Stimulus repetition reduces neural response in cortical areas. Such adaptation is used in functional magnetic resonance imaging to infer the selectivity of neuronal populations; however, the mechanisms of adaptation remain elusive, especially in higher areas. We measured adaptation of spiking activity and local field potentials (LFPs) in macaque inferior temporal (IT) cortex for parameterized shapes by comparing tuning for test stimuli following a brief adaptation with predictions derived from different models of adaptation. Adaptation was similar during passive fixation or an attention-demanding task. We found consistent adaptation of spiking activity and LFP power in high- (gamma) but not low-frequency bands when repeating shapes. Contrary to sharpening models, repetition did not affect shape selectivity. The degree of similarity between adapter and test shapes was a stronger determinant of adaptation than was the response to the adapter. Adaptation still occurred when adapter and test stimuli did not spatially overlap, but adaptation was stronger for same, compared with different, adapters and test stimulus positions. These adaptation effects were similar for spiking and for gamma activity. In conclusion, adaptation of IT spiking activity and LFPs in IT is strongly dependent on feature similarities in the adapter and test stimuli, in agreement with input, but not firing-rate fatigue models.
The responses of neurons in macaque inferior temporal (IT), a visual cortical area coding for object properties (Gross et al. 1969; Dean 1976; Logothetis and Sheinberg 1996; Tanaka 1996; Afraz et al. 2006), decrease with the repetition of a stimulus (Gross et al. 1967, 1969; Baylis and Rolls 1987; Miller et al. 1991a, 1991b; Riches et al. 1991; Sobotka and Ringo 1993; Vogels et al. 1995; Sawamura et al. 2006; McMahon and Olson 2007; Verhoef et al. 2008; Liu et al. 2009). This “repetition suppression” (Desimone 1996) or “adaptation” effect (Ringo 1996) has recently aroused considerable interest because of the use of functional magnetic resonance imaging (fMRI) adaptation to estimate the stimulus selectivities of neuronal populations in humans (Grill-Spector et al. 2006). An understanding of the mechanism(s) of adaptation is thus essential for the interpretation of fMRI adaptation studies.
In the present study, we examine spiking and local field potential (LFP) responses in one sort of adaptation paradigm that has been widely used in fMRI adaptation studies: the rapid, zero-lag adaptation paradigm in which 2 stimuli, an adapter stimulus S1 and a test stimulus S2, are presented sequentially, separated by a brief interval (e.g., Kourtzi and Kanwisher 2000). Various models of adaptation have been suggested (Grill-Spector et al. 2006) and can be distinguished by their effect of S1 response strength, the relationship between S1 and S2 on the tuning for S2, and the latency and duration of adaptation effects. Figure 1 shows, for several adaptation models, the predicted effect that adaptation would have on stimulus tuning for various sequences of adapter and test stimuli. We computed the expected responses to 6 different test stimuli (see Materials and Methods for details), that is, the tuning function for these stimuli, following exposure to several adapter stimuli (i.e., following stimulus 1, following stimulus 2, and so on), using implementations of the adaptation models. Figure 1 shows the results of simulations for 4 different models of adaptation. Responses to the test stimuli (S2) 1–6 are plotted as a function of S1 by the colored curves, and the responses to S1 are shown by the black curve. Each colored curve indicates the tuning for the S2 following an S1 indicated by the same color on the abscissa. The first model, which we label the “firing-rate fatigue” model, assumes that the responses of a neuron to repeated stimuli decline due to firing rate–dependent “fatigue” of the neuron (Fig. 1A). A commonly made assumption about such a model is that the degree of repetition suppression is proportional to the firing rate with which the neuron responds to S1 (Andresen et al. 2009; Fig. 1E). In this sort of fatigue model, the responses to S2 would decrease in proportion to the response to the preceding S1. As shown in Figure 1I, a proportionally attenuating mechanism such as this produces only a weak decrease in the degree of selectivity when repeating a stimulus (degree of selectivity quantified by the depth of selectivity [DOS] index). The tuning for S2 is unaffected by whether S1 is a more (3 in Fig. 1) or less effective (6 in Fig. 1) stimulus for the neuron. The overall shape of the tuning function for S2 is therefore unaffected by adaptation.
Adaptation of a single neuron could also result from reduced efficacy of synapses or inputs with repetition, in agreement with the greater degree of adaptation observed when a stimulus is preceded by the same stimulus compared with a different one (Sawamura et al. 2006). Figure 1 shows the predictions of simple implementations of 2 such “input-fatigue” models. The first is a pure model (Fig. 1B,F,J) in which the degree of adaptation depends on the absolute parametric difference between S1 and S2 (as long as the neuron responds to S1): The smaller the parametric distance between S1 and S2, the stronger the adaptation effect. The second is a hybrid model (Fig. 1C,G,K) in which the response to S2 also depends on the response strength to S1 and is thus a mixture of the simple firing-rate fatigue (Fig. 1A,E,I) and the pure input-fatigue model (Fig. 1B,F,J). Repetition of the same stimulus produces either no effect on the degree of selectivity to S2 (pure input-fatigue model) or else a weak decrease (hybrid model), in a manner similar to the firing-rate fatigue model. However, unlike the firing-rate fatigue model, the input-fatigue models predict that the shape of the tuning function for S2 is affected by the value of S1: Tuning curves for S2 show a dip in this curve when S1 and S2 have the same parametric values. For each of the S2 stimuli, the largest adaptation occurs when S1 equals S2; this can be seen in Figure 1F by the consistently smaller responses to S2 for the curves with the color identical to the stimulus indicated on the abscissa. As a result, the degree of selectivity for S2 is strongly affected by whether S1 was a more or less effective stimulus such that the stronger the response is to S1, the weaker the selectivity for S2.
A third model, the “sharpening model” (Desimone 1996; Wiggs and Martin 1998; Fig. 1D,H,L) suggests that adaptation sharpens the tuning. This produces stronger adaptation when repeating less compared with more effective stimuli. In the simple version of this model simulated in Figure 1, equivalent tunings for S2 are predicted for both more and less effective adapter stimuli. This is why in Figure 1H, unlike the tuning curves for the other models, the tunings for S2 are indicated by a single stippled curve. More complex sharpening mechanisms are possible, for instance models in which only part of the tuning curve is sharpened (e.g., Fig. 5 of Grill-Spector et al. 2006), in which the degree of selectivity for S2 will depend on the effectiveness of S1. Nonetheless, a sharpening model will predict an increase in S2 selectivity when a stimulus is repeated, in contrast to other models. A fourth model, the “facilitation model” (e.g., James and Gauthier 2006), assumes that repetition decreases response latency or response duration, a hypothesis that can be tested by examining the time courses of adaptation effects.
As shown by the simulations in Figure 1, measurements of stimulus tuning before and after adaptation can be used to distinguish among the various adaptation models. The predicted effects on the tuning are summarized in Table 1. Previous studies have shown that the responses of IT neurons display orderly changes with regard to systematic variations of shape (Op de Beeck et al. 2001; Kayaert et al. 2005; De Baene et al. 2007), allowing measurements of at least portions of the IT tuning functions. In following these studies, the present report examines the effect of adaptation on the tuning of IT neurons, as measured using parameterized shapes. We employed a zero-lag, short stimulus and short interstimulus duration for our adaptation paradigm, with timing parameters similar to those used in rapid, event-related, short-duration fMR-A paradigms (e.g., Kourtzi and Kanwisher 2000).
|Model||Strongest adaptation?||Effect on selectivity in “same” trials?||Selectivity for test stimulus?|
|Firing-rate fatigue model||After effective adapter||No effect||Following effective = following ineffective|
|Input-fatigue model||Test stimulus = adapter||No effect||Following ineffective > following effective|
|Sharpening model||After ineffective adapter||Increase||Following effective = following ineffective|
|Model||Strongest adaptation?||Effect on selectivity in “same” trials?||Selectivity for test stimulus?|
|Firing-rate fatigue model||After effective adapter||No effect||Following effective = following ineffective|
|Input-fatigue model||Test stimulus = adapter||No effect||Following ineffective > following effective|
|Sharpening model||After ineffective adapter||Increase||Following effective = following ineffective|
Previous studies of adaptation in IT suggest that a mechanism based on firing-rate fatigue cannot fully explain adaptation effects observed in single IT neurons (Sawamura et al. 2006; Liu et al. 2009) and that the degree of adaptation does not decrease with stimulus effectiveness (McMahon and Olson 2007), in contrast to the sharpening model. However, previous studies have used only a single S2 stimulus (Sawamura et al. 2006), precluding a detailed assessment of S1 value on S2 selectivity, or adaptation effects proved rather weak and variable (McMahon and Olson 2007), perhaps due to the uncontrolled number of intervening stimuli. In other instances, only portions of the response were examined in a passive fixation task (Liu et al. 2009). The present study, for which preliminary data have been published earlier in abstract form (De Baene and Vogels 2007, 2008), shows that the tuning for S2 is affected by adaptation in a manner that agrees with input-fatigue models of adaptation at the single neuron level. Measurements were performed during 2 different tasks: a passive fixation task as in Sawamura et al. (2006) and Liu et al. (2009), or, which is novel, during a demanding luminance-change detection task. Because in the latter task attention to S1 and S2 is equalized, adaptation effects in this task cannot be ascribed to differences in attention to S1 and S2. Unlike previous studies, we also varied stimulus positions of S1 and S2 to determine how much of the adaptation observed in IT is inherited from earlier areas with small receptive fields (RFs) and whether position differences between S1 and S2 affect adaptation in the same manner as differences in the shapes of S1 and S2.
In addition to studying spiking activity, we also examined the effects of adaptation on a second measure of neural activity, LFPs. LFPs represent a population measure of neuronal, mainly synaptic, activity in the local cortical network (Mitzdorf 1987; Logothetis 2003). Measuring adaptation in LFPs and comparing adaptation effects for LFPs and spiking activity is important for at least 2 reasons. First, because adaptation will affect the response to the stimulus within an entire network of neurons and may alter network properties (e.g., decreasing interneuronal correlations; Gutnisky and Dragoi 2008), it is of interest to examine how adaptation affects the population response as captured by LFPs. Second, studies suggest that LFPs are better correlated with the blood oxygen level-dependent (BOLD) signal than is spiking activity (Logothetis et al. 2001; Viswanathan and Freeman 2007; Maier et al. 2008; Rauch et al. 2008).
Materials and Methods
Adaptation Models Simulations
In the simulations shown in Figure 1, the response to the adapter was defined as a Gaussian function: with and (for monotonic tuning [data not shown], and ). In the firing-rate fatigue model (Fig. 1E), the level of adaptation is proportional to the neuron's initial response. The response to the test stimulus i following a particular adapter x, yi(x), was modeled as (Andresen et al. 2009): with and . The response to the test stimulus i, following a particular adapter x, yi(x), as predicted by the input-fatigue model (Fig. 1F), was modeled as with and c = 0.50. The response to the test stimulus i, following a particular adapter x, yi(x), as predicted by the hybrid model (Fig. 1G), was modeled as with and . The response to the test stimulus i following a particular adapter x, yi(x), as predicted by the sharpening model (Fig. 1H), was modeled as with . Note that the predictions of the models do not depend on the tuning curves of the neurons for the adapters: Monotonic (data not shown) and nonmonotonic tuning curves lead to qualitatively equivalent predictions. We determined whether the effects of adaptation on S2 tuning were tolerant to changes in the parameter values over a wide range. We varied between 0.1 and 20 for all adaptation models. For the firing-rate fatigue model in particular, we varied the value of c, by dividing the by values between 1 and 20. For the input-fatigue model, we varied c between 0.1 and 1 and substituted by . These 2 latter variations were also examined for the hybrid model. For the sharpening model, the of the Gaussian function defining the tuning width for S2 was varied between 0.1 and 20. These parameter variations all produced qualitatively similar adaptation effects on the tuning.
Subjects and Recording
Data were collected from 2 rhesus monkeys (Macaca mulatta). All animal care and experimental protocols complied with national and European guidelines and were approved by the K.U. Leuven Ethical Committee for animal experiments.
Positioning of the recording chamber was guided by structural magnetic resonance (MRI) images of each animal. Estimations of the recording positions were obtained by visualization of copper sulfate–filled tubes, inserted into the grid at one or more recording positions, using MRI scans obtained at the beginning of the recording sessions and at several times thereafter. The depths of the recording positions were estimated using depth readings of white–gray matter transitions and of contacts with the base of the skull. The estimated recording locations extended from 11 to 15 mm anterior to the ear canal in monkey M and from 16 to 22 mm anterior in monkey G. The latter animal was a 7.5 kg male, whereas monkey M was a 4.5 kg female. The brain anatomy at any given A/P level differed between the 2 animals. After correcting for differences in anatomy, using Saleem and Logothetis (2007) as a reference, the A/P levels of the recording sites ranged from 15 to 19 mm anterior in monkey M, overlapping with those of monkey G (16–22 mm). The recording locations were at the A/P level of the anterior middle temporal sulcus (AMTS) and included the lower bank of the superior temporal sulcus (STS) and the IT cortex in and lateral to the AMTS (estimated recording positions ranged from 18 to 24 mm lateral to the midline). Because there were no significant differences in adaptation index (AI, see below) between recordings in the STS and the lateral IT convexity, we pooled results from the different recording positions.
LFPs and spikes were recorded simultaneously from the same microelectrode using a Plexon data acquisition system. The input impedance of the head stage was >1 GΩ. The grounded guide tube served as a reference. The signal was amplified and split into spiking activity (band passed between 250 Hz and 8 kHz) and LFPs (band passed between 0.7 and 170 Hz, sampled at 1 kHz). Single spikes were isolated online using the Plexon software. Timings of the single units were stored, together with stimulus and behavioral events, in computer files for later offline analysis. Eye position was measured online using an ISCAN infrared eye tracking system (Iscan, Inc., Woburn, MA; sampling rate 120 Hz).
We generated 4 sets of shaded 3D stimuli (Fig. 2A) using “3ds Max 7” (Autodesk, Montreal, Canada). First, we designed 2 dissimilar, 3D shapes with the same number of vertices per set. Next, we applied the morphing algorithm implemented in the 3ds Max 7 software, resulting in a set of 6 shapes that altered their appearance smoothly from 1 of the 2 original shapes (the base) to the other (the target) by translating the position of the vertices from their configuration in the base shape to the configuration in the target shape in 5 equal steps. The shape heights ranged from 5.4° to 7.2° and the area occupied by the shapes within a given set was made virtually equal (to within a maximum difference of 1% of the total area). The stimuli of each set were gamma corrected to account for the monitor gamma. They were of equal mean luminance and were presented on a uniform gray background (17 cd/m2). Scrambled versions of each image were made by dividing the original image into blocks of equal size (16 × 16 pixels) and repositioning them in a random way. This scrambling procedure was repeated 5 times per stimulus, resulting in 120 scrambled images.
Monkeys had to keep their gaze within 0.8° (monkey G) or 1.0° (monkey M) of a red fixation target (size: 0.18°). A fluid reward was given when the animals maintained fixation throughout a trial. Responsive neurons were sought by using a search procedure in which all 6 stimuli of every shape set (N = 6 × 4 = 24 conditions) were presented. After a fixation period of 500 ms, a stimulus was presented for 300 ms. A trial ended after another 300 ms of fixation. The shapes were presented in pseudorandom order with the constraint that any 2 subsequent shapes were from a different shape set. The stimuli were presented with their centers of mass at the fixation point.
In the main adaptation procedure (Fig. 2B), after a stable fixation of 500 ms, the stimuli were presented foveally for 300 ms with an interstimulus interval (ISI) of 300 ms. After another 300 ms of stable fixation, the trial ended. The different stimulus sequences were shown in pseudorandom order so that a stimulus presented in trial n could not reoccur in trial n + 1. In order to unadapt the neuron, we presented 2 unadaptation trials between 2 of these so-called adaptation trials. In a unadaptation trial, 2 randomly chosen scrambled versions of the images were presented. To assure that a minimum of 4 scrambled images were presented between 2 adaptation trials, an unadaptation trial was repeated (with different scrambled images) whenever this trial was aborted. The main adaptation procedure always started with a unadaptation trial (Fig. 2B). Each stimulus combination was, on average, presented 8.11 times (minimum = 3, maximum = 14) per neuron/site.
In the position procedure (Fig. 3), test stimuli were presented at an eccentricity of 4° above or below the fixation point. In this fully crossed design, any 2 selected shapes could serve as adapter and test stimuli and could be presented at either position as the adapter or the test stimulus, resulting in 16 (2 shapes × adapter vs. test stimulus × 2 adapter positions × 2 test stimulus positions) conditions. The timing was identical to that of the main adaptation procedure. Randomly chosen scrambled images were presented above and below the fixation point in unadaptation trials. Each condition was, on average, presented 9.55 times (minimum = 4, and maximum = 13) per neuron/site.
The monkeys were trained to detect a 150-ms dimming of either the adapter or test stimulus that began randomly between 0 and 150 ms after stimulus onset. During the first 83 ms of the dimming period, we linearly decreased the luminance of every pixel in the object. Difficulty was manipulated by modulating the amplitude of this decrease (0.86, 0.90, or 0.94 of the original intensity). The luminance then increased linearly over the next 67 ms back to the original level. All stimuli, including scrambled images could be subjected to the dimming. The dimming occurred in about 25% and 13% of the adaptation and unadaptation sequences, respectively, and was equally probable for adapter and test stimuli. Whenever a dimming occurred, a saccade had to be made toward a black target point within 500 ms after the start of the dimming. This target (0.56° in diameter) appeared together with the fixation point, at 9° eccentricity to the right of the fixation spot. When no dimming occurred, the monkey had to maintain fixation throughout the trial. Correct saccades in dimming trials, and fixation maintenance in the other trials, were rewarded. Maintenance of fixation or saccades before dimming onset in dimming trials were scored as errors. Saccades to the target point in nondimming trials were scored as false alarms. Adaptation sequences were separated by 2 nonaborted unadaptation trials with scrambled images and were not preceded by a unadaptation sequence in which a dimming occurred. Each stimulus combination was, on average, presented 8.30 times (minimum = 4 and maximum = 14) per neuron/site.
Eye Movement Analysis
The within-trial standard deviation (SD) of the eye position was measured within a 300-ms interval starting at the onset of the test stimulus presentation and averaged across all trials having the same parametric distance between adapter and test stimuli. The mean standard deviation averaged across parametric distances, measured in the main procedure, equaled 0.07° and 0.08° for the x direction in monkeys G and M, respectively, and 0.08° for both monkeys in the y direction. There was no significant difference between the SDs of the “same” conditions and the conditions with maximal parametric differences between adapter and test stimuli (mean SD differences <0.002° and 0.001° in the x and y directions, respectively). Similar nonsignificant results were obtained in the position procedure. These results show that any differences in adaptation effects were not due to differences in eye movements.
Unless otherwise stated, net neuronal responses were used in all analyses and were computed as the mean firing rate, in the response window extending from 50 to 350 ms relative to stimulus onset, after subtraction of the mean firing rate in the baseline window 300–0 ms before onset of the adapter. For each cell, the responsiveness to each stimulus presented as an adapter was tested with a Wilcoxon test (P < 0.05). For the analyses of the position procedure, we defined the response to a particular adapter at a particular position as the average net response across all conditions in which that particular shape, presented at that particular position, served as an adapter.
For the LFPs, we applied a digital 50-Hz notch filter offline (fourth-order Butterworth finite impulse response filter; Fieldtrip Toolbox, F.C. Donders Centre for Cognitive Neuroimaging, Nijmegen, The Netherlands; http://www.ru.nl/fcdonders/fieldtrip). Trials in which the signal exceeded the 5–95% window of the total input range were excluded from the analyses. The visually evoked potentials (VEPs) were computed by stimulus-locked averaging of the LFPs (Fig. 4A,B). The VEPs were analyzed only at the population level. More elaborate analyses of the LFPs were based on a time–frequency wavelet decomposition of the signal between 15 and 100 Hz for a spectral analysis of the LFPs. By convolving single-trial data using complex Morlet wavelets (Tallon-Baudry et al. 1997) and taking the square of the convolution between wavelet and signal, we obtained the time-varying power of the signal for every frequency (Fig. 4C,D). Per frequency and site, we took the median power across trials. The complex Morlet wavelets had a constant center frequency–spectral bandwidth ratio (f0/σf) of 7, with f0 ranging from 15 to 100 Hz in steps of 1 Hz. We conservatively excluded frequencies below 15 Hz from our analyses to avoid overlap of responses to test and adapter stimuli for the long-duration wavelets at these low frequencies. For Morlet wavelets, temporal resolution decreases with decreasing center frequency, precluding valid power estimates at low frequencies for the stimulus durations and ISI used in the present study.
The LFP power response was computed by taking the average energy at each frequency in a 50- to 350-ms response window relative to stimulus onset and averaging across the frequencies of the frequency band of interest. We employed 2 frequency bands: 15–30 and 60–100 Hz. Intermediate frequencies (31–60 Hz) were not included in the analyses, because only weak differences, compared with the baseline power, were observed for these frequencies throughout the duration of the stimulus sequences. Frequencies above 100 Hz were not examined due to the possible contribution of low-frequency components of spikes at these higher frequencies.
where Rai = mean response to the ith adapter; Rti = mean response to ith test stimulus, with i ranging from 1 (stimulus A) to 6 (stimulus F) for the main adaptation procedure. For the position adaptation procedure, Rai = mean response to adapter on ith position and Rti = mean response to test stimulus on the ith position.
The degree of adaptation for the VEPs was computed by averaging the VEPs across all “same” trials, that is, trials in which the same stimulus was repeated, across all sites per monkey. Next, we measured the peak-to-peak amplitudes for 2 components (see Fig. 4) by computing the amplitude difference between peak 1 and peak 2 (component 1) and between peak 2 and peak 3 (component 2) for the adapter and between peak 4 and peak 5 (component 1) and between peak 5 and peak 6 (component 2) for the test stimulus. The peaks were defined using visual examination of the population VEPs. The adaptation indices for the VEP components were computed as follows:
The DOS (Rainer and Miller 2000) was used as a metric of shape selectivity:
Besides computing the DOS index as a metric of shape selectivity, we used the best–worst B–W index defined as the difference between the maximum and minimum raw responses for the 6 shapes divided by the maximum response. Additionally, we computed d′ values, defined as the difference between the maximum and minimum raw responses for the 6 shapes divided by the square root of the mean of the variances in the response strength to these stimuli (Afraz et al. 2006; De Baene et al. 2008). These d′ values take into account the difference in mean spike counts as well as the variance of the response over trials.
We subdivided the cells/sites into 3 groups based on the responses to the different adapters. The first group of cells/sites showed maximum responses to one of the extremes of the shape set when these were presented as adapters (shape A or F, Fig. 2). For the cells/sites with shape A as the most effective adapter, both adapter and test stimuli were ranked in ascending order (i.e., ABCDEF). If shape F was the most effective adapter, a descending ranking was used (i.e., FEDCBA), for both the adapter and test stimuli. The second group of cells/sites showed a maximum adapter response to shape B or E, whereas the third group of cells/sites had a maximum adapter response for shape C or D. For the cells/sites with shape B or C as the most effective adapter, both adapters and test stimuli were ranked in ascending order (i.e., ABCDEF). If the maximum adapter response was found for shape E or D, a descending ranking was used (i.e., FEDCBA).
To assess the influence of response strength and S1–S2 parametric distance on the responses to S2, we employed a multiple regression approach. We predicted the response to the test stimuli in the 36 conditions using 3 independent variables:
R(S2(i)) = b1*R(Shape(j))+b2*R(S1(i))+b3*(distance S1–S2). The first variable, R(Shape(j)), corresponded to the response to each shape, presented as S1 (j = 1…6) and is the response to the shape of the test stimulus measured when that particular shape was presented as an adapter (averaged across the 6 conditions in which the adapter had the same shape). This variable corresponds to the shape tuning of the neuron. The second variable, R(S1(i)), was the response strength to the adapter S1 that preceded the test stimulus in that particular sequence (i = 1…36). An effect associated with this variable indicates an influence of the strength of the S1 response on the response to S2. The third variable was the parametric distance between the adapter and the test stimulus in each of the 36 conditions. An effect here indicates that the distance between S1 and S2 influences the response to S2. We computed squared semipartial correlations (Cohen et al. 2003) between the responses to test stimuli and each of the 3 variables. The squared semipartial correlation estimates the portion of the variance in the dependent variable (the response to S2 in the 36 sequences) explained by the independent variable. Semipartial correlations were computed for those neurons and LFP sites in which the multiple regression analysis provided a statistically significant prediction (P < 0.05). Raw responses were used in the regression analysis.
We employed 4 sets of shapes consisting of 6 shapes each (Fig. 2A). In a preliminary test, the shape selectivity of an isolated IT neuron was measured using all 6 of the stimuli in each of the 4 shape sets (24 stimuli). Based on visual inspection of the online peristimulus time histograms (PSTHs) of the responses of the neuron to these stimuli, we selected a shape set for which the neuron appeared to show shape selectivity. In the main adaptation test (Fig. 2B), each of the 6 shapes (adapters) could be followed by each of the 6 shapes (test stimuli) of the selected shape set. Thus, we presented all 36 possible stimulus sequence combinations (sequences A–A, A–B, A–C, …, F–E, and F–F). We measured spiking and LFP activity simultaneously using the same microelectrode. Spiking and LFP activities for each of the 36 shape sequences are shown for a single neuron and its corresponding LFP site in Figure 5.
We recorded single-unit activity and LFPs from 80 sites (40 in each of the 2 monkeys) in the main adaptation test. All 80 neurons were responsive to at least one stimulus presented as an adapter (Wilcoxon test, P < 0.05). The spiking activity decreased when the stimulus was repeated: The average response to the adapter was significantly larger than the average response to the test stimulus when both stimuli had the same shape (“same” trials; Wilcoxon test: both monkeys, P < 0.001; in each monkey P < 0.001). The median AI (see Materials and Methods) for the spiking activity was 37% and was similar for the 2 monkeys (monkey G: 40% and M: 34%; Fig. 6A).
After removing the outlier trials (see Materials and Methods) in the LFP data, 3 LFP sites were excluded from further analyses because of insufficient data in some shape sequence conditions. Figure 4A,B shows the mean VEP averaged across sites for each monkey. The peak-to-peak amplitude of 2 components (see Materials and Methods) of the population VEPs (Fig. 4A,B) decreased substantially with repetition in monkey M (AI component 1 = 70% and AI component 2 = 44%). However, in monkey G, the amplitude decreased only slightly (AI = 7%) for component 1 and even increased with stimulus repetition (AI = −19%) for component 2. Spectral analysis of the LFPs (Fig. 4C,D) showed consistent suppression of power at high frequencies (>60 Hz) in both animals, but as for the VEPs, differences between the 2 animals were observed for frequencies below 30 Hz. To quantify this observation, for each LFP site, we computed the power in each trial in 2 frequency bands: a low band extending from 15 to 30 Hz and a high one extending from 61 to 100 Hz (gamma). In monkey G, power in the low-frequency band increased significantly with stimulus repetition (median AI = −14%; Wilcoxon test, P < 0.001), whereas the power significantly decreased with repetition in monkey M (median AI = 24%; Wilcoxon test, P < 0.001). Negative AIs for the low-frequency band were observed at all guide tube positions and in the STS and the lateral convexity of monkey G, whereas positive AIs were observed at all guide tube positions and in the lateral convexity and STS of the other animal. The opposite effects present in the 2 animals resulted in an across-monkey median AI of 10% for the low-frequency band (Fig. 6B). However, the power in the higher, gamma band decreased substantially with repetition (Fig. 6C). The average gamma response to the adapter was significantly larger than the response to the test stimulus in the “same” trials (Wilcoxon test across monkeys, P < 0.001; in each monkey, P < 0.001). The median gamma power AI was 22% across monkeys, and adaptation was present in both subjects (median AI = 27 and 14% for monkey G and M, respectively). Using raw, instead of net, responses to calculate the AI for the spiking activity resulted in a median AI of 22%, which was very similar to the value obtained for the gamma power. Thus, spiking activity and gamma power demonstrated similar levels of adaptation.
The same stimulus sequences were also presented when the animals performed a “dimming” task (see Materials and Methods) in which stimulus attention was controlled. In this task, the monkeys were required to detect a subtle global luminance change in the stimulus. The luminance change could occur for either the test or the adapter stimulus, and was introduced in a small proportion of the trials. To detect the random luminance changes, the animal was thus required to attend to both test and adapter stimuli. The animals were trained extensively in this task for several months. The overall performance of both monkeys was well below ceiling rate during the recordings (average detection rate 73%) and decreased with decreasing luminance steps (detection rates of 94%, 85%, and 40%, averaged across monkeys for the 3 interleaved luminance steps, respectively). False-alarm rates were very low in both monkeys (<1%). In both monkeys, the detection rate of the dimming of the test stimulus did not depend on the parametric distance between the test and the preceding adapter stimulus (Fig. 7). Thus, the level of stimulus attention is expected to be similar for the “same” and “different” trials.
For the dimming task, we recorded spiking activity and LFPs from 29 responsive neurons (Wilcoxon test, P < 0.05) and sites (16 and 13 in monkeys G and M, respectively). We will present only the analyses of the responses in trials in which no dimming occurred that were randomly interleaved with the dimming trials. The median AI for the spiking activity was 31% and the response to the adapter was significantly larger than the response to the test stimulus in these “same” trials (Wilcoxon test in each monkey: P < 0.001). The AI for the spikes did not differ significantly between the dimming task and the main adaptation procedure in which the animals passively fixated (Mann Whitney U test; P = 0.40). The AI for the power in the low (median AI: 6%) and gamma bands (20%) in the dimming task were also similar to those during the fixation task (Fig. 6D–F). As in the latter task, there was a strong discrepancy in repetition effects between the 2 animals for the low-, but not the high-, frequency bands.
Effect of Repetition on Neuronal Selectivity for Spiking and LFP Responses
The parametric variation of shape allowed us to construct tuning curves and determine the effect of repetition on the shape tuning for spiking and LFP responses. In the first analysis, we compared the tuning for adapter and test stimuli when both stimuli had the same shape. This allowed us to test the predictions of the different models: The sharpening model predicts an increase in selectivity with repetition, whereas the other models predict either no change or a decrease in selectivity.
The response to a particular adapter was defined as the average response across all “different” conditions (i.e., different test and adapter stimuli in the sequence) in which the adapter shape was shown. For the test stimuli, tuning curves were based on the average responses in the “same” conditions (i.e., identical test and adapter stimuli in the sequence). Because the tuning curves for the adapter and for the test stimuli were thus based on independent trials, any similarity between these tuning curves could not be induced by common artifacts, a consideration important for LFPs. Only cells showing significant differences in spiking activity across the 6 shapes of the selected set (i.e., selective cells; analysis of variance [ANOVA]; P < 0.05) were included in the analyses (N = 71).
We subdivided the cells/sites into 3 groups based on the parametric values of their preferred shapes (see Materials and Methods), that is, those preferring shapes A or F, B or E, and C or D. The selectivity of spiking activity for the test stimuli in the “same” trials was significant in each group (repeated-measures ANOVA; P < 0.001 in each group), indicating tuning for the test stimulus, and the tuning curves were similar for adapter and test stimuli (Fig. 8A). The gamma band activity also showed reliable tuning for the test stimuli for group 1 (Fig. 8C; repeated measures ANOVA: group 1: P < 0.01; groups 2 and 3: n.s), demonstrating that IT gamma band activity can show reliable shape tuning, at least when the distance between effective and ineffective stimuli is sufficiently large (as in group 1). Notably, for gamma activity, as for spike activity, the most effective shape was the same for adapter and test stimulus presentations. (Low-frequency power data were not analyzed because these did not show consistent adaptation; see above).
To compare the degree of shape selectivity for the adapter and test stimuli, we computed per cell/site the DOS (see Materials and Methods) for both sets of stimuli. Again, we used independent trials to compute the selectivity indices for the adapter and test stimuli: The DOS index for the test stimuli was computed on the responses in “same” trials whereas the DOS index for the adapters was computed for a different condition which was randomly selected per shape. There was a significant correlation between the DOS indices for the adapter and test stimuli, for both spiking (r = 0.80, P < 0.001; N = 71; Fig. 8B) and gamma activity (r = 0.42, P < 0.001; N = 77; Fig. 8D). The average differences between the DOS for the adapters and the test stimuli were small and not significantly different from 0 for both the spiking (median difference adapter − test stimuli = −0.01, P = 0.60, Wilcoxon test) and gamma band activities (median difference = 0.01, P = 0.25, Wilcoxon test). Making this sort of comparison of the DOS for the lowest and highest quartiles of cells/sites using subsets based on their maximum net response strengths gave very similar results, for both spiking (respectively, median difference = −0.02, P = 0.62 and −0.04, P = 0.30, Wilcoxon test) and gamma band activities (respectively, median difference = 0.03, P = 0.35 and −0.02, P = 0.40, Wilcoxon test). Controlling for stimulus attention in the dimming task also resulted in average differences in DOS values for adapter and test stimuli in “same” trials that were also small and not significant, for both spiking (median difference = 0.00, P = 0.97, Wilcoxon test; N = 26) and gamma band activity (median difference = 0.00, P = 0.69, Wilcoxon test; N = 29).
These results were not specific to the selectivity index used: Very similar results were also found using different measures of selectivity, the B–W index, computed as the difference between maximum and minimum activity divided by the maximum, as well as d′ values, computed as the difference between maximum and minimum activity divided by the square root of the mean of the variances. As for the DOS indices, the average differences between the B–W indices for the adapter and the test stimuli were also small and not significantly different from 0 for both the spiking (median difference adapter − test stimuli = −0.03, P = 0.29, Wilcoxon test) and gamma band activity (median difference = 0.03, P = 0.09, Wilcoxon test). This was also true for the d′ values for spiking (median difference adapter − test stimuli = 0.24, P = 0.07, Wilcoxon test) and gamma band activity (median difference = 0.06, P = 0.43, Wilcoxon test).
Thus, repetition of a shape did not increase the selectivity and as shown by the tuning curves in Figure 8, nor did the adaptation effect in the “same” trials increase with decreasing stimulus effectiveness. Thus, the adaptation effects observed in same trials argue against a sharpening model but, rather, favor other models of adaptation. This holds true for both spiking and gamma band activity. The latter models can be further distinguished by examining the tuning for the test stimuli as a function of the adapter stimulus.
Effect of Adapter Value on Tuning for Test Stimulus for Spiking and LFP Responses
Our design allowed us to examine tuning for the test stimulus as a function of the adapter value. For both spiking activity and gamma power, the degree of adaptation tended to be greatest when adapter and test stimulus were identical, and this observation held even for the less effective shapes. In fact, the tuning curves for the test stimuli, sorted according to adapter stimulus, tended to show a dip when test and adapter shape were identical (Fig. 9), similar to the predicted tuning changes for the input-fatigue models (Fig. 1F,G). For Figure 9, we defined the response to a particular adapter as the average response across the 6 conditions in which that particular shape served as an adapter. Note that this definition is different from that used in the previous analysis, in which the adapter responses in the “same” trials were not included. This explains the slight difference in number of recording sites for the same-preferred-shape group in the gamma band data shown in Figures 8 and 9.
The effect of adapter value on test stimulus selectivity can be demonstrated most clearly by considering 2 shapes for each neuron: the most effective shape presented as an adapter (“effective stimulus”) and the shape which parametrically differed from that shape the most (“ineffective stimulus,” e.g., shape F when the most effective shape was stimulus A, B, or C). These analyses were restricted to those selective cells responding significantly to both effective and ineffective stimuli presented as an adapter (Wilcoxon test; P < 0.05; only excitatory responses). The spiking activity (N = 71; Fig. 10A) showed significantly more adaptation when the effective stimulus was repeated than when this test stimulus followed the ineffective adapter (P < 0.001; Fisher least significant difference [LSD] test). For the ineffective test stimulus, however, the opposite was true: significantly more adaptation was found when this test stimulus followed the identical, but ineffective adapter, than when it followed the different, but more effective stimulus (P < 0.01; Fisher LSD test). The interaction between test and adapter stimulus value proved significant (ANOVA; P < 0.001; crossing of gray and dashed curves in Fig. 10A), which agrees with the input-fatigue models but not with the firing-rate fatigue model. Indeed, if the latter had been the case then the dashed and gray curves should run parallel and should not cross.
For the gamma power (N = 77; Fig. 10D), significantly more adaptation was also found when the ineffective shape was repeated compared with when this shape followed the most effective adapter (P < 0.001). More adaptation was also found when the most effective stimulus was repeated compared with when this stimulus followed the ineffective adapter (P < 0.001). As for the spiking activity, the interaction between test and adapter stimulus value was significant for the gamma power for the test stimulus (ANOVA; P < 0.001; crossing of gray and dashed curves in Fig. 10A,D). Note that very similar results were found when the above described analyses were executed for each animal individually: For both the spiking activity and the gamma power for the test stimulus, the interaction between test and adapter stimulus value was significant in each monkey (ANOVA; P < 0.001). In the dimming task, the interaction between test and adapter value was also significant for both the spiking activity (N = 22) and the gamma power (N = 29) of the test stimulus (ANOVA P < 0.01 for both spiking and gamma activity; Fig. 10B,E).
Examination of the degree of selectivity for the test stimuli as a function of the adapter stimulus value also agrees with an input-fatigue model. As shown in Figure 1, input-fatigue models predict that the degree of selectivity for the test stimuli will be smaller when the test stimuli are preceded by the effective stimulus, compared with the ineffective stimulus, for a given neuron. This was indeed the case for spiking (N = 71) and gamma power (N = 77) activity in the main task shown in Figure 11: The mean DOS, B–W and d′ indices for the test stimuli were significantly smaller when the test stimuli were preceded by the effective stimulus (mean DOS = 0.33 and 0.21; mean B-W index = 0.50 and 0.32; mean d′ values = 2.22 and 0.99 for spiking and gamma activity, respectively) compared with when these were preceded by the ineffective stimulus (mean DOS = 0.37 and 0.24; mean B–W index = 0.55 and 0.36; mean d′ values = 2.77 and 1.33 for spiking and gamma activity, respectively. Difference for spiking and gamma activities, respectively: P < 0.001 and P < 0.05 for DOS index; P < 0.001 and P < 0.01 for B–W index and P < 0.001 and P < 0.001 for d′, Wilcoxon test).
The effect of the adapter stimulus value on the degree of selectivity for the test stimuli did not differ significantly between the main adaptation procedure, requiring only passive fixation, and the dimming procedure. This was true for both the spiking activity (ANOVA; interaction between tasks and DOS indices following the effective and ineffective stimulus: F < 1; mean DOS in dimming procedure after effective vs. ineffective stimulus = 0.27 vs. 0.31) and for the gamma power (F < 1; mean DOS in dimming procedure after effective vs ineffective stimulus = 0.20 vs. 0.21).
Relative Contribution of Adapter–Test Distance versus Adapter Response Strength for Spiking and LFP Responses
To quantify the relative contributions of adapter–test stimulus distance and adapter response strength to modifications in the tuning to the test stimulus, we performed a multiple regression analysis in which we predicted the response to the test stimuli in the 36 conditions based on 3 variables. The first variable corresponded to the tuning for the adapter shape (6 values) and was computed as the response to the shape of the test stimulus measured when that particular shape was presented as adapter (averaged across the 6 conditions in which the adapter has the same shape). The second variable was the response strength to the adapter that preceded the test stimulus for each of the 36 conditions (36 values). The third variable was the parametric distance between adapter and test stimulus in each of the 36 conditions. We computed squared semipartial correlations (Cohen et al. 2003) between the test stimuli responses and each of the 3 variables. The pure input-fatigue model (Fig. 1B) predicts that the squared semipartial coefficients will be larger for the stimulus distance variable than for the response strength variable (0.04 vs. 0.00 for the simulations of Fig. 1), whereas the firing-rate fatigue model (Fig. 1A) predicts just the opposite (0.00 vs. 0.04). The predictions of hybrid firing rate–input-fatigue models (Fig. 1C) depend on the relative contributions of the 2 factors (response strength and input similarity) to the adaptation. Note that the predictions for the simulations based on monotonic tuning curves are qualitatively identical to those of the bell-shaped tuning curves of Figure 1.
Semipartial coefficients were computed for those neurons and LFP sites in which the multiple regression analysis provided a statistically significant prediction (P < 0.05). Across the 77 neurons that fulfilled this criterion, a mean of 64% of the total variance in spiking activity elicited by the test stimuli was accounted for by these 3 variables. The shape-tuning variable explained the bulk of the variance (54%). On average, the response strength variable explained a much lower proportion of the variance (3%) than did the adapter–test stimulus distance (6%), a difference that proved significant (Wilcoxon test; P < 0.01; Fig. 11C). The adapter–test stimulus distance variable explained little or no variance in only a minority of the neurons, whereas the response strength variable contributed to the variance (neurons close to the Y axis of Fig. 11C). Hence, only a minority of the neurons showed a pattern that conformed to the firing-rate model and for the majority of IT neurons, the stimulus distance was a stronger determinant of the adaptation than the strength of the response to the adapter.
For the gamma power, 38 sites fulfilled the criterion with a mean total explained variance of 41%. Only about half of this variance was due to the tuning variable (22%), which is less than that for the spiking activity, and fits the overall lower shape selectivity for LFPs compared with spike counts. Similar to the spiking activity results, stimulus distance explained 13% of the variance while the response strength explained only 4%, a highly significant difference (Wilcoxon test; P < 0.001; Fig. 11D). Similar results were obtained by analyzing all 77 LFP sites (showing significant difference [Wilcoxon test; P < 0.001] between the response strength [3% explained variance] and adapter/test stimulus distance [9%; total explained variance 25%] variables).
The relative contributions of response strength and adapter–test stimulus distance did not differ significantly between the main adaptation procedure with passive fixation, as described above, and the dimming procedure. This was true for both the spike rate (ANOVA; interaction between tasks and squared semipartial correlation coefficients for the 2 variables; P = 0.26) and gamma power (P = 0.81). The squared semipartial correlation coefficients for the response strength variable were very similar for these 2 tasks (t test; dimming task: 0.03 for spikes and gamma power). This was also true for adapter–test stimulus distance (squared semipartial coefficients 0.04 and 0.11 for spikes and gamma power, respectively, in the dimming task). These data suggest that the degree of similarity between adapter and test stimuli is a more powerful determinant of both spiking activity and gamma power adaptation than is the strength of the response to the adapter.
Adaptation for Nonoverlapping Adapter-Test Stimulus Positions
One possible explanation of the input-dependent adaptation is that the adaptation in IT is inherited from adaptation at earlier stages of the visual stream. To determine whether the adaptation effects observed in IT might originate in earlier visual areas having small RFs, we presented the adapter and test stimuli at different, nonoverlapping locations within the RF of an IT neuron. If adaptation is simply an inherited property from such earlier areas, no adaptation would be expected if both adapter and test stimuli do not fall within the same RFs at those earlier levels.
Two shapes, chosen on the basis of a preliminary test, were selected for this position procedure: the shape eliciting the best response and a shape that differed from the best shape as much as possible while still eliciting a clear response. We recorded single-unit activity and LFPs from 40 sites (20 in each monkey) using this position procedure. All single neurons tested were significantly responsive to at least one stimulus presented at one of the 2 positions. Only cells (N = 37) that were responsive to both shapes at both positions (Wilcoxon test; P < 0.05) were included in further analyses to ensure that both stimulus locations were within the RF of the neurons tested.
To examine the effect of adapter and test stimulus positions on adaptation, we restricted our analyses to those conditions in which the same shape had served as both adapter and test stimuli. The “best” position was defined as the position at which the most effective shape elicited the greatest response when presented as the adapter. The other position was designated “worst” position. We computed the AI index, averaged across best and worst positions and across both shapes, in conditions where the same shape was presented at different positions as an adapter and as a test stimulus. The median AI for these “different position” conditions was 21% for the spiking activity (monkeys G: 21% and M: 23%) and 14% for the gamma power (monkeys G: 8% and M: 20%). The response decrease after stimulus repetition was significant (Wilcoxon test, P < 0.001 for both spiking activity and gamma power), but smaller than in conditions where adapter and test stimuli were presented at the same position (AI: 36% and 16% for spiking and gamma activity, respectively). The robustness of the response decreases even when adapter and test stimuli have different positions suggests that the adaptation was, at least partially, generated at a higher-level visual area rather than being inherited from earlier stages with smaller RFs.
We also compared the responses with the shapes presented as test stimuli at the 2 positions as a function of the position of the adapter (best or worst position). To this end, we used a repeated-measures ANOVA with adapter position (best vs. worst), test position (best vs. worst) and shape (effective vs. less effective) as within-neuron or within-site factors and response to the test stimuli as the dependent variable. The interaction between adapter and test stimulus position was significant (P < 0.001) for both the spiking activity and the gamma power of responses to the test stimuli. The spiking and gamma activity (Fig. 10C,F) showed more adaptation when the shape was repeated at the best position compared with when this shape at the best position followed the shape at the worst position (although this proved significant only for the spiking activity: P < 0.05; gamma activity: P = 0.13; Fisher LSD test). However, for the test stimulus presented at the worst position, significantly more adaptation was found when this test stimulus followed the adapter at the worst compared with the best position (spiking and gamma activity: P < 0.05; Fisher LSD test).
These results demonstrate that the degree of adaptation depends upon the positions of both adapter and test shapes. This is similar to the manner in which adaptation depended upon the relationship between the shape features of the first and second stimuli in the main adaptation procedure. Thus, the strongest adaptations are obtained when the shapes or positions of the adapter and test stimuli match.
Time Course of Adaptation Effects
Figure 12 shows population PSTHs of the spiking and gamma activity in same shape trials of the main and position procedures for the fixation task and for the dimming task. All cells/sites we recorded from in the respective condition were included in the analyses. In agreement with Sawamura et al. (2006), suppressed spiking and gamma responses to the repeated stimulus was present throughout the duration of the stimulus in both tasks. Note that the duration of adaptation is shorter than the duration of the response to the stimulus, because at 350-ms poststimulus onset, the difference in response between adapter and test stimuli is no longer present although the responses are still above baseline. This implies that, at this late part of the response, the tuning is similar for the adapter and test stimuli.
An inspection of Figure 12 suggests that for both spiking and gamma band activity, suppression was statistically significant over longer durations in the fixation compared with the dimming task. However, this apparent difference in the duration of significant adaptation did not prove to be reliable when taking into account the lower number of neurons in the dimming task. This was tested as follows: For the spiking activity test, we randomly selected 29 out of the 80 cells of the main test 500 times and then compared adaptation durations of these samples of neurons to the adaptation duration of the 29 neurons tested in the dimming task. Adaptation duration was computed for the 2 tasks using the following procedure. First, we tested which 10-ms bins showed significant differences in mean response between adapter and test stimuli, using a Wilcoxon matched pairs test (P < 0.05; Fig. 12). Second, only where at least 2 out of 3 successive time bins were significant were the significant bins defined as showing a reliable difference. The offset of the adaptation was defined as the last reliable time bin. The computed adaptation duration of the dimming task was at percentile 33 of the distribution of the 100 adaptation durations of the main task, thus showing no significant difference in adaptation duration between the fixation and dimming task. For the gamma power, the procedure for determining the adaptation duration was the same as that for the spiking activity except that bin width was set to 1 ms and the requirement of a reliable, significant difference consisted of a significant difference of 20 of 25 successive bins. As for the spiking activity, the gamma power showed no significant difference in adaptation duration between the fixation and dimming task (P = 0.22).
To examine whether the relative contributions of adapter–test stimulus distance and response strength changed during the course of the response, semipartial correlations between these variables and the responses to S2 (see above) were computed for each of 4 successive time windows of 70 ms, starting 70 ms after stimulus offset. These regression analyses were done for each neuron and task separately. A repeated measure ANOVA with adapter–test stimulus distance versus response strength and time window as within-neuron factors and taking task as a between-neurons factor showed no significant interaction between the task factor and the other factors (F < 1 and P > 0.36, for both spiking activity and gamma power, respectively). As shown in Suppl. Figure 1, the explained variance for the adapter–test stimulus distance variable exceeded that for the response strength variable at each of 4 time windows for both the spiking activity and the gamma power (main effect of adapter–test stimulus distance versus response strength factor; P < 0.001). The interaction between time window and the adapter–test stimulus distance versus response strength factor was not significant (P > 0.09 and P > 0.08, for spiking activity and gamma power, respectively).
According to the facilitation model of adaptation, the processing of a repeated stimulus is faster and of shorter duration than a stimulus that differs from the preceding stimulus. This can be tested by comparing the time courses of the responses in same and different shape trials. Suppl. Figure 2 shows population PSTHs of the spiking and gamma activity averaged across the test stimuli A and F in same shape trials (A–A and F–F) and in different shape trials (A–F and F–A). Shorter test stimulus latencies were not observed for either spiking or gamma activity in trials in which adapter and test stimuli were identical compared with trials in which they differed. This suggests that, at least in the present paradigm (see Woloszyn and Sheinberg 2009 for a different result when the monkey performs a short-term memory task), repetition of a shape does not shorten the response latency compared with when 2 distinct shapes are presented in succession.
Although a comparison of latencies of spiking and gamma response is difficult, given the different filtering of the signals, the time courses for the various positions of adapter and test stimuli showed some distinctions between spiking and gamma activity. The peaks of both spiking and gamma activity showed position-dependent adaptation but, in the later part of the response (>200 ms), the degree of adaptation for the gamma power was equal whether test and adapter were at the same or different positions, but for the spiking activity, the adaptation effect remained position-dependent (Fig. 12).
Correlations of Gamma and Spiking Activity
Except for differences in their time courses (Fig. 12; position task), the gamma and spiking activity showed overall similar adaptation effects. This led us to examine whether our data show a general correlation between gamma activity and spiking activity. The average “signal correlation” between mean spiking and gamma activity for the 6 adapters, computed for each neuron and corresponding site measured with the same electrode, was significantly greater than 0 (mean r = 0.42; N = 77; P < 0.001; main task). These positive “signal correlations” between gamma and spiking activity are lower than those reported in studies of V1 (Belitski et al. 2008) and MT (Liu and Newsome 2006) but are larger than those reported by Kreiman et al. (2006) in IT for the 40- to 100-Hz frequency band.
There was a median correlation of 0.09 (min r = −0.16; max r = 0.51; N = 77; main task) between trial-to-trial variations in spike count and gamma power (computed across responses of each trial that were standardized for each stimulus condition [mean response for each stimulus = 0 and SD = 1]), which was significantly different from 0 (P < 0.001; Wilcoxon test). These “noise correlations” were significant (P < 0.05) for 34 of the 77 cell–site pairs.
It is unclear to what degree gamma activity between 60 and 100 Hz) reflects synaptic activity versus multiunit spiking activity in the neighborhood of the electrode (Viswanathan and Freeman 2007; Gieselmann and Thiele 2008; Ray et al. 2008). However, we can address whether the similar results for the gamma and spiking activity might be due to contamination of the gamma power by spiking activity of the simultaneously recorded single neuron. To address this possibility we reanalyzed the gamma power activity effects using only those sites for which the noise correlation of gamma and single-unit spiking activity was not significant. For these 43 sites, the median AI was 24% and was similar between monkeys (monkey G: 33% and M: 20%). As for the whole population of sites, the tuning curves for the adapters and for the test stimuli of this subset of sites were very similar, and the average differences in the DOS for the adapters and the test stimuli were small and not significantly different from 0 (median difference = 0.02, P = 0.23, Wilcoxon test). Additionally, the interaction between test and adapter shapes was significant for the gamma power of the test stimulus (ANOVA P < 0.001), just as it was without applying this selection criterion. These results suggest that the similarity between the adaptation effects for the gamma and for the spiking activity is not merely a result of contamination of the LFP signal by the spiking activity of the single cell we were recording from. Neither do noise correlations for spiking and gamma activity necessarily point to single-cell spike contamination. Indeed, the weak noise correlations may reflect cross-neuronal correlations in activity.
Models of adaptation differ in their predictions regarding the effect of adaptation on the tuning for the test stimulus. We examined the effects of adaptation on the shape tuning of single IT neurons and LFPs and the results conform to the various predictions of the input-fatigue models: no detectable narrowing of the tuning when the shape is repeated (Fig. 8), stronger adaptation when test and adapter stimulus are identical compared with when these differ irrespective of the strength of the response to the adapter (Fig. 10), and broader tuning for the test shapes following effective adapters compared with those that are less effective (Fig. 11A,B). Although most of the variance in the mean response to the test stimulus stemmed from the stimulus tuning of the neuron, relatively more variance was explained by the similarity between adapter and test stimuli than by the strength of the response to the adapter stimulus (Fig. 11C,D). These effects were qualitatively similar for gamma and spiking activity and were present during passive fixation as well as during an attentionally demanding dimming task.
The present findings agree with and extend those of Sawamura et al. (2006) and Liu et al. (2009) insofar as they show that the degree of adaptation is not merely determined by the firing rate elicited by the adapter stimulus but that the relationship between test and adapter stimulus is a strong determinant of the degree of adaptation. Our results extend the findings of Liu et al. (2009) by showing that this holds true when considering the entire response, a period equal to the duration of the stimulus, and is not restricted to some subset of the response showing maximal adaptation. Also, unlike in the Liu et al. (2009) study, we were able to examine different tuning patterns (“bell-shaped” and “monotonic” tuning curves). In addition, we show that changes in tuning following different adapters are similar during passive fixation and during a task in which attention to the adapter and test stimuli were equalized. This suggests that the effects of adaptation on shape tuning are not the result of reduced attention to repeated stimuli. Also, the importance of similarities between adapter and test stimuli with regard to the strength of the adaptation effect was not limited to their shapes, but was also related to their positions: adaptation was stronger when adapter and test stimuli occupied identical positions, irrespective of the strength of the response at the adapter position (Fig. 10). Nonetheless, adaptation was still elicited when adapter and test stimuli were presented at different positions, suggesting that the adaptation is not entirely inherited from earlier areas with small RFs.
There was a striking difference in adaptation for lower (<30 Hz) and higher (>60 Hz; gamma) frequency bands that was due mainly to interanimal differences in the degree of adaptation at the low frequencies (beta range). We do not know the source of this interanimal variation, but it turned out to be a fortunate circumstance, because it allowed us to dissociate gamma (and spiking) activity from the low-frequency LFP power. The 2 animals showed similar degrees of adaptation in the gamma power, and also in spiking activity, but this was not the case for the low frequencies. Unlike the high frequencies, LFP activity in the low-frequency bands did not correlate with the spiking activity, nor with the gamma activity, which agrees with results obtained in V1 (Belitski et al. 2008; Maier et al. 2008) and with MT data (Liu and Newsome 2006). These low-frequency components dominate the VEP and thus, as we observed, the adaptation of VEPs does not necessarily correlate with adaptation of spiking activity.
The adapter-dependent effects on shape tuning were also present in the gamma power of the LFP. It is important to verify that any effects observed in LFPs are comparable with those in the spiking activity, because this would extend the validity of the adaptation model to the population of neurons beyond the sample of neurons (which is biased for high-amplitude spiking activity) from which we recorded. Overall, the gamma power showed stronger stimulus-dependent adaptation effects than did the spiking activity. This can be explained by the fact that LFPs represent the activity of a large population of neurons, which can have heterogeneous stimulus preferences. Indeed, that population may contain neurons that are not activated by the effective or ineffective stimulus but do respond to the ineffective or effective stimulus, respectively. Such a heterogeneous population, taken as a whole, will show less adaptation when the effective adapter stimulus precedes the ineffective one or vice versa, compared with a single neuron, because the effective and ineffective stimuli will have driven sets of neurons that only partially overlap. Indeed, if a neuron does not respond to the adapter, adaptation will be absent (Sawamura et al. 2006). Note that in the present analysis of the spiking data, we required that a single neuron responded significantly to both effective and ineffective stimuli, thus ruling out spiking fatigue as a possible explanation of the greater adaptation for same compared with different adapter–test stimulus sequences.
The present electrophysiological study did not measure fMRI activity in our monkey subjects using the stimulus presentation paradigm. However, previous human fMRI studies using timing parameters similar to the ones used here have reported adaptation of the BOLD response for shaded shapes in area LOC (e.g., Kourtzi and Kanwisher 2000; Kourtzi et al. 2003; Murray and Wojciulik 2004), which might be the homologue of monkey IT (Denys et al. 2004; Sawamura et al. 2005). Furthermore, we have previously shown similar adaptation in a block design in monkey and human fMRI responses in IT and LOC, respectively (Sawamura et al. 2005). Thus, there is every reason to believe that consistent fMRI adaptation would be found in monkey IT for the stimuli and paradigm used in the present study. The consistent presence of adaptation in the gamma but not low-frequency band would therefore suggest that the fMRI response more faithfully reflects the high- rather than low-frequency LFP activity, as has already been suggested (Mukamel et al. 2005; Niessing et al. 2005; Viswanathan and Freeman 2007; but see Maier et al. 2008).
The repetition of a shape did not decrease the tuning width, a result inconsistent with the sharpening model of adaptation. One might object that the stimuli we employed could have been suboptimal for the neurons we recorded from, explaining why we also found strong adaptation for these suboptimal, “effective” shapes. However, even for suboptimal stimuli, the sharpening model predicts that the degree of adaptation would decrease with stimulus effectiveness (Fig. 1), resulting in a narrower tuning, which we did not observe. Furthermore, no sharpening of the tuning was present when only highly responsive neurons were considered. Thus our results do not support a sharpening mechanism. This extends the results of McMahon and Olson (2007), who mixed different time lags and numbers of intervening stimuli, and those of Woloszyn and Sheinberg (2009) in a recent study of short-term memory match effects in IT. The gamma power also showed strong adaptation for the effective stimuli. However, this cannot be used as evidence against sharpening, because the LFP response to our effective stimulus may reflect the activation of a mixture of neurons that prefer our effective but not “ineffective” stimuli and neurons that prefer the ineffective stimulus but also respond to the effective stimulus. Because the sharpening model predicts that neurons tuned to the ineffective stimuli but that are still (though less) responsive to the effective stimulus will show adaptation, one would also expect adaptation in the LFP activity to the effective stimuli. Thus, single unit spiking activity data are essential for distinguishing the different adaptation models.
It is important to consider that we employed a short-lag adaptation paradigm using familiar stimuli. Our spiking data show that in such a paradigm—which has been used in many fMRI studies—sharpening does not occur, at least to a measurable degree. However, a sharpening of selectivity may nonetheless occur in “adaptation” paradigms with long interstimulus intervals after exposure to novel stimuli. In fact, the sharpening concept was originally developed to explain the reduced activity seen when a previously novel stimulus becomes familiar (Li et al. 1993). Repeated presentations of stimuli can, under some circumstances, enhance the selectivity of IT neurons for highly familiar stimuli compared with novel stimuli (Freedman et al. 2006; Anderson et al. 2008). This sort of sharpening mechanism, however, differs from the mechanism that underlies the repetition suppression seen in the present study using short interstimulus intervals, familiar stimuli, and no intervening stimuli.
Comparison of the time courses of the responses to the test and adapter stimuli provided no evidence for the facilitation model: There was no faster response for the test stimulus compared with the adapter and the response to the test stimulus was similar to that of the adapter in duration. In fact, the strongest difference between the responses to the adapter and test stimuli is the difference in response amplitude, which is in accord with fatigue models.
Our data demonstrate that in general repetition suppression in IT using short stimulus presentations and ISIs shows little dependence on the strength of the response to the adapter, and by implication, little dependence upon firing-rate fatigue mechanisms in the recorded neuron. Thus, the degree of feature similarity—either shape or position, and most likely size (Verhoef et al. 2008; in that study, the neurons were not required to respond to all the sizes employed, thus a spiking-dependent mechanism could not have been ruled out)—between the adapter and test stimuli is a crucial factor with the strongest suppression occurring when adapter and test stimuli are matched. Such stimulus-specific adaptation effects have also been observed in area V1 using both long- (Movshon and Lennie 1979; for review see Kohn 2007) and short-duration adapters (Müller et al. 1999) but differ with regard to the effect of long-duration adaptation on the direction and speed tuning of MT neurons (Kohn and Movshon 2004; Krekelberg et al. 2005). Thus, this particular stimulus-dependent adaptation mechanism appears to be specific to particular cortical areas.
Note that for the spiking activity, we observed position-dependent adaptation throughout the response, unlike Lueschow et al. (1994). These different results may be due to differences in recording region (more medial/ventral, encroaching perirhinal cortex in Lueschow et al. 1994) and/or task (passive fixation vs. delayed matching to sample position-invariant stimulus recognition task).
There are 2 general mechanisms that can underlie input fatigue as modeled here: 1) an input-specific, synaptic depression mechanism (Abbott et al. 1997) that operates at the level of the neuron and 2) suppressed input from afferent neurons. In the latter case, the adapted input neurons can be part of the local IT network, from a lower level input region (i.e., bottom up input; with RFs that span the horizontal meridian), from higher-level regions (i.e., top down modulations), or any combination thereof. In all these cases however, the bulk of the adaptation of a single neuron is the result of adapted input of other neurons. The presence of gamma adaptation at onset suggests that the initial adaptation response depends on a synaptic suppression mechanism or adaptation of the feedforward input. However, the greater position invariance present in the later part of the gamma response may point to feedback from higher-level regions with larger RFs or recurrent activity in IT. Also, it is likely that the adaptation of the initial response, related to, for example, synaptic depression, is further amplified during the course of the response by recurrent activity within the cortical network of which the neuron is a part. Thus, the relative contributions of synaptic depression and input from other neurons to adaptation can shift during the course of the response, which may explain time course–dependent aspects of adaptation (Verhoef et al. 2008; Liu et al. 2009).
Part of the input that determines adaptation may result from cognitive factors such as attention. Similarity in the degrees of adaptation observed in the fixation and dimming tasks rules out differential attention to the adapter and test stimuli as a major factor in the present paradigm. However, it is possible and even likely that selective attention can modulate the degree of adaptation in IT. In fact, several human fMRI studies have documented effects of attention on adaptation suggesting that adaptation, at least that of the BOLD response, is less pronounced for competing unattended stimuli (Eger et al. 2004; Murray and Wojciulik 2004; Chee and Tan 2007; but see Bentley et al. 2003). Other high-level cognitive factors may also affect adaptation in IT, such as “expectancy” (Summerfield et al. 2008), but further research is needed to assess the contributions of such factors to IT adaptation.
It is becoming clear that adaptation is a complex phenomenon that can depend on several mechanisms, each of which can contribute differently under different paradigms (e.g., short vs. long-duration adaptation; short vs. long ISI adaptation) and within the course of the response. The unknowns regarding the degree to which adaptation is locally synaptic or that it stems from adapted local, downstream or upstream input makes it particularly challenging to examine the link between the selectivity of the adaptation effect, as estimated in BOLD fMR-A studies, and the tuning shown by a neuron or by population of neurons within a region. If the input to a neuron is predominantly local and the adaptation is primarily synaptic, the tuning of the adaptation effect can reflect the overall selectivity of the local population of neurons. Thus, the usefulness of adaptation for inferring local tuning may depend on the organization of the inputs of neurons within a region, which may differ between regions (Tanigawa et al. 2005).
Fonds voor Wetenschappelijk Onderzoek Vlaanderen (G0.0644.08), Geneeskundige Stichting Koningin Elisabeth (GSKE-Rufin Vogels 2008-2010), Geconcerteerde Onderzoeksacties (GOA/2005/18), Interuniversitaire attractiepolen (P6/29-C), Detection and Identification of Rare Audio-visual Cues (DIRAC-027787), and Excellentiefinanciering K.U. Leuven (EF/05/014).
The technical assistance of P. Kayenbergh, G. Meulemans, I. Puttemans, M. De Paep, W. Depuydt and S. Verstraeten is gratefully acknowledged. We thank R. Peeters and H. Kolster for their help in taking the MRI scans. Conflict of Interest: None declared.