, or 8 seconds. To measure control by number, tests consisted of a 4-second presentation of 2, 3, 4, 5, 6, 7, or 8 flashes. Psychophysical curves plotted proportion of trials on which the large number/long time key was chosen as a function of time and number. These curves showed equivalent control by time and number. Meck and Church (1983) found similar results with rats, while Breukelaar and Dalrymple-Alford (1998) found that rats� attention to time may be stronger than attention to number.

While the abstract dimensions of time, space, and number have been investigated as elements in compound sample stimuli, only a few studies have directly focused on attention to these dimensions. The few studies reviewed here suggest that attentional processes may interact in a complex fashion with the types of dimensions placed in compound with one another. Future research on the processing of multiple dimensions and the limits to that processing may further demonstrate how pigeons process different types of information simultaneously and may provide more clues to the organization of information processing systems. Pigeons� attention is most certainly divided between widely varying stimulus types in nature, and survival may depend on the appropriate division of attentional resources. Future studies may also address the limits to divisions of attention in nature as well as in the laboratory.

Bitterman, M. E., & Coate, W. B. (1950). Some new experiments on the nature of discrimination learning in the rat. Journal of Comparative and Physiological Psychology, 43, 198-210.