| HOME | HELP | FEEDBACK | SUBSCRIPTIONS | ARCHIVE | SEARCH | TABLE OF CONTENTS |
|
|
|
|||||||
|
|||||||||
1Howard Hughes Medical Institute, Systems Neurobiology Laboratories, The Salk Institute for Biological Studies, La Jolla, California; and 2Helmholtz Institute, Functional Neurobiology, Utrecht University, Utrecht, The Netherlands
Submitted 15 July 2005; accepted in final form 20 September 2005
 |
ABSTRACT |
|---|
 TOP ABSTRACT INTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
|
INTRODUCTION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
10–2 cd/m2) with that at high noon (105 cd/m2). Most artificial devices are unable to function over such a range because they lack a fundamental property of the visual brain: adaptation.
It has been hypothesized that the visual system adapts to make efficient use of its limited resources (Barlow and Földiák 1989
): if the environment is dark, encoding information on bright objects is useless and sensitivity for dark objects is critical. In other words, to represent information with neurons that have only a limited response range, it pays to zoom in on the properties of the environment and only represent properties that occur often enough to matter but not too often as to be meaningless. Such a change in the neural representation, however, may also have negative consequences: some of the incoming signals may be misinterpreted. We studied this dual view of adaptation at the single cell level.
In particular, we focused on the role of adaptation in the perception of speed. After adaptation, a moving stimulus may appear to move at a different speed (Goldstein 1957; Rapoport 1964; Thompson 1981). We consider this an example of an undesirable side effect of adaptation: a misinterpretation of subsequent visual stimuli. Speed adaptation, however, also has a beneficial effect: after prolonged exposure to moving stimuli, speed discriminability is increased (Bex et al. 1999; Clifford and Langley 1996; Clifford and Wenderoth 1999; Huk et al. 2001). The latter effect is much more subtle than changes in perceived speed, which may explain why it has not been studied in as much detail. Nevertheless, we believe that the visual system's main goal in motion adaptation is not a reduction in perceived speed but rather improvements in speed discriminability.
We hypothesized that these dual perceptual changes are a consequence of neural changes in cortical visual area middle temporal area (MT). To study this hypothesis, we developed a paradigm that was amenable to testing in awake, behaving monkeys. The main hurdle we faced in doing so was that the long adaptation periods (30s) used in human research are not feasible with awake monkey subjects. We thus resorted to much shorter adaptation periods (2 s). In the first part of our report we document that these brief adaptation periods elicit small but robust perceptual effects in both humans and monkeys. The qualitative features of brief adaptation were identical to those observed with long adaptation: a reduction in perceived speed, and an improvement in speed discriminability. These findings provided the basis for studying the neural correlates of speed adaptation in macaque MT. We found that MT neurons generally reduced their firing rates even after brief adaptation. Surprisingly, these firing rate reductions were largest for speeds that were different from the adaptation speed. Therefore the changes in the speed tuning curves with adaptation could not be described as a simple scaling factor. Using an ideal observer analysis, we could relate the changes in the shape of the tuning curves to both the reduction in perceived speed and the improvement in speed discriminability.
|
METHODS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
Two adult male rhesus monkeys (Macaca mulatta; monkeys M and S) were used in the psychophysical and electrophysiological experiments. Experimental protocols were approved by the Salk Institute Animal Care and Use Committee and conform to U.S. Department of Agriculture regulations and to the National Institutes of Health guidelines for humane care and use of laboratory animals.
A total of five naïve human subjects and one author participated in the psychophysical experiments. The participants gave their informed consent, and all procedures were in accordance with international standards (Declaration of Helsinki) and National Institutes of Health guidelines. All subjects had normal or corrected-to-normal visual acuity.
Surgical preparation
The surgical procedures have been described in detail elsewhere (Dobkins and Albright 1994). In short, a head post and a recording cylinder were affixed to the skull using stainless steel rails, screws, and dental acrylic (monkey M) or CILUX screws and dental acrylic (monkey S). Recording chambers were placed vertically above the anatomical location of area MT as determined from structural MR scans (typically 4 mm posterior to the interaural plane and 17 mm lateral to the midsagittal plane) to allow for a dorso-ventral electrode trajectory. After surgical recovery and attainment of criterion performance on the visual fixation task, a craniotomy was performed to allow for electrode passage into area MT. All surgical procedures were conducted under sterile conditions using isoflurane anesthesia.
Visual stimulation
All visual stimuli were generated with in-house OpenGL software using a high-resolution graphics display controller (Quadro Pro Graphics card, 1,024 x 768 pixels, 8 bits/pixel) operating in a Pentium class computer. In the experiments with monkey subjects, stimuli were displayed on a 21" monitor (Sony GDM-2000TC; 75 Hz, noninterlaced). In the experiments with human subjects, we used a Sony Trinitron E500 monitor. The output of either video monitor was measured with a PR650 photometer (Photo-Research, Chatsworth, CA), and the voltage/luminance relationship was linearized independently for each of the three guns in the monitor. Stimuli were viewed from a distance of 57 cm in a dark room (<0.5 cd/m2).
Monkeys were seated in a standard primate chair (Crist Instruments, Germantown, MD) with the head post rigidly supported by the chair frame. Eye position was sampled at 60 Hz using an infrared video-based system (IScan, Burlington, MA), and the eye position data were monitored and recorded with the CORTEX program (Laboratory of Neuropsychology, NIMH; http://www.cortex.salk.edu/), which was also used to implement the behavioral paradigm and to control stimulus presentation.
Stimuli and experimental paradigms
The stimuli used to study speed adaptation were random dot patterns. Each pattern consisted of 100 dots moving coherently in one direction within a 10°-diam circular aperture. The dots were 0.15° in diameter and were 70% more luminous than the gray background (5 cd/m2). The direction of motion of the dots could be adjusted to match the preferred direction of motion of the cells rounded to the nearest multiple of 45°. The speed of the dots was 1, 2, 4, 8, 16, 32, or 64°/s.
Training paradigm
The monkeys were trained to report which of two moving random dot patterns on the screen moved faster. During training, the monkey fixated a central red dot, and two random dot patterns appeared to the left and right of fixation. One moved at least twice as fast as the other. After 0.5 s, the patterns disappeared and were replaced by two red dots. Monkey M was rewarded with a small amount of liquid for making a saccade to the dot at the position of the faster moving pattern. The monkey was required to execute this response within 2,000 ms after onset of the dot target and to maintain fixation on the chosen target for 500 ms. Monkey S was trained to hold two touch bars during stimulus presentation and to release the touch bar below the faster stimulus to indicate his choice. After the monkeys reached criterion (80% correct) performance on this task, we introduced the adaptation phase. At first, we used only adaptation to stationary patterns to introduce the monkey to the longer fixation times required in these adaptation trials. Then, we switched to the complete behavioral paradigm.
Behavioral paradigm to measure changes in perceived speed
Humans and monkeys were tested in almost identical paradigms that consisted of five phases, referred to as the prepare, adaptation, blank, test, and response phases (Fig. 1). We will describe these phases in turn. The differences between the human and monkey paradigms were restricted to the response phase.
|
After fixation had been achieved, the adaptation phase started. Two random dot patterns appeared 10° left and right of fixation for 2 s. The two experimental conditions (AdaptNone, AdaptOne) differed only in terms of the kind of stimuli presented during this adaptation phase. In the control condition (AdaptNone), both random dot patterns were stationary. Because there is no adaptation to motion in this condition, we refer to it as the nonadapted condition. In the AdaptOne condition, one of the random dot patterns was stationary whereas the other random dot pattern moved to the right at a constant adaptation speed (Fig. 1, speed A). Stationary and moving patterns were positioned left and right of the fixation point in a pseudorandom manner and combined in the off-line analysis. In the 0.5-s blank phase that followed the adaptation phase, only the fixation dot was on the screen.
In the test phase, two random dot patterns appeared on the screen, also at 10° left and right of fixation, both moving rightward. For the AdaptOne condition, one pattern, called the reference pattern, was positioned where the stationary pattern was during the adaptation phase. Because of the spatial specificity of motion adaptation (Kohn and Movshon 2003), this reference pattern was therefore nonadapted. In this set of experiments, the speed of the reference pattern was equal to the speed of the adaptation stimulus. In the notation of Fig. 1, R = A. To determine the influence of adaptation on perceived speed, we presented a test pattern at the position where the moving pattern was during the adaptation phase. The dots in this test patch moved at 60, 80, 90, 95, 100, 105, 110, 120, or 140% of the reference speed. In the notation of Fig. 1, the speeds T were arranged around R to measure a psychometric curve for speed discrimination. For the monkey, we replaced the 95 and 105% speeds with the easier 30 and 170% speeds.
After the test patterns had been switched off, the response phase started. Human subjects pressed one of two buttons in a two-alternative forced choice to indicate which of the two patterns (left or right) moved faster. Monkey M communicated his decision regarding which stimulus was faster by making a saccade to one of two small red dots that appeared 10° left and right of fixation. Monkey S was trained to release one of two touch bars to indicate his choice. We performed the same experiment at a range of adaptation speeds (Fig. 1A): humans at 4, 8, and 16°/s, monkey M at 4 and 16°/s, and monkey S at 8°/s. All subjects repeated each adaptation speed and test speed combination 
20 times.
In the AdaptNone trials we knew what the monkeys should perceive, and they were rewarded for correct decisions only. In the AdaptOne conditions, we tried to measure the monkey's illusion of speed perception. Not knowing how large the monkey's illusion was, we could not rely on physical speed differences to determine the reward. Based on human pilot data, we estimated that the illusory reduction in perceived speed after adaptation would be at most 20% of the adaptation speed. Therefore we decided to reward all trials with a physical speed difference of >20% for (physically) correct decisions only. Trials with a speed difference of <20% were rewarded randomly at a rate of 60%.
Behavioral paradigm to measure changes in speed discrimination
We used a third condition to determine whether speed discrimination improved with a brief 2-s adaptation to motion. This AdaptBoth condition had the same five phases discussed above. In the adaptation phase of the AdaptBoth condition, both patches of random dots moved at the adaptation speed.
In the test phase of this paradigm, two dot patterns (called the reference and the test pattern) that moved at different speeds were shown. For the monkey experiments, the adaptation speed (A) was always 16°/s and the reference speed (R) was 1, 2, 4, 8, 16, 32, or 64°/s. To map out the psychometric curve around this reference speed, the test pattern was assigned a speed (T) that was 5, 10, 20, 40, or 80% slower or faster than the reference speed (R). In all trials of this AdaptBoth condition, both of the patches were adapted. Hence, the reduction in perceived speed should apply equally to both patches. This allowed us to reward Monkey M, who performed this experiment, for veridical speed discrimination in all trials. This consistent reward was crucial to get a consistent level of performance for small speed differences. For the same reason we did not interleave these conditions with the AdaptOne conditions in which a considerable fraction of trials had to be rewarded randomly. As a control, we did interleave AdaptNone conditions with reference (R) and test (T) speeds identical to those in the AdaptBoth conditions.
For the human subjects, the AdaptBoth trials were interleaved with the AdaptNone and AdaptOne conditions. The choice of adaptation, reference, and test stimuli was therefore the same as in those conditions described above.
Electrophysiological paradigm
We recorded the activity of single units in area MT using tungsten microelectrodes (FHC, 3- to 5-M
base impedance), which were driven into cortex using a hydraulic micropositioner (model 650, David Kopf). Neurophysiological signals were filtered, sorted, and stored using the Plexon system (Plexon, Dallas, TX). Off-line spike sorting based on principal components analysis of the waveforms was used to separate up to three cells from a single electrode. We identified area MT physiologically by its characteristically high proportion of cells with directionally selective responses, receptive fields that were small relative to those of neighboring area medial superior temporal area (MST), and its location on the posterior bank of the superior temporal sulcus. The typical recording depth agreed well with the expected anatomical location of MT that was determined from the structural MR scans.
We used automated methods to determine cells' directional selectivity and receptive fields (for details, see Krekelberg and Albright 2005). The receptive field (RF) center and the preferred direction of motion revealed by these methods were used to optimize stimuli for subsequent neuronal response measurements. We determined cells' speed tuning with a random dot pattern positioned at the center of the RF moving in the preferred direction. The dots moved for 500 ms at a constant speed chosen from the set (1, 2, 4, 8, 16, 32, 64°/s). This tuning curve was used to determine the preferred speed. Based on an on-line analysis of the speed tuning curve, we chose one of these speeds as the adaptation speed for the subsequent adaptation paradigm. We aimed to choose a speed that was on the ascending or descending flank of the tuning curve.
To maximize neural data collection, the monkeys did not perform the psychophysical task during recording. Therefore we removed those elements of the display that were relevant for the behavioral task but not for the measurement of the neural responses (Fig. 1). The adaptation stimulus was positioned in the receptive field and it moved in the preferred direction at the chosen adaptation speed (A) for 2 s. After a blank phase of 0.5 s, the test phase consisted of a single test pattern, also positioned at the center of the receptive field, and moving in the preferred direction at one of five speeds. These test speeds (T) were chosen to lie around the adaptation speed. One test speed was identical to the adaptation speed (100%); the other speeds were 20, 60, 140, or 180% of the adaptation speed.
Data analysis
PSYCHOPHYSICAL DATA.
We calculated the percentage of trials in which the subjects responded that the test stimulus was faster for each speed of the test stimulus. We used the psignifit Matlab toolbox (Wichmann and Hill 2001) to fit these psychometric functions with cumulative Gaussians. This fit provided us with an estimate of the point of subjective equivalence (PSE) and the slope of the psychometric function at that point.
To ensure that the monkey's behavioral was under adequate experimental control, we required that the behavioral performance for the easiest conditions (speed comparisons with >70% speed difference) was significantly better than 65% correct. For both the perceived speed experiments and for the discriminability experiment, one session did not match this criterion and was removed from the analysis.
PHYSIOLOGICAL DATA. We analyzed neuronal responses with Matlab (The Mathworks). The measure of neuronal response was the mean spike rate computed within a window of 500 ms (the test stimulus duration) after response onset. Response onset for a given stimulus was defined as the start of the first 50-ms bin in which the firing rate was >3 SD from the baseline firing rate (computed during the initial 250-ms interval during which fixation was maintained but no stimulus was presented in the RF). Response latency was defined as the minimum difference across conditions between stimulus onset and response onset.
Significance of the cell's speed tuning and adaptation was assessed with a Friedman two-factor (speed and adaptation state) ANOVA.
To summarize a cell's speed tuning with three parameters, we used a nonlinear least squares method to fit functions of the following form to the speed tuning data
 |
is the preferred speed, 
is the tuning width, and 
scales the peak firing rate. We defined an adaptation index (AI) to quantify the relative changes in firing rate for every stimulus speed. If the response in an adapted condition is A and the response to the same stimulus without adaptation NA, then AI = 100% x (A – NA)/NA.
Following Liu and Newsome (2003), we defined a cell as having band-pass speed tuning if the firing rate for speeds both slower and faster than the preferred speed dropped <90% of the peak firing rate. A cell whose firing rate for all speeds below the preferred speed was >90% of the peak firing rate was considered low-pass. A cell whose firing rate for all speeds above the preferred speed was >90% of the peak firing rate was considered high-pass.
Linking neural and behavioral data
We used ROC analysis (Green and Swets 1966) to determine whether an ideal observer using a single threshold firing rate could determine whether a neural response was more likely to have been caused by the nonadapted reference stimulus or 1 of the 10 test stimuli (2 adaptation states x 5 speeds).
The area under the receiver operating characteristic (ROC) curve was transformed into a perceptual measure reflecting the percentage of trials in which the ideal observer would decide that the test stimulus was faster. This transformation was based on a simplified vector average model of the representation of speed in MT in which neurons cast their vote not for a particular speed, but for whether a stimulus is faster or slower than the preferred speed. For neurons with a preferred speed above the reference speed, the "fraction of trials in which the test stimulus moved faster" was equated with the area under the ROC curve. For neurons with a preferred speed below the reference speed, the fraction of trials in which the test stimulus moved faster was equated with 1 – area under the ROC curve. This ordinal code reflects our current ability to relate neural to perceptual data (see DISCUSSION).
In the discriminability analysis, we compared the distribution of nonadapted neural responses to a stimulus with speed S1 to the nonadapted neural responses to a stimulus with speed S2. The resulting ROC curve gave a measure of discriminability of S1 and S2 before adaptation. Specifically, we estimated discriminability as the absolute value of the difference between the area under the ROC curve and 0.5. This analysis was repeated using postadaptation distributions of neural responses to S1 and S2.
To quantify the influence of adaptation on discriminability, we subtracted the discriminability of the nonadapted test stimuli from the discriminability of the adapted test stimuli. A value of 0% indicates no effect of adaptation; positive values mean that an ROC observer would be better at discriminating the speeds of adapted versus nonadapted stimuli.
|
RESULTS |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
). The main advantage of the brief adaptation paradigm is that we could use it to show that monkey subjects exhibit similar perceptual effects and study the neural correlates of these effects with intracortical recordings. The second section reports the changes in firing rates of macaque MT cells using the same adaptation paradigm. In the third section, we use an ideal observer analysis to relate changes in firing rate to changes in perception. Perceptual changes
Our psychophysical paradigm is described in detail (see METHODS). Briefly, two patches of random dots were presented for 2 s during the adaptation phase. After a 0.5-s blank interval, two new patches appeared. One patch was called the reference stimulus; the other was the test stimulus. The subject indicated which patch contained the faster moving dots (Fig. 1). There were three conditions, which differed only with respect to the patches that were shown during the adaptation phase. In the control condition (AdaptNone), both patches were stationary during the adaptation phase. In the AdaptOne condition, one patch moved at the adaptation speed, and the other was stationary. This allowed us to measure the change in the perceived speed caused by adaptation. In the AdaptBoth condition, both patches moved at the adaptation speed during the adaptation phase. This third condition allowed us to measure the change in discriminability that follows adaptation, without contamination from a change in perceived speed.
Adaptation reduces perceived speed
Figure 2A shows the performance of one human subject on a speed discrimination task that involved a reference speed of 16°/s and many test speeds. The graph plots the probability that the subject reported the test speed to be faster than the reference speed, as a function of the actual test speed. The green curve shows the data from the nonadapted conditions. The curve reaches the 75% test faster response for a test speed near 17°/s and the 25% test faster for test speeds near 15°/s. Hence, in this experiment, the subject's 75%-threshold for speed discrimination was about 1°/s, which corresponds to 6% of the reference speed. The speed difference at which the psychometric curve crosses 50% test faster is the PSE: a test stimulus moving at this speed was perceived to have the same speed as the reference stimulus. In the AdaptNone condition, the PSE was 15.8°/s, which is close to the actual reference speed of 16°/s. The blue curve represents the data from the AdaptOne condition in which the adaptation speed was 16°/s. Relative to the nonadapted condition (green curve), the blue curve is shifted to the right (PSE = 17.7°/s). In other words, after adaptation, a test stimulus with a physical speed of 17.7°/s was perceived to have the same speed as the nonadapted reference stimulus moving at 16°/s. Adaptation in this subject at the adaptation speed of 16°/s thus led to an underestimation of the test speed by about 10%.
|
20% in the AdaptOne condition, which indicates that the monkey—like the human subjects—perceived a reduced speed after adaptation. To quantify these effects across the population, we determined the PSE from the fitted psychometric curves for each subject and expressed these as a percentage of the adaptation speed. Although the shift in the PSE in humans tended to be somewhat smaller for higher adaptation speeds, this effect was nonsignificant (P = 0.06) and even entirely absent for the monkey subjects (P = 0.7). We therefore pooled adaptation effects over all adaptation speeds. Figure 3A shows a histogram of the adaptation induced shifts in PSE. In all but one experiment the PSE shifted to the right, indicating that subjects typically perceived a reduced speed after adaptation. This effect was statistically significant in both humans and monkeys. (humans mean effect: 16.9 ± 2.8%, P < 0.01; monkeys: 5.7 ± 4.3%, P < 0.05; signed-rank test).
|
Figure 2, A (human subject) and C (monkey subject), documents how adaptation leads to improved discriminability. We first discuss the results obtained in the human subject. The red data points and curve in Fig. 2A show the performance on the speed discrimination task in the AdaptBoth condition with a 16°/s reference speed after adaptation to a 16°/s moving pattern. Compared with the AdaptNone condition (green data points), the psychometric curve was clearly steeper. For test speeds either faster or slower than the reference speed, the subject responded with the correct answer on a larger fraction of the trials. To quantify this effect, we simply determined the difference between the percentage correct answers in the adapted and unadapted conditions for each test speed. For this subject, the average of this difference over all test speeds was a particularly large improvement of 16%.
Figure 2C shows the results of one behavioral session for monkey M. Monkey M's overall performance was clearly not as good as that of the human subjects. To perform at 75% correct, this monkey required a speed difference of 20%. Monkey S had very similar thresholds (data not shown). As is evident from the figure, after adaptation (to 16°/s), monkey M's psychometric curve (red) was steeper than without adaptation (green). This means that—after adaptation—smaller speed differences led to higher levels of performance. Moreover, the higher asymptote of the red curve shows that, after adaptation, the monkey reached levels of performance for test speeds >16°/s that were never reached in the unadapted conditions. Both the steeper slope and the higher asymptote indicate an improvement in speed discrimination. We quantified this improvement again by simply comparing the percentage correct in the adapted and nonadapted conditions. In this experiment, adaptation improved Monkey M's performance on the speed discrimination task by 17%.
The improvement data are summarized across the population in Fig. 3B. Averaged across all conditions, the human subjects showed a median improvement of 3% (P < 0.01; signed-rank test), whereas monkey M showed an improvement of 6% (P < 0.01; signed-rank test).
Making use of the large number of trials obtainable from the monkey subject, we also studied whether adaptation at one speed improves discrimination performance at different speeds. Specifically, for a single adaptation speed (16°/s), the monkey was tested on speed discriminations involving reference speeds that were more than twice as fast or slow (see METHODS). No statistically significant improvement was found for any of those speeds (data not shown). In other words, adaptation at 16°/s only improved discrimination performance significantly (P < 0.05, 2-way ANOVA) for comparisons that involved the 16°/s reference speed and speeds in a range close to it. In the physiological experiments, we only studied the range of speeds close to the adaptation speed (see METHODS).
To summarize, our behavioral experiments show that 2 s of adaptation to a moving stimulus leads to a perceptual underestimation of absolute speed but an improved ability to discriminate the adapting speed from other speeds. This phenomenon was observed in both humans and monkeys. Having thus documented that the perceptual phenomenon of interest occurs in both humans and monkeys, we turned to the neural signals that we hypothesized to be responsible for these perceptual changes.
Neural changes
To study the neural basis of changes in speed perception with adaptation, we recorded from 83 cells in area MT of two monkeys (S: 41 cells; M: 42 cells) at eccentricities ranging from 3 to 15°, with a median eccentricity of 8°. We analyzed and classified the basic speed tuning properties of these cells using standard procedures explained in METHODS. As expected, and in agreement with earlier findings (Rodman and Albright 1987), a majority of the cells (69/83; 83%) was significantly speed tuned. We restricted our analyses to these cells. Most cells (77%) were band-pass speed-tuned, 4% were low-pass tuned, and 19% high-pass tuned. These relative proportions of tuning types agree well with the proportions found in another recent study of speed tuning in area MT (Liu and Newsome 2003). The distribution of preferred speeds was broad; the quartile range extended from 7.5 to 30°/s, with a median of 19°/s. This too is in qualitative agreement with previous findings (Churchland and Lisberger 2001; DeAngelis and Uka 2003; Rodman and Albright 1987; Van Essen 1985) from area MT.
Adaptation reduces firing rate
Adaptation typically reduced the firing rate of MT neurons. Averaged over all cells and all test speeds, a reduction in firing was observed in 70% of cases. To assess the statistical significance of rate changes with adaptation, we performed a two-way ANOVA with speed and adaptation state (adapted/nonadapted) as the factors (see METHODS). This analysis showed that 27 cells (39%) showed a statistically significant change in response rate after adaptation. For cells exhibiting a statistically significant adaptation effect, a rate reduction was observed in 87% of cases; hence adaptation almost never led to an increase in firing rate.
To control for the possibility that the animals' fixational eye movements contributed to the adaptation effects, we analyzed eye position and microsaccades for each recording. The average eye position did not depend significantly on stimulus condition in any recording. For the speed and direction of small eye movements (within the fixation window), we found a significant effect for only 3 of 34 recordings. There was no indication that these recordings with a significant change in microsaccade distribution also were the recordings with the significant adaptation effect. In fact, of the cells with a significant adaptation effect, none had a significant microsaccade effect. Overall, there was no indication that eye movements were in any way correlated with the presence or absence of adaptation effects.
The magnitude of the adaptation effect often varied with the test speed. Four typical examples, covering the range of effects we observed, are shown in Fig. 4. These graphs plot the average firing rate in a 500-ms response window as a function of the physical speed of the random dot stimulus (moving in the preferred direction). Solid lines connect data points that were recorded without previous adaptation to motion; dotted lines connect data points recorded after 2 s of adaptation to motion (in the preferred direction). The arrows indicate the adaptation speed. Figure 4, A and B, shows data from cells that preferred higher speeds; after adaptation, the firing rate showed a general decrease. Figure 4, C and D, shows data from cells that preferred lower speeds; firing rates were reduced after adaptation.
|
Adaptation changes speed tuning
In this section, we aim to find a parsimonious description of the relation between the adaptation effect and the other properties of the neurons. The analyses reported in this section were applied to all 69 cells, but we found qualitatively similar results in the subset of significantly adapting cells. We first tested two simple intuitive hypotheses. The first hypothesis was that the adaptation is a fixed, test-speed independent, reduction in firing. This kind of subtractive relationship would, for instance, be expected if the response during the adaptation phase caused an increase in firing threshold. To test this hypothesis, we determined the average (over all test speeds) adaptation-induced reduction in firing per cell, and subtracted this from the unadapted response for each test speed. We compared this predicted adapted response to the actual adapted response. If the subtractive model was correct, one would expect the residuals (the difference between predicted and actual response) to be "noise." That is, they should be normally distributed around zero. This was not the case (Lillie test; P < 0.01). Notably, the adaptation was systematically underestimated at high firing rates. This suggests that a divisive, not a subtractive mechanism is at work.
This leads us to the second hypothesis, which we refer to as neural fatigue. In this model, the adaptation effect increases with the firing rate evoked during the test period. We tested this divisive model in a similar way. First, we determined the best divisive factor per cell by averaging the adaptation ratio over all speeds. Then we predicted the adapted response by multiplying the unadapted responses with the same factor. The residuals of the divisive model were smaller than those of the subtractive model, hence the divisive model was better. However, the residuals were not normally distributed (Lillie test; P < 0.01). This suggests that this model also fails to capture a systematic component in the data. In other words, the dependence of adaptation on the test speed is more complex than a simple subtraction or division by a constant factor.
To study the dependence of adaptation on test speed more thoroughly, we determined the adaptation index for each test speed, sorted the indices by the unadapted response at that test speed and binned and averaged the adaptation indices over all cells. These average adaptation indices are represented by the solid curve in Fig. 5A. The rightmost data point contains the average adaptation index for conditions in which the unadapted neural response for a given cell was 85–100% of the maximum response for that cell. Hence, it represents the typical magnitude of adaptation found near the peak of a cell's tuning curve. The leftmost data point represents the average adaptation found for test speeds that were far from a cell's optimal speed. Figure 5A strongly argues against adaptation as neural fatigue. There was a significant (Kruskal-Wallis ANOVA; P < 0.01) effect of the normalized firing rate on the adaptation index, but it was opposite to what would be expected if fatigue were the underlying mechanism. The adaptation was largest (the index most negative) at nonoptimal speeds, reached a minimum for stimuli with intermediate efficacy then became larger again at the peak of the tuning curve.
|
In Fig. 5B, we tested this interpretation of the test speed dependence of adaptation more directly. The scatter plot shows a cell-by-cell comparison of the AI for the condition in which the test speed equaled the adaptation speed versus the AI observed at the other four test speeds. Red dots represent cells that have statistically significant adaptation effects. The first observation is that the data points are not homogeneously distributed (
2 test, P < 0.001): most data points are found in the bottom left quadrant. This confirms that, for most conditions, firing rates decreased with adaptation. Within the bottom left quadrant, 50 data points are in the lower octant, whereas nearly twice as many (94) are in the upper octant. This shows that most cells exhibited less adaptation when the test stimulus moved at the adaptation speed than when the test stimulus moved at any of the other four test speeds. This was confirmed statistically with a signed-rank test (P < 0.01).
Figure 5B suggests that the adaptation effect increases when the test speed is different from the adaptation speed. However, the alternative model suggested by Fig. 5A, that adaptation increases with distance from the flank of the tuning curve, has not yet been disproved. To address this issue, we first defined the flank as a firing rate of 75 ± 10% of the peak firing rate. We selected the cells in our population where the adaptation speed and the flank of the tuning curve did not coincide. Only these recordings provide information to distinguish the two models. For these cells (n = 32), we repeated the analysis in Fig. 5B and confirmed a statistically significant reduction of adaptation for test speeds near the adaptation speed (P < 0.01). We performed the same analysis to compare the adaptation effect on the flank with that found for the other four test speeds. There was no statistically significant effect (P = 0.31). This analysis shows that the adaptation depended on the difference between test and adaptation speed, not on whether the test speed was close to the flank of the tuning curve.
This trend can be observed directly in the single cell examples of Fig. 4. For instance, the cell in Fig. 4C adapted least at the adaptation speed. There were, however, clear exceptions to this rule. For instance, adaptation in the cell shown in Fig. 4D clearly increased with distance from the preferred speed rather than the adaptation speed.
To determine how general the test speed dependence of adaptation was, we performed the analysis shown in Fig. 5 on various subsets of the data. First, we studied whether the response window used for the analysis was relevant by using only the transient (1st 100 ms) or only the sustained (last 400 ms) part of the response to the test stimulus. The adaptation effects in the transient and sustained windows were not significantly different (P = 0.4, signed-rank test) and both response windows showed effects that were qualitatively similar to those shown in Fig. 5. Second, we analyzed the adaptation separately for the subset of cells with low preferred speeds (<16°/s) and high preferred speeds (16°/s). The test speed dependence of adaptation for both subsets was qualitatively similar to that shown in Fig. 5. To summarize, these analyses show that the finding that adaptation is largest for test speeds other than the adaptation speed is a general one; it does not depend on the window of analysis, nor is it restricted to particular subclasses of cells.
Parametric changes in speed tuning
The analysis in the previous section allowed us to compare the adaptation effects evoked at various test speeds. Additionally, we performed a parametric analysis to quantify global changes in tuning at the population level. We fitted tuning curves (see METHODS) to the mean responses, separately for adapted and nonadapted conditions. Because the test speeds were chosen per cell to lie on the flank of the tuning curve, not every recording resulted in a good coverage of the tuning curve. To avoid over interpreting poorly fitted curves, we used only cells for which the fitted curve explained >75% of the variance (n = 49). For these cells, the fitted tuning curves provided an estimate of the peak firing rate, the preferred speed, and the width of the tuning curve. Figure 6 compares these three quantities in the adapted and nonadapted state. In Fig. 6A, we plotted the adapted peak firing rate as a function of the nonadapted peak firing rate. Most cells lie below the diagonal line; hence adaptation reduced peak firing. This effect was small, but significant at the population level (n = 49; 0.7 Hz; P < 0.01) and for the subset of significantly adapting cells (n = 19; 2.9 Hz; P < 0.01). Figure 6B shows that adaptation did not significantly effect the preferred speed (P > 0.7). Finally, Fig. 6C shows that the width of the tuning curve was reduced after adaptation. This was a nonsignificant trend at the whole population level (n = 49; 2.7°/s; P = 0.09), but significant for the cells that showed a statistically significant adaptation effect (n = 19; 4.3°/s; P < 0.05). These results show that adaptation led to narrower tuning curves, with a lower peak, but the same preferred speed.
|
To compare the speeds of two stimuli—A and B—a decision mechanism must compare a neural response to stimulus A with a neural response to stimulus B. One technical difficulty is that on any given trial, we, the experimenters, only obtained information about the response to one stimulus (say A, the stimulus that was in the receptive field of the cell under study in this trial). To study speed discrimination with such data, we used information from trials in which stimulus B was in the receptive field. Of course, the brain of an observer doing a discrimination task cannot solve the problem this way. One conceptual way to link this analysis to a decision process that could be implemented in the brain is to use a so-called antineuron. In this conceptualization one assumes that there is a neuron that is identical to the neuron under study in all respects, except for its tuning in the stimulus dimension about which a decision is to be made. Britten et al. (1992) were the first to use this technique to relate the firing of direction tuned neurons to perceptual decisions about direction. In that case, direction was the relevant stimulus dimension; hence they introduced antineurons with a direction tuning opposite to the direction tuning of the neuron under study. Because the decision we tried to model was one of position (where is the faster stimulus?), our antineuron had the same speed tuning but a different spatial receptive field. Hence for the decision A versus B, we used the responses of our neuron from trials in which stimulus A was in the receptive field and for the antineuron we used the trials in which stimulus B was in the receptive field of the neuron. With this foundation, we modeled speed discrimination as a comparison of the neural responses to stimulus A versus those to stimulus B. If these responses were very different, it should be possible for a decision mechanism that compares inputs from the neuron and its antineuron to infer that the stimuli were different.
We first wanted to determine whether it is reasonable to suppose that the monkey used a population of MT cells to perform a speed discrimination task in which a test and a reference speed are compared. We compared the monkey's behavioral performance with the performance of an ideal observer that based its decisions on the responses of individual MT neurons. To be specific, we determined the expected performance of an ideal observer by calculating an ROC that compared the distribution of responses to the reference stimulus with the distribution of responses to a test stimulus. The area under the ROC curve estimates the percentage correct performance of an observer that uses a single (but arbitrary) threshold firing rate to determine whether a particular response is more likely to have been caused by stimulus A than stimulus B (Green and Swets 1966). The ROC approach is reasonably general, biologically plausible, and has been used in other contexts to link neural firing with perceptual decision processes (Britten et al. 1992, 1996; Croner and Albright 1999; Thiele et al. 2000). The percentage correct (discriminability) of the ideal observer is of course bounded by 50 (chance level) and 100% (perfect performance). Thirty-eight cells were tested with the same reference speed of 16°/s. In Fig. 7, we compared the performance of the ideal observer based on those 38 cells with the behavioral performance of a monkey subject presented with the same set of reference and test stimuli.
|
80% of the trials correct when the difference in speed was 40%. However, there was a considerable spread in the performance over days. In comparing these behavioral data to the neural data, it is important to recognize that not every cell is suitable for every speed comparison. In particular, if the two speeds to be compared are near the (flat) peak of a cell's tuning curve, the firing rate difference will be small, and hence, the ideal observer will not obtain much information from this cell. Compare for instance the cell in Fig. 4B with that in Fig. 4D. The cell in Fig. 4B peaked near 22°/s and there was only a small difference in firing evoked by a 22.4 and 16°/s stimulus. Hence, this cell provided little information for this particular speed discrimination. The speed tuning curve of the cell in Fig. 4D, on the other hand, peaked at 3.2°/s, and it had its steepest slope near 16°/s. This cell provided most information for comparisons of 16°/s with higher speeds. To show this, we calculated each cell's performance for all speeds and divided the cells into two subpopulations on the basis of their discrimination performance for speeds lower and higher than the reference speed. If a cell was best at discriminating higher speeds from the 16°/s reference speed, it was assigned to the "above" population (n = 11). If a cell performed best at discriminating lower speeds from the 16°/s reference speed, it was assigned to the "below" population (n = 27). Note that this subdivision was based on the neurons' performance on a perceptual task (the ROC analysis) and not whether these cells are tuned for high or low speeds. This subdivision is clearly arbitrary and only serves to show the point that subdividing the population can be beneficial. On the basis of our data, we cannot make any claims whether the brain actually follows the procedure we used to define the subpopulations.
Figure 7 shows the performance of the below population in blue. For the discrimination that involved the lowest test speed (3.2°/s) and the 16°/s reference, the ROC observer's performance spanned a range that was similar to that of the monkey. Typically, however, these cells' tuning curves reached their peaks in the range of 16–32°/s. The resulting shallow slope of the speed tuning curve in the range of high speeds explains why an ROC observer that relies on these below cells performs relatively poorly on the discrimination of 16°/s with 22.4 or 28.8°/s. The performance of an ROC observer using cells from the above population is shown in green. These cells peaked at lower speeds and exhibited a large difference in firing rate between 16 and 28.8°/s. This difference in firing rate enabled the above population to provide information where the below population could not. Even for these two small populations, the differences in performance was statistically significant (P < 0.01) for all but the 9.6°/s test speed (2-way ANOVA; interaction; P < 0.01, followed by Tukey-Kramer multiple comparison tests).
This analysis shows that for part of the range of behavioral performances of the monkey, there were single MT cells that provided enough information. On some days and in particular for the comparisons involving the smallest speed differences, however, the monkey outperformed most of its cells. This suggests that the monkey's decision must be based on pooled information provided by a number of cells and that this number may be larger for the more difficult discriminations. Figure 7 also shows that different pools of cells provide pertinent information for different speed discriminations. Hence selecting specific subpopulations for pooling would be a worthwhile strategy. With current knowledge of MT, however, it is not clear how the speed information is pooled nor is it clear whether and how the decision process selects suitable subpopulations for pooling.
In what follows, we will side-step these pooling issues and simply determine the average performance based on single cells. This average provides insight into the performance that could be expected from the average cell. If a population of such cells were independent, the performance would simply scale with the number of cells. Correlations between the cells, however, would limit this increase in performance (Shadlen et al. 1996). It seems likely that the true decision process uses more sophisticated methods than simple summation. Exploration of the simplest pooling and selection mechanisms, however, provides insight into the representation of speed in MT, as well as the influence of adaptation on this representation.
Adaptation reduces perceived speed
In the psychophysical tasks that studied changes in the percept of speed with adaptation (AdaptNone and AdaptOne), subjects compared the nonadapted reference stimulus to a range of nonadapted and adapted test stimuli. At the level of an MT cell, this means that the relevant ingredients for the decision process are the response to the nonadapted reference stimulus and the response to one of the test stimuli.
We used a simplified version of the vector average labeled line model to relate neural activity to speed perception (Churchland and Lisberger 2001; Priebe and Lisberger 2004). In the vector average model, each cell votes for its preferred speed with a weight proportional to its firing rate, and the speed percept is decoded as the weighted average of these votes. The precise nature of the weighting and distribution of the labels (preferred speeds) have a strong influence on the quantitative outcome of this model. In fact, in previous work (Churchland and Lisberger 2001; Priebe and Lisberger 2004) this model has not been used to decode veridical speed information. The model can, however, successfully account for ordinal effects of stimulus changes on speed perception. For example, entering MT firing rates recorded for high spatial frequency gratings into the model leads to a lower decoded speed than entering the firing rates recorded for low spatial frequency gratings. This ordinal relationship (high spatial frequency looks slower) also holds perceptually (Priebe and Lisberger 2004). Typically, psychophysical experiments provide direct access only to ordinal relationships. In our behavioral experiments, for instance, subjects were asked to judge which of two stimuli was faster. Hence, to link the neural data to the behavioral data we only had to specify how neural firing is related to the concept of "faster."
Our simplification of the vector average model therefore used only ordinal labels. If the preferred speed of the cell was higher than the reference speed, the cell voted for faster. We refer to these cells as high-ps cells. Conversely, if the preferred speed of the cell was lower than the reference speed, it voted for slower (low-ps cells). This direct mapping of the neural activity to the two alternative forced choice (2AFC) task also allowed us to use the ROC analysis to estimate the weight of the vote: the larger the area under the ROC curve, the stronger the vote. In other words, the area under the ROC curve mapped directly onto an estimate of the fraction of trials in which an ideal observer would decide that the test stimulus was faster (or slower).
To show this decoding process, first consider the unadapted tuning curve of the high-ps cell in Fig. 4B. The ROC analysis showed that, in 80% of trials, a stimulus moving at 28.8°/s evoked a larger response than a stimulus moving at 16°/s. The preferred speed of the cell was higher than the reference speed; hence we interpreted the increase in firing rate as a vote for an increase in speed. Using this code, we could therefore translate the statement: in 80% of trials, the stimulus moving at 28.8°/s evoked a higher firing rate than the stimulus moving at 16°/s into the statement: in 80% of trials the ideal observer concluded that the 28.8°/s test stimulus moved faster than the 16°/s reference stimulus.
We applied the same analysis and the same code to all test speeds to generate the nonadapted neurometric curve (solid line) in Fig. 8A. The same ROC analysis and the same code were then used for comparisons of the adapted responses to the test stimuli and the nonadapted responses to the reference stimulus. This resulted in the adapted neurometric curve (dashed line). For this cell, the comparison of a 28.8°/s test stimulus after adaptation with the 16°/s reference stimulus resulted in a "test faster" response in only 50% of trials. In other words, the perceived speed of the test stimulus was reduced after adaptation. This reduction in perceived speed was found for all test speeds presented to this cell, as witnessed by the rightward shift in the neurometric curve. This rightward shift matched the rightward shifts seen in the psychometric curves of the AdaptOne condition in Fig. 2, A and B, and it can be compared with the change in the tuning curve of this cell shown in Fig. 4B.
|
Figure 8B shows the neurometric curves based on this linking assumption for the cell in Fig. 4D. Note how the psychometric curves are inverted compared with the firing rate tuning curves of Fig. 4D. For this cell, adaptation shifts the neurometric curve to the left, suggesting that, for the ideal observer relying on this cell, adaptation should lead to an increase in perceived speed.
As the neurometric effects clearly vary across cell types, it is important to quantify the behavior of the population. To average over cells, we expressed all test speeds as a percentage of the adaptation/reference speed. We determined the neurometric curves per cell (as shown for 2 cells in Fig. 8) and averaged those over all cells. Figure 9A shows the average neurometric curve for the high-ps neurons in our sample (n = 52). The average neurometric curve undergoes a rightward shift with adaptation that corresponds to 29% of the adaptation speed. A two-way ANOVA showed that the effect of adaptation was significant (P < 0.01). An ROC observer relying on this subpopulation should report test stimuli to be slower after adaptation. Figure 9B shows the same analysis applied to the low-ps cells in our sample (n = 17). Here adaptation also shifts the neurometric curve, but now 22% to the left, which suggests that the ROC observer relying on this subpopulation should report test stimuli to be faster after adaptation. This effect was also statistically significant (P < 0.01).
|
We also performed the analysis shown in Fig. 9 to the subset of significantly adapting cells (n = 27). The left and rightward shifts corresponding to those in Fig. 9, A and B, were larger in this subset (74%, n = 17; 29%, n = 10, respectively). Averaging over all cells that showed significant adaptation effects after adaptation at 16°/s (i.e., the equivalent of Fig. 9C), resulted in a PSE shift of 19%.
Although the speed tuning properties of our population were qualitatively similar to those reported in the literature, it is of interest to know whether this reduction in perceived speed is robust to changes in the composition of the population. In the recorded population, 65% of cells were high-ps cells. By resampling cells (with replacement) independently from the low-ps and high-ps subpopulations, we could simulate populations with different fractions of high-ps cells. In bootstrap fashion, we performed the same analysis used to create Fig. 9C on 1,000 simulated populations of 38 cells with varying compositions. This allowed us to determine the relationship between the shift in PSE (i.e., the change in perceived speed after adaptation) and the high-ps fraction. Over the range of high-ps cell fractions from 10 to 90%, the change in PSE varied almost linearly from –10 to +30%. These numerical values should be interpreted with caution because they depend on the size of the adaptation effects we found in our sample of 38 cells. Qualitatively, however, the result of this analysis shows that a decrease in perceived speed with adaptation occurred as long as the fraction of high-ps cells was >40%.
Adaptation improves discriminability
To determine whether two stimuli are different, an ROC observer need only determine whether the neural responses are different. This avoids the difficulty with interpreting a difference in rates in terms of faster or slower. In other words, modeling of discriminability can be accomplished without assuming a code. We performed ROC analysis to calculate the discriminability for each combination of two speeds in both the adapted and nonadapted state.
Figure 10 shows a scatter plot for all cells of the discriminability before and after adaptation, averaged over all speed comparisons. Most (41/69: 59%) data points fall above the diagonal line, showing that adaptation typically led to an improved discriminability for speeds in the range (±80%) around the adaptation speed. The median improvement over all cells for all possible speed comparisons was only 0.4% (P = 0.12, signed-rank test). The subset of 27 cells that individually showed a significant change in firing rate with adaptation exhibited an average improvement in discriminability of 1.7% per cell (P < 0.01, signed-rank test).
|
|
DISCUSSION |
|---|
|
TOPABSTRACTINTRODUCTIONMETHODSRESULTSDISCUSSIONGRANTSACKNOWLEDGMENTSREFERENCES |
|---|
