Distinct early and late neural mechanisms regulate feature-specific sensory adaptation in the human visual system

Significance

Sensory adaptation is widely assumed to reflect a form of efficient coding in the brain, but its underlying neural mechanisms are debated. On one account, sensitivity to prolonged or repeated input is reduced due to neuronal fatigue. Alternatively, adaptation might reflect sharpening of neuronal selectivity over time. To adjudicate between these accounts, we recorded whole-brain activity from human observers before and after they adapted to visual grating stimuli, and used inverted encoding to characterize multivariate changes in neural representations. We identified patterns of neural activity consistent with both fatigue and sharpening accounts. Critically, however, these two mechanisms operate at different points in the sensory processing cascade, likely reflecting distinct feedforward and feedback interactions, to produce qualitatively distinct effects on perception.

Abstract

A canonical feature of sensory systems is that they adapt to prolonged or repeated inputs, suggesting the brain encodes the temporal context in which stimuli are embedded. Sensory adaptation has been observed in the central nervous systems of many animal species, using techniques sensitive to a broad range of spatiotemporal scales of neural activity. Two competing models have been proposed to account for the phenomenon. One assumes that adaptation reflects reduced neuronal sensitivity to sensory inputs over time (the “fatigue” account); the other posits that adaptation arises due to increased neuronal selectivity (the “sharpening” account). To adjudicate between these accounts, we exploited the well-known “tilt aftereffect”, which reflects adaptation to orientation information in visual stimuli. We recorded whole-brain activity with millisecond precision from human observers as they viewed oriented gratings before and after adaptation, and used inverted encoding modeling to characterize feature-specific neural responses. We found that both fatigue and sharpening mechanisms contribute to the tilt aftereffect, but that they operate at different points in the sensory processing cascade to produce qualitatively distinct outcomes. Specifically, fatigue operates during the initial stages of processing, consistent with tonic inhibition of feedforward responses, whereas sharpening occurs ~200 ms later, consistent with feedback or local recurrent activity. Our findings reconcile two major accounts of sensory adaptation, and reveal how this canonical process optimizes the detection of change in sensory inputs through efficient neural coding.

Prolonged exposure to an oriented edge or bar alters the observer’s perception of subsequently viewed nearby orientations. This sensory adaptation phenomenon, called the tilt aftereffect (Fig. 1A), is thought to reflect changes in the neural population code for orientation (1). There are two dominant explanations for how population responses change following adaptation, which are supported by partially overlapping behavioral and neural phenomena. The “fatigue” account describes how neural responses are reduced following adaptation, with the amount of reduction maximal around the adapted feature position (2, 3). The “sharpening” account describes how neural responses are more selective following adaptation, as tuning functions around the adapted feature position become narrower (4–6). To illustrate the difference between these accounts, consider the responses of orientation-selective neurons. For simplicity, we assume that prior to adaptation, the tuning curves for these neurons indicate similar responsiveness (Fig. 1 B, Upper). According to the neural fatigue account, adaptation to an anticlockwise-from-vertical orientation (Fig. 1B, magenta) reduces the responsivity of neurons to this orientation, resulting in a nonuniform distribution of responsiveness where the adapted orientation elicits the weakest response (Fig. 1 B, Left). By contrast, according to the neural sharpening account, adaptation results in narrowing of tuning curves around the adapted orientation (Fig. 1 B, Right). Despite decades of research on adaptation, it is unclear how the phenomena that support these disparate accounts of adaptation can be explained by a single unifying framework.

Fig. 1.

The tilt aftereffect and theoretical accounts of sensory adaptation. (A) Demonstration of the tilt aftereffect. Maintain fixation for 30 s on the blue spot inside the leftmost pattern (adapt), then view the middle pattern (test) and observe its orientation. Although the test pattern is vertically oriented, the orientation of the test pattern should now appear more similar to that of the rightmost pattern. (B) Illustration of neural “fatigue” and “sharpening” accounts of sensory adaptation.

Results

To adjudicate between fatigue and sharpening accounts, we recorded observers’ brain activity with electroencephalography (EEG) before (Fig. 2A, pre-adaptation sequence) and after adaptation (Fig. 2A, adaptation sequence), and then used inverted encoding to characterize multivariate changes in neural representations. In the pre-adaptation sequence, observers viewed a series of randomly oriented sinewave gratings (Fig. 2A, neural probe) and we characterized EEG responses to these orientations as a baseline measurement of neural activity. In the adaptation sequence, observers viewed gratings of the same orientation for a prolonged period to induce adaptation (Fig. 2A, adaptation) prior to the neural probe. In both the pre-adaptation and adaptation sequences, participants completed a detection task in which they counted the number of target (low spatial frequency) gratings presented in the neural probe (Fig. 2A, detection task). The detection task was intended to maintain the observer’s attention on a feature of the stimulus that was orthogonal to orientation (Fig. 2B).

Fig. 2.

Experimental design and perceptual orientation judgements. (A) Schematic of experimental design. To assess neural effects of adaptation we recorded observers’ brain activity with EEG while they viewed a series of randomly oriented gratings (5-Hz presentation; neural probe). In an initial pre-adaptation phase, participants performed a simple detection task that involved counting the number of low-spatial frequency target gratings within a sequence, the goal of which was to ensure participants maintained their attention on a feature of the stimulus that was orthogonal to orientation. In a subsequent adaptation phase, we additionally included an “adaptation” period in which gratings of the same orientation were presented (2-Hz presentation; initially 30 s, followed by 10-s top-up periods) prior to the “neural probe”. We also included an orientation reproduction task at the end of the trial (behavioral probe) to assess the effect of adaptation on perception; the red arrows and white lines are included in the schematic for illustrative purposes. (B) Performance on the detection task in pre-adaptation (gray) and adaptation (purple) trials; chance-level performance is indicated by the horizontal dashed line. (C) Aggregate responses to the behavioral probe, relative to the test orientation (vertical dashed line); distribution mean is shown by the gray arrow and positive values represent a repulsive bias away from the adapted orientation.

To measure adaptation in perception, we instructed observers to reproduce the orientation of a subsequently presented test grating, oriented ±10° from the adapted orientation (Fig. 2A, behavioral probe). As expected, we found a significant repulsive bias (1.43°) relative to the adapted orientation (CI_95% = [0.2431°, 2.6258°]; Fig. 2C). The magnitude of the bias was similar to values previously reported for similar stimuli (7). To measure corresponding neural changes associated with adaptation, we compared the matched recordings of responses with the neural probe taken prior to and post adaptation.

From the pre-adaptation neural probe data, we established that orientation information was primarily represented in parietal and occipital EEG sensors (SI Appendix, Fig. S2 and Method–Orientation Decoding Analysis), which was consistent with the spatial mapping of the voltage map (SI Appendix, Fig. S1); thus, we only included the signal from these 20 sensors in our analyses. We first characterized adaptation-related univariate changes in EEG activity (Fig. 3 A and B). Based on previous electrophysiological (8–13) and functional magnetic resonance imaging (fMRI) work (14–17), which has shown that prolonged/repeated exposure results in reduced neural responses, we would expect adaptation to reduce the amplitude of evoked EEG signals. We found significant differences between the univariate signals produced by oriented gratings following adaptation (Fig. 3C). To test whether these changes represented increases or decreases in amplitude, we expressed the differences between pre-adaptation and post-adaptation signals as the absolute difference from the baseline (zero) amplitude. We found that the response evoked by adapted orientations was both significantly smaller and larger than the same signals prior to adaptation (Fig. 3D).

Fig. 3.

Univariate neural responses. (A) Event-related potentials for stimuli oriented from ±16° around ±[0° (purple), 32° (cyan), 72° (green)] from the adapted orientation for pre-adaptation (A), and post-adaptation neural probes (B). Orientation bins are indicated by the colored disk, relative to the adapted orientation (purple). Solid and dashed boxes indicate the period of stimulus presentation for the binned and subsequent orientations, respectively. Shaded error regions indicate ± SEM. (C) The difference between pre-adaptation and post-adaptation. Black horizontal bars indicate cluster corrected periods that showed a significant main effect of orientation. (D) Same as C, but expressed as the distance from the baseline (zero) amplitude, i.e., positive and negative values indicate larger and smaller amplitude signals following adaptation, respectively. Colored horizontal bars indicate cluster corrected periods that showed a significant difference from zero. The plots at the Bottom of C and D show the Bayes factors associated with significance testing.

We next sought to characterize how adaptation alters the neural representation of orientation using inverted encoding analyses (18, 19). We obtained an inverted model using the pre-adaptation EEG signals. We then compared the results yielded from decoding orientation from the same pre-adaptation signals (pre) with those yielded from decoding of orientation from the post-adaptation signals (post). We sorted the gratings into five orientation bins (±16° around −72°, −32°, 0°, +32°, and +72°) relative to the adapted orientation. For each stimulus presentation, we decoded the orientation from the EEG signal. From this, we calculated average decoding accuracy (indexed from −1 to 1, where 1 indicates 0 angular distance between presented and decoded orientations, and −1 indicates 90° distance), precision (the inverse of the SD between decoded orientations within each bin), and bias (the average decoded angle for each orientation bin relative to the presented orientation).

Consistent with previous work on decoding of orientation from EEG signals (20, 21), pre-adaptation and post-adaptation decoding accuracy rose sharply from ~50 ms following stimulus onset and gradually reduced over the following 400 ms. However, while orientation could be reliably decoded from both sets of signals, we found that accuracy was significantly higher for pre-adaptation signals (average F_1,71 = 22.92, P = 8.94e⁻⁶; Fig. 4A). By contrast, the precision of neural responses, which also rose sharply at ~50 ms following stimulus onset, was significantly higher for post-adaptation signals (average F_1,71 = 26.04, P = 2.68e⁻⁶; Fig. 4B). Three adaptation-induced differences in decoding precision are readily observable in Fig. 4B. First, post-adaptation precision is significantly higher than pre-adaptation precision even before the stimulus is presented. This indicates that adaptation biased the neural representation of orientation such that some orientations were more likely to be decoded, even in the period prior to the epoch-locked stimulus presentation. Second, there are periods of increased post-adaptation precision, relative to pre-adaptation, which exceed the difference observed before ~50 ms following stimulus onset. Third, there appears to be a double peak in precision between 50 and 250 ms following stimulus onset, which may indicate that the mechanism responsible for the increased precision is subject to oscillatory dynamics. In contrast to accuracy and precision, we found no differences in bias (Fig. 4C), which is what we would expect when averaged across all orientations. Broadly, these results suggest that adaptation decreases the accuracy, while increasing the precision, of neural representations.

Fig. 4.

Adaptation-related changes to neural representations. (A) Decoding accuracy, (B) precision (inverse SD), and (C) bias, averaged across all orientations, of the inverted model tested on pre-adaptation and post-adaptation data as a function of time from stimulus onset. Solid and dashed boxes indicate the period of stimulus presentation for the binned and subsequent orientations, respectively. Shaded error bars in A and B indicate SEM; shaded error bars in C indicate CI_95%; black horizontal bars indicate cluster corrected periods that showed a significant difference between pre-adaptation and post-adaptation. Note that this difference was significant for the entire period displayed in B. The plots at the Bottom of A and B show the Bayes factors associated with significance testing. The gray arrow highlights the period before stimulus presentation in which there is increased precision for post-adaptation signals.

To further explore adaptation-related changes in neural representations, we compared pre-adaptation and post-adaptation data using two sets of analyses: One that focused on differences between orientations, and another that examined differences over time. To test how adaptation altered the representation of different orientations, we obtained an inverted model using a variation of a recently described method (22), which allowed us to sort responses to the presented grating stimuli into 180 orientation bins (one for each orientation rounded to the nearest degree) from which we could calculate decoding accuracy, precision, and bias parameters. To maximize the signal-to-noise ratio for each bin, we averaged results over the period in which there was above-chance decoding (50 to 450 ms).

As we expected, for pre-adaptation signals, there were no clear differences in accuracy, precision, or bias as a function of angular distance from the adapted orientation (Fig. 5 A–C, Top Row). By contrast, for post-adaptation signals, accuracy was reduced, and precision was increased around the adapted orientation (Fig. 5 A and B, Bottom Row). Of note, precision was increased across all orientations of post-adaptation relative to pre-adaptation. There were also changes in bias around the adapted orientation (Fig. 5C). Model comparison of the changes in parameter estimates as a function of orientation confirmed that while the pre-adaptation parameters could be adequately described by a uniform function (all ΔAIC < 5), the post-adaptation parameters were better described by von Mises (accuracy and precision) and first-order Gaussian derivative functions (bias; all ΔAIC > 38).

Fig. 5.

Feature-specific adaptation-related changes to neural representations. (A–C) The time-averaged (A) accuracy, (B) precision, and (C) bias of the inverted model responses, at all orientations, for pre-adaptation (Top, black) and post-adaptation (Bottom, red) data. Note the increased precision across all orientations following adaptation, which is consistent with the results from the previous orientation-averaged analysis. Vertical magenta line indicates the adapted orientation; black and red lines show the model fits to the data. (D) Results of data simulations assuming different models of adaptation. The left column shows the neural response functions from which pre-adaptation and post-adaptation data were derived. The columns to the right are the results, displayed in the same format used in A–C.

The shape of the bias decoded from the neural responses matches previously reported behavioral biases associated with the tilt aftereffect (7), but the increased precision observed around the adapted orientation is a surprising result, without precedent. To test whether we could identify this effect in the behavioral responses, we assessed the relationship between the precision and bias of participants’ orientation estimates. We reasoned that if adaptation results in increased precision around the adapted orientation, in addition to the conventional repulsive bias, then participants with more precise judgements should have larger biases. In line with the results from the neural data, there was a significant correlation between participants’ behavioral bias and their precision (r = −0.36, P = 0.0298), such that those with larger repulsive biases tended to make more precise judgements.

Having demonstrated reliable changes in the neural representation of orientation following adaptation, we sought to determine whether these results were more consistent with the fatigue or sharpening accounts of sensory adaptation. We simulated EEG data from neural tuning functions that were either reduced in amplitude (fatigue account) or narrowed around a selected orientation (sharpening account), and used inverted encoding to estimate accuracy, precision, and bias as a function of orientation (Fig. 5D). While both models predicted reduced accuracy around the adapted orientation, the sharpening model failed to predict increased precision and the fatigue model failed to predict a repulsive bias, suggesting that neither model alone is sufficient to explain adaptation. These results are consistent with previous modeling work by Kohn and Movshon (4), who applied similar models of fatigue and sharpening, which also included shifted tuning preferences, and found that fatigue produces no bias while sharpening produces a repulsive pattern of biases. We then tested a model which combined both reduced and narrowed tuning and found that this captured the changes associated with accuracy, precision, and bias observed in the empirical data. The combined model also best captured the pattern of results, i.e., increased precision, observed prior to stimulus onset (SI Appendix, Fig. S3). These results provide a unifying explanation for the effects of adaptation on neural activity. Specifically, the simulations reveal that both neural fatigue and sharpening occur and, perhaps counterintuitively, that they are associated with increased precision and repulsive biases, respectively.

Single-unit recordings in animal models indicate that neurons’ preferred tuning is shifted following sensory adaptation (4, 11, 23, 24). Although our primary aim was to adjudicate between sharpening and fatigue models of adaptation, for completeness, we also simulated shifting the preferred orientation of neural response functions. The simulation results suggest that neither shifted tuning preference in isolation, nor when combined with sharpening or fatigue, captured the empirical results (SI Appendix, Fig. S4). This apparent inconsistency may indicate that even though adaptation changes the preferred tuning of some neurons, the effects of adaptation at the population level are better explained by sharpening and fatigue.

We next sought to investigate the temporal dynamics of adaptation. To test when adaptation altered the neural representation of orientation, we sorted neural probe presentations into five bins (±16° around −72°, −32°, 0°, +32°, and +72°) relative to the adapted orientation and used inverted encoding to calculate accuracy, precision, and decoded angle at each time point within the period around stimulus presentation. Prior to examining temporal dynamics, we expressed the channels as a bank of tuning curves, which represent the average response between 50 and 450 ms after stimulus onset (Fig. 6A). The pre-adaptation tuning curves were of similar amplitude and centered on their corresponding channel orientations (Fig. 6A, dashed vertical lines). In contrast, for the post-adaptation responses, the amplitude of the adapted orientation tuning curve was reduced, and the flanking orientation curves were biased away from the adapted orientation. To quantify these changes, we calculated accuracy, precision, and bias for each of the five channels. We combined data from channels with positive and negative offsets (e.g., –32° and +32°) to create three offset bins ±16° around 0°, |32°|, and |72°|, relative to the adapted orientation. We found the same pattern of results as in the previous orientation-focused analysis. That is, accuracy was reduced (CI_95% = [−0.1445, −0.3019]; Fig. 6B, purple) and precision increased (CI_95% = [1.9360, 0.9413]; Fig. 6C, purple) for the adapted orientation, and there was a repulsive bias for the flanking orientation (CI_95% = [1.4218°, 6.8395°]; Fig. 6D, cyan). Encouraged by the consistency of these results, we calculated these parameters as a function of time relative to stimulus onset to examine their temporal dynamics.

Fig. 6.

Dynamic adaptation-related changes in neural representations. (A) Time-averaged responses of each of the five inverted encoding channels, at all possible orientations, decoded from pre-adaptation and post-adaptation signals. Note, these curves are distinct from the neural tuning functions used to generate the simulated EEG data; see Methods–Orientation Decoding Analysis for further details. Average difference in decoding (B) accuracy, (C) precision, and (D) bias parameters between pre-adaptation and post-adaptation for channels at three absolute offsets (0, 32, and 72°) from the adapted orientation. Orientation bins are indicated by the colored disk, relative to the adapted orientation (purple). Error bars indicate CI_95%. Average (E) accuracy, (F) precision, and (G) bias, as a function of time from stimulus onset for channels at three absolute offsets from the adapted orientation decoded from (Left) pre-adaptation and (Right) post-adaptation signals. Note the bias represents the center of the channels. Shaded error regions in A, E, and F and G indicate ± SEM and CI_95%, respectively; black horizontal bars in E and F and black filled in region in G indicate cluster corrected periods that showed a significant difference between orientations or pre-adaptation and post-adaptation, respectively. Solid and dashed boxes indicate the period of stimulus presentation for the binned and subsequent orientations, respectively.

As expected, there were no differences in pre-adaptation accuracy or precision between orientations (Fig. 6 E, Left). By contrast, there was a period of differences (~50 ms) in post-adaptation accuracy around 350 ms after stimulus onset (average F_2,71 = 4.60, P = 0.013; Fig. 6 E, Right). In this period, the accuracy was highest for orientations orthogonal to the adaptor (i.e., |72°| bin) and lowest for orientations similar to the adaptor (0° and |32°| bins). The timing of accuracy modulation is unexpectedly late, given that orientation is reliably decodable 300 ms earlier. However, there is evidence from previous EEG work using more complex visual stimuli indicating that effects of adaptation first become detectable around this time (25–28). This delay has previously been interpreted as indicating that adaptation effects are not present during initial feedforward processing of visual inputs, but arise instead as a consequence of feedback from higher level (e.g., prefrontal) brain regions. Single-unit recordings suggest adaptation may arise, at least partially, from local recurrent activity (4, 29), which might also explain the delayed changes in accuracy observed here. Another possibility is that there are changes in accuracy earlier than 300 ms that we failed to detect with our cluster correction analyses; indeed, there appear to be brief periods of notable differences in accuracy between 50 and 200 ms following stimulus onset. To test this, we applied a different cluster correction method (Threshold Free Cluster Enhancement) that is more sensitive to high frequency effects (30). This analysis, however, also yielded no significant clusters within the first half of the epoch (SI Appendix, Fig. S5).

As with accuracy, there were no differences in pre-adaptation precision between orientations, but there were three periods during which there were differences in post-adaptation precision (average F_2,71 = 6.41, P = 0.003; Fig. 6F). In all periods, the precision was highest for the orientations that were similar to the adaptor (i.e., 0° bin). Analysis of the bias revealed a period of repulsion in the near-to-adapted orientations (i.e., |32°| bin) at ~350 ms, consistent with the difference observed in accuracy (Watson–Williams test, average z = 2.02, P = 0.022; Fig. 6G). This shows that the change in accuracy observed at ~350 ms after stimulus onset represents the well-known tilt aftereffect bias.

None of the periods in which there were changes in precision aligned with those in which there were changes in accuracy and bias. Specifically, we found strong evidence against a positive relationship between the timing of changes in precision and accuracy (r = −0.052, BF₊₀ = 0.049). Moreover, the changes in precision were present from the very earliest moments at which orientation was reliably decodable. The topographic patterns of orientation responses showed slight differences over the course of the epoch (0 to 500 ms following stimulus onset; SI Appendix, Fig. S2C), the spatial mapping of which may relate to the difference between pre-adaptation and post-adaptation univariate responses. However, the spatial resolution of EEG is not sufficient to draw precise inferences about the spatial origins of these effects and the signal-to-noise ratio is too low to reliably decode orientation when the data are split in both spatial and temporal dimensions. The spatial mapping of these effects could be assessed in future work using a neuroimaging technique with higher spatial resolution, e.g., fMRI or magnetoencephalography (MEG). Taken together, these results suggest there are two distinct processes that operate during adaptation: An early mechanism that influences precision during feedforward processing, and a later mechanism that influences accuracy (possibly due to feedback or local recurrent connections). Combined with the results from our simulations, these findings indicate that neural fatigue operates during the initial feedforward processing of a stimulus and produces more precise representations, whereas neural sharpening operates at a later stage and produces repulsive shifts in the representation.

Unlike the neural responses recorded throughout each trial, observers’ behavioral responses comprise a single snapshot of perception, which presumably represents the culmination of perceived orientation over time; thus, we cannot directly test the temporal dynamics of the behavioral response. We can, however, test whether behavioral responses support the late-stage repulsion effect found in the neural data by comparing the relationship between observers’ perceptual judgements and either early- or late-stage neural responses. We reasoned that if the late-stage differences observed in the neural representation reflect those that are responsible for the perceptual repulsion effect, then late- but not early-stage adaptation-related neural changes should be more likely to be predictive of behavioral biases. In line with this reasoning, we found that late- (250 to 450 ms) but not early-stage (50 to 250 ms) differences in the pre–post neural representation significantly predicted trial-level behavioral biases (early: average r = −0.004, t₃₅ = −0.162, P = 0.861; late: average r = 0.046, t₃₅ = 2. 644, P = 0.012). The late stage neural-behavioral correlation, although significant, has a relatively small effect size. This may be because the behavioral responses from the reproduction task are subject to sources of variance that the neural data are not subject to, namely, variance associated with the perceptual decision and motor response execution.

Discussion

Adaptation shapes perception by making neurons sensitive to the temporal context in which stimuli are embedded. This process is evident through phenomena such as the tilt aftereffect (Fig. 1A), and has been reliably measured across the brain using a variety of techniques sensitive to a broad range of spatiotemporal scales of neural activity (5, 9, 14, 25, 27). While our understanding of the neural mechanisms of adaptation has largely been informed by single-cell recordings in animal models, there have also been demonstrations of the phenomenon in humans using noninvasive techniques that measure neural activity at the population level such as fMRI and EEG. However, these studies have typically measured univariate changes in neural activity, e.g., repetition suppression (14, 31), which overlooks more nuanced effects of adaptation. Here, we used inverted encoding to establish converging lines of behavioral and neural evidence that adaptation results in heightened precision around the adapted orientation. This finding provides empirical evidence for the hypothesis that adaptation supports perception by shaping neural responses in a manner that increases the salience of changes to relevant stimuli over time (32–34).

Previous electrophysiological work has reported mixed evidence on the time course of adaptation. Single-cell recordings in animal models show that the activity of some neurons in response to adapted stimuli is changed at the earliest stages of (feedforward) processing (5, 9, 11, 35), whereas EEG recordings of neural population activity in humans have suggested differences in univariate responses to adapted stimuli are first observed at later (feedback) stages of processing (25–28). Consistent with previous neurophysiological work, here we demonstrate univariate changes in EEG activity that scale with similarity to the adapted stimulus feature and emerge during the initial stages of sensory processing. Critically, however, our multivariate analyses reveal adaptation-related changes to the representation of visual grating stimuli that seem to have distinct temporal dynamics. During early feedforward processing, changes in the neural representation are associated with increased precision around the adapted orientation, which our simulations show can be explained by neural fatigue. Changes during later feedback processing, by contrast, are associated with neural and perceptual biases, which can be explained by neural sharpening. These results are in agreement with single-unit recording work suggesting that adaptation produces dynamic, complex, patterns of both attenuation and enhancement (36). They are also consistent with our univariate analyses, which revealed that adaptation produces a dynamic pattern of feature-specific increases and decreases in amplitude.

In conclusion, we have proposed a unifying framework for the behavioral and neural bases of sensory adaptation. Our work reconciles neural sharpening and fatigue accounts of visual adaptation, and demonstrates how adaptation supports brain function by increasing sensitivity to change. Our findings reveal the possibility of two discrete adaptive neural mechanisms that operate at different points in the sensory processing cascade to produce qualitatively distinct perceptual outcomes: Early increased precision and late reduced (biased) accuracy.

Methods

Participants.

Thirty-seven neurotypical human adults (mean ± SD age, 23.8 ± 4.6 y; 23 females) participated in the experiment. Observers were recruited from The University of Queensland, had normal or corrected-to-normal vision (assessed using a standard Snellen eye chart), and were required to pass an initial screening session to qualify for the experiment (see Stimuli, Task, and Procedure section for details). Data from one participant were omitted from analyses due to hardware failure. All participants were naïve to the aims of the experiment and gave informed written consent. The experiment was approved by The University of Queensland Human Research Ethics Committee.

Apparatus.

The experiment was conducted in a dark, acoustically and electromagnetically shielded room. The stimuli were presented on a 24-inch ViewPixx monitor (VPixx technologies, Inc., Montreal) with 1,920 × 1,080 resolution and a refresh rate of 144 Hz. Viewing distance was maintained at 45 cm using a chinrest, meaning the screen subtended 61.18° × 36.87° (each pixel 2.4’ × 2.4’). Stimuli were generated in MATLAB (The MathWorks, Inc., Matick, MA) using Psychophysics Toolbox ((37, 38); see http://psychtoolbox.org/). EEG signals were recorded using 64 Ag-AgCl electrodes (BioSemi, Amsterdam, Netherlands).

Stimuli, Task, and Procedure.

The stimuli comprised sinewave gratings (1 cycle/°, 0.5 contrast, random phase) presented centrally within a circular aperture (radius 4.2°), which was smoothed at the edges, on a mid-gray background. A centrally positioned green fixation dot (radius 0.25°) was presented to reduce eye movements. To maintain attention, participants were instructed to count the number of “target” stimuli, in which the spatial frequency of the grating was reduced (0.66 cycle/°). Between 0 and 3 targets appeared during each trial, selected at random.

There were two types of trial: pre-adaptation and adaptation. Pre-adaptation trials consisted of grating stimuli (orientations randomly selected between 0 and 180°) presented for 0.05 s each, separated by a blank 0.15-s interstimulus-interval (ISI) for 10 s (neural probe). The numbers 0 to 3 were then displayed on the screen and participants were given 2 s to indicate the number of targets presented, using a computer mouse (detection task). This was repeated 12 times per block. Participants performed 6 blocks of pre-adaptation trials (~20 min), receiving feedback on their detection accuracy at the end of each block. Immediately following the pre-adaptation blocks, adaptation blocks were run. Adaptation blocks were the same as pre-adaptation blocks, except that they included an adaptation phase prior to the neural probe sequence, and an additional orientation reproduction task that followed the detection task report. The adaptation phase comprised repeated presentations of a grating with a fixed orientation (0.4 s duration, 0.1 s blank ISI) for either 30 s (first block trial) or 10 s (subsequent block trials). The orientation reproduction task consisted of the brief presentation (0.1 s) of a grating oriented at ±10° from the adapting orientation, followed by presentation of a straight line (random start angle) intersecting the fixation point, partially occluded by an imaginary disk equal to the size of the stimulus. Participants were instructed to use the mouse to rotate the line so that it matched the orientation of the preceding grating. When psychophysically measuring adaptation aftereffects, it is common to measure the behavioral response directly after the adaptor, in order to estimate the effects of adaptation during the period in which they are likely to be strongest. In our experimental design, there was a ~12-s delay between the adaptor and the behavioral measurement, which likely reduced the magnitude of the behavioral estimate. We designed the experiment as such to match the structure of the pre-adaptation and adaptation blocks as closely as possible, and to measure neural changes to adaptation when they were most likely to be strongest, i.e., immediately following adaptation. Adaptation blocks comprised 6 trials, and participants completed 12 blocks (~40 min). In pre-adaptation trials, targets could appear during the neural probe; in adaptation trials they could additionally appear during the adaptation period.

Each participant was assigned one of 16 adaptor orientations, linearly spaced between 0 and 180°; the orientations were counterbalanced across the 37 participants. Prior to the main experiment, participants were required to pass an initial screening session in which they completed the same task but without EEG. To pass the screening, participants were required to perform above chance (25%) on the detection task and show a significant repulsive bias on the orientation reproduction task (assessed using a one-tailed t test). This resulted in approximately half of the participants being screened out (32 due to insufficient repulsive bias, and 8 due to both insufficient repulsive bias and low detection task performance). Screening and experiment sessions were separated by a minimum of 24 h. Note that the individuals who were screened out are not included in the Participants section.

EEG.

The EEG signals were digitized at 1,024-Hz sampling rate with a 24-bit A/D conversion. The 64 active scalp Ag/AgCl electrodes were arranged according to the international standard 10 to 20 system for electrode placement (39) using a nylon head cap. As per BioSemi system design, the common mode sense and driven right leg electrodes served as the ground, and all scalp electrodes were referenced to the common mode sense during recording. Offline EEG preprocessing was performed using EEGLAB in accordance with best practice procedures (40, 41). The data were initially down-sampled to 512 Hz and subjected to a 0.5-Hz high-pass filter to remove slow baseline drifts. Electrical line noise was removed using pop_cleanline.m, and clean_artifacts.m in EEGLAB (42) was used to remove bad channels (identified using Artifact Subspace Reconstruction), which were then interpolated from the neighboring electrodes. Data were then re-referenced to the common average before being epoched into segments around each neural probe stimulus (−0.25 s to 0.5 s from the stimulus onset). Systematic artifacts from eye blinks, movements and muscle activity were identified using semi-automated procedures in the SASICA toolbox (43) and regressed out of the signal.

Univariate Analysis.

To characterize the changes in the univariate response to orientation, we averaged responses separately for pre-adaptation and post-adaptation data across sensors and across gratings that were positively and negatively offset from the adapted orientation for three absolute offsets (±[0, 32, 72]°). We then repeated this for the difference between pre-adaptation and post-adaptation responses. Finally, to characterize how the absolute amplitude of responses changed following adaptation, we calculated the difference between the absolute pre-adaptation and post-adaptation responses:

$$\mathbf{A}=\left|pre\right|-\left|post\right|$$ [1]

Orientation Decoding Analysis.

To characterize sensory representations of the stimuli, we used an inverted modeling approach to reconstruct the orientation of the gratings from the EEG signal (44). Briefly, a theoretical (forward) model was nominated that described the measured activity in the EEG sensors given the orientation of the presented grating. The forward model was then used to obtain the inverse model that described the transformation from EEG sensor activity to stimulus orientation. The forward and inverse models were obtained using the pre-adaptation neural probe data. The inverse model was then applied to the post-adaptation neural probe data to test the effects of adaptation on the neural representation of orientation. As a baseline comparison, we also applied the inverse model to the pre-adaptation data, using a ten-fold cross-validation approach in which 90% of the data were used to obtain the inverse model on which the remaining 10% were decoded.

Consistent with previous work (18), the forward model comprised five hypothetical channels, with evenly distributed idealized orientation preferences between 0° and 180°. Each channel consisted of a half-wave rectified sinusoid raised to the fifth power. The channels were arranged such that a tuning curve of any orientation preference could be expressed as a weighted sum of the five channels. The observed EEG activity for each presentation could be described by the following linear model:

$$\mathbf{B}=\mathbf{WC}+\mathbf{E}$$ [2]

where $\mathbf{B}$ indicates the (m sensors × n presentations) EEG data, $\mathbf{W}$ is a weight matrix (m sensors × 5 channels) that describes the transformation from EEG activity to stimulus orientation, $\mathbf{C}$ denotes the hypothesized channel activities (5 channels × n trials), and $\mathbf{E}$ indicates the residuals errors.

To compute the inverse model, we estimated the weights that, when applied to the data, would reconstruct the underlying channel activities with the least error. Similar to previous MEG work (21, 45), when computing the inverse model, we deviated from the approach proposed by (18) by taking the noise covariance into account to optimize it for EEG data, given the high correlations between neighboring sensors.

We ran two sets of orientation decoding analyses; one that focused on differences between orientations and one that focused on differences over time. In the orientation focused analyses (Fig. 5), we modified a method recently introduced by (22), which uses sparsely distributed channels to produce a dense representation of channel responses. Briefly, the standard inverted modeling approach typically involves training and testing a model using a relatively small number of channels (e.g., ref. 5) to represent the entire spectrum of identities across a given feature (e.g., orientation). The ‘enhanced inverted encoding model’ (eIEM) method involves repeatedly computing the inverse model while rotating the forward model channel orientation preferences until each orientation is represented by a channel response (e.g., 180 channels, one for each orientation). The eIEM method was established using data with feature identities sampled from discrete bins, which determines: a) the number of channels used in each model, and b) the number of feature identities from which predictions could be derived from the dense output. By contrast, we presented observers with orientations sampled from a continuous uniform distribution of all orientations. Thus, while we sorted presentations into five orientation bins to maintain sufficient data to estimate the weights for each channel, after rotating channels to produce 180 channel responses for each trial, we then binned the responses according to the orientation nearest to that presented in each trial. This allowed us to produce high-resolution model predictions for each orientation. We performed this analysis at each time point within the epoch, but to improve the signal of the estimates we averaged predictions across time points when there was reliable decoding accuracy (50 to 450 ms following stimulus onset).

For the analyses that focused on temporal dynamics (Fig. 6), we only used the original five channel responses, and we did not average channel responses across time. Channel orientation preferences were aligned at five positions (±16° around [−72, −32, 0, 32, 72]°) relative to the adapted orientation. Note that we found cardinal biases in the model predictions, but these were mitigated by averaging across participants after aligning the channels to the 16 different adapted orientations used. For temporal analyses, we averaged responses from positively and negatively offset channels to produce decoded orientations associated with three absolute offsets from the adapted orientation (±[0, 32, 72]°). Prior to assessing temporal dynamics of the neural representation, as a sanity check, we constructed tuning curves from each of the five channel responses. The tuning curves shown in Fig. 6A were computed by taking the dot product of the channel responses and the forward model.

For each presentation, we decoded orientation by converting the channel responses to polar form:

$$\widehat{\theta } = \frac{arg\left(z\right)}{2}, z=\mathit{c}{e}^{2i\varphi }$$ [3]

where $\mathit{c}$ is a vector of channel responses and $\varphi$ is the vector of angles at which the channels peak (multiplied by two to project 180° orientation space onto the full 360° space). From the decoded orientations, we computed three estimates: accuracy, precision, and bias. Accuracy represented the absolute angular difference between the decoded and presented (binned) orientation, and was expressed by projecting the difference onto a vector with 0°:

$$r_{\theta }=\cos \left(\delta z_{\theta }\right), \delta z_{\theta } = i\left(\widehat{\theta }-\theta \right)$$ [4]

where $r$ indicates the accuracy of the decoded signal. Thus, accuracy ranged from −1 to 1, indicating that the decoded orientation was the opposite to or the same as the presented (binned) orientation, respectively. Accuracy estimates were averaged across trials within the same bin to produce one measurement for each orientation. Precision was estimated by calculating the circular SD of the decoded orientations within each bin and expressed in degrees as the reciprocal of this value, such that higher values indicated better precision. Chance-level precision was estimated as the reciprocal of the SD of uniformly distributed angles. Bias was estimated by computing the average signed angular difference between the decoded and presented orientation such that positive values indicated a repulsion effect.

Prior to orientation decoding analyses, we established the sensors that contained the most orientation information by treating time as the decoding dimension and obtaining inverse models for each sensor, using ten-fold cross-validation with the pre-adaptation data. This analysis revealed that orientation was primarily represented in posterior sensors; thus, for all subsequent analyses we only included signals from the parietal, parietal-occipital, occipital, and inion sensors to compute the inverse model. Specifically, we included signal from the following sensors in the decoding analyses: P_1-10, P_z, PO_3-4, PO_7-8, PO_z, O_1-2, O_z, and I_z.

For the neural-behavioral correlations, we quantified the effect of neural adaptation as the difference between pre-adaptation and post-adaptation inverted model responses evoked during the neural probe prior to the behavioral response. To test the relationship between early and late adaptation effects, we split these data into responses between 50 to 250 ms and 250 to 450 ms following each stimulus onset within the neural probe sequence.

Model Simulations.

To characterize changes in neural tuning which might have given rise to the empirical results observed, we used inverted modeling to decode orientation from simulated EEG data produced by neural populations that were either sharpened, fatigued, or both sharpened and fatigued around a common orientation. Simulated data were produced by assuming variable weights between a bank of response curves, each of which represented the aggregate responses of neural populations with similar tuning, with evenly distributed orientation preference (n = 16) and EEG sensors (m = 32). As with the forward model, the neural functions were arranged such that a tuning curve of any orientation preference could be expressed as a weighted sum of the 16 functions. Thus, for each simulated presentation, the EEG activity was computed as:

$$s_j={\mathit{c}}_{\theta }\mathbf{w}+\mathbf{u}$$ [5]

where $s_j$ denotes the activity of sensor $j$, ${\mathbf{c}}_{\theta }$ indicates the tuning curve of orientation $\theta$, $\mathbf{w}$ denotes the weights between the sensor $j$ and the n response functions, and $\mathbf{u}$ indicates Gaussian noise (SD = 2). Neural tuning consisted of von Mises functions (κ = 2), normalized such that they ranged from 0 to 1. Neuron-to-sensor weights were randomly assigned from a uniform distribution between 0 and 1. Neural sharpening was modeled by increasing the κ parameter of neural tuning functions according to a von Mises function (κ = 4), centered on the adapted orientation, normalized between 0 and 4. This was intended to capture the sharpening account of adaptation, in which prolonged/repeated exposure results in narrower tuning functions around the adapted feature and is claimed to reflect a reduced number of neurons firing in response to stimuli (5, 6). Neural fatigue was modeled by subtracting a von Mises function (κ = 4), centered on the adapted orientation and normalized to between 0 and 0.5, from the amplitude of each tuning function. This was intended to capture the fatigue account of adaptation, in which prolonged/repeated exposure leads to reduced responsivity that varies as a function of proximity to the adapted stimulus feature (2, 3). Combined neural sharpening and fatigue were modeled by first applying sharpening and then fatigue manipulations.

In line with the empirical experiment, the orientation of each simulated presentation was drawn from a uniform distribution between 1 and 180°. We simulated 3,600 pre-adaptation presentations from the unmodified neural tuning functions and 3,600 post-adaptation presentations from each of the modified tuning functions. We then applied the same orientation focused analyses as used on the empirical data to estimate accuracy, precision, and bias as a function of orientation. The final presented results are the average of parameters estimated from 36 simulated datasets.

In the empirical EEG data, we found differences in the post-adaptation parameter estimates, relative to those of the pre-adaptation, prior to stimulus onset and processing. As a further test of the adaptation models, we simulated data that represented the period before stimulus information was available, and compared the results with the empirical observations. Stimuli in the empirical experiment were presented at 5 Hz, and we showed that orientation information is still reliably decodable at 450 ms after stimulus onset; thus, prestimulus neural activity reflects the response to the previous stimulus orientation, which was (by design) uncorrelated with the epoch-locked stimulus. To simulate the prestimulus period, therefore, we shuffled the labels of the pre-adaptation and post-adaptation presentations and estimated accuracy, precision, and bias from the responses.

Statistical Analyses.

Statistical analyses were performed in MATLAB. The behavioral responses from the orientation reproduction task were expressed as their angular distance from the test orientation. Where the test orientation was negatively offset from the adapted orientation, the sign of the angular distance was reversed. Thus, responses were normalized to represent the angular distance from the test orientation, where negative and positive values indicate attractive and repulsive biases, respectively, relative to the adapted orientation. Prior to performing statistical inferences, outlier rejection was performed on the data. In particular, normalized behavioral orientation responses that were >2.5 SD from each observer’s mean response were omitted. A circular mean test, equivalent to a one-sample t test, was used to infer whether the average bias across participants differed significantly from zero, i.e., was within the 95% CI. A circular-linear correlation test was used to infer the significance of relationships between behavioral and neural responses (46). Specifically, we calculated the correlation between neural and behavioral estimates of adaptation separately for each participant, then used a one-sample t test to assess whether the distribution of correlations differed significantly from zero.

For analyses of differences in univariate responses and parameter estimates between orientations as a function of time, a cluster correction was applied to remove spurious significant differences. First, at each time point, the effect size of orientation was calculated. For linear data, a repeated measures analysis of variance was applied to calculate the F statistic associated with orientation; for the angular bias data, a Watson-Williams test (47) was used to calculate the t statistic associated with the difference between the pre-adaptation and post-adaptation bias. Thus, for accuracy and precision, we sought to detect differences between orientations within pre-adaptation or post-adaptation phases, whereas for bias we sought to determine whether there was a significant difference between pre-adaptation and post-adaptation separately for each orientation. Next, we calculated the summed value of these statistics (separately for positive and negative values) within contiguous temporal clusters of significant values. We then simulated the null distribution of the maximum summed cluster values using permutation (n = 1,000) of either the orientation labels (linear data) or the sign (angular data), from which we derived the 95% percentile threshold value. Clusters identified in the data with a summed effect-size value less than the threshold were considered spurious and removed.

To characterize the difference in parameter estimates as a function of distance from the adapted orientation, we used the Akaike information criterion (AIC) to compare the fit of either a von Mises function (accuracy and precision) or a first-order Gaussian derivative function (bias) with a uniform function defined by a single offset parameter. We reasoned that if there were no reliable differences in parameter estimates, the data should be fit best by a uniform function.

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

EEG recordings and behavioural responses data have been deposited in OSF (https://osf.io/5ba9y/) (48).