Early Cortical Sensitivity and Speeded Target Selection Underlie Incidentally Learned Prioritization of Visual Features
- 1Psychological and Brain Sciences, Dartmouth College, Hanover, New Hampshire 03755-3565
- 2Psychology Department, Gonzaga University, Spokane, Washington 99258
-
Author contributions: K.O., D.A.A., and V.S.S. designed research; K.O. performed research; K.O. analyzed data; K.O., D.A.A., and V.S.S. wrote the paper.
Abstract
Adaptive behavior relies on prioritizing relevant sensory information, and decades of research have shown that current task goals and stimulus saliency influence this prioritization. Recent behavioral work indicates that incidental experience with frequently relevant locations or nonspatial features also shapes behavioral prioritization. The present study investigates the neural processing stages affected by incidental learning of nonspatial visual features. We recorded neural activity with high temporal resolution using electroencephalography while human participants (female and male) searched for visual targets that had predictable features (i.e., that appeared more frequently in a particular color). We found that incidental learning of the statistical structure was accompanied by an early differentiation of neural activity for relevant compared with irrelevant features beginning at ∼120 ms post stimulus onset, followed by an earlier-onset selection of the target item, as indexed by a latency shift of the N2pc (∼200 ms), and changes in later memory and response-related processes, marked by amplitude modulations of the LPC (>400 ms). Importantly, the magnitude of the effects across all three neural measures strongly tracked individual differences in the behavioral benefits of learned prioritization, suggesting that successful learning of feature regularities depends on modulating the flow of information across multiple processing stages.
Significance Statement
The human brain is highly sensitive to the statistical structure of the sensory environment. Behavioral work has shown that participants rapidly learn what spatial locations or visual features are more frequently relevant in a task and can exploit these regularities to improve task performance. In the present study, we tested how incidental learning of statistical regularities influences neural processing. By measuring the brain's responses with high temporal resolution, we found that incidental learning changes patterns of neural activity in early processing stages, accelerates subsequent attentional selection, and modulates later memory-related activity. These neural changes were stronger in participants who learned more effectively. Overall, this demonstrates that incidental experience can shape the flow of information across multiple stages of neural processing.
Introduction
Statistical regularities in the environment influence how our brains process and prioritize information in the future. For example, participants are faster at finding a target in a visual search task when the configuration of the display is repeated (Chun and Jiang, 1998; Jiang and Chun, 2001) or when targets appear more frequently at a particular location (Shaw and Shaw, 1977; Miller, 1988; Hoffmann and Kunde, 1999; Jiang et al., 2014; Addleman et al., 2018, 2019; for similar effects with distractors, Wang and Theeuwes, 2018). Participants learn these statistical regularities quickly and can form long-lasting biases in what to prioritize, often without being aware of them (for a review, see Jiang and Sisk, 2019). Participants not only learn about spatial regularities in the environment, but also about nonspatial visual regularities, for example, the frequency of target and distractor colors during visual search (Sha et al., 2017; Witkowski and Geng, 2022; Addleman and Störmer, 2023).
To date, the majority of this research has relied on behavioral responses, leaving open the question of what mechanisms support incidentally learned selection biases, in particular for nonspatial features. One possibility is that the observed behavioral effects are due to similar processes as those involved during endogenous or exogenous attention, where explicit task goals and stimulus saliency drive attentional prioritization, respectively. Indeed, theoretical accounts have proposed that an individual’s selection history contributes to a common priority map that integrates goals, salience, and recent experience (Awh et al., 2012; Theeuwes, 2019). If this were the case, the neural effects due to incidental learning should mirror those effects commonly observed during exogenous or endogenous attention, such as changes in the amplitude or speed of target processing at early processing stages (Anllo-Vento and Hillyard, 1996; Müller et al., 2006; Störmer et al., 2009; Foster et al., 2020). Another possibility is that incidental learning of statistical regularities operates at later decision and response stages, for example, through lowering decision thresholds or decreasing decision and/or response uncertainty for frequently experienced features (Huang et al., 2004; Huang and Pashler, 2005; Wang et al., 2023). Behavioral measures are poorly suited for disambiguating these competing hypotheses, as response times would be accelerated for frequently experienced target features in both cases.
To fill this gap and resolve what mechanisms underlie incidental learning of frequent target features, we recorded the brain’s electrophysiological responses to investigate whether incidental learning influences early perceptual and selection processes. We used a visual search task in which, unbeknownst to the participants, target items, relative to nontarget items, appeared more frequently in one specific color. Our analyses focused on two well-established event-related potential (ERP) components: the N2pc, a negative deflection over the contralateral occipital cortex with respect to the target location that emerges ∼200 ms post stimulus onset and that indexes early attentional selection and target–distractor individuation (Luck and Hillyard, 1994; Eimer, 1996; Woodman and Luck, 1999; Mazza and Caramazza, 2011), and the late positive complex (LPC), a broadly distributed sustained positivity that typically emerges at ∼500 ms post stimulus and that has been linked to postperceptual memory and decision-making processes (Woodruff, et al., 2006; Rugg and Curran, 2007; Yang et al., 2019). Furthermore, we employed time-resolved multivariate decoding techniques to examine how early the brain would show sensitivity to the statistical regularities of colors. Our results revealed an early sensitivity (∼120 ms) to the presence of the frequent target color, followed by speeded target selection (as indexed by a latency shift of the N2pc) and an increased amplitude for frequent target colors during later processing stages (indexed by the LPC). Each of these neural effects was stronger in participants who showed greater behavioral effects of learning, insinuating their functional relevance. Together, these results reveal that effective learning about relevant features shapes the flow of information across multiple stages of the processing stream.
Materials and Methods
Participants
We collected data from 45 participants (age range, 18–28; M = 20.0 years; 42 right-handed; 26 female), with the goal of having at least 30 usable participants after exclusions (based on an exclusion rate of ∼33% in a previous pilot version of the task). Data from participants were excluded from the final sample if their behavioral accuracy in the search task was below 85% or if >40% of their trials were excluded due to EEG artifacts (blinks, eye movements, muscle noise, or excessive alpha). Data from four participants were excluded due to poor overall accuracy, and data from an additional seven were excluded because of excessive EEG artifacts. An additional two participants failed to complete the entire experiment and were also excluded. This resulted in a final sample size of 32 participants. Participants received course credit or monetary compensation for participating ($20/1 h). All participants gave written informed consent and had normal or corrected-to-normal color vision, assessed with the 14-plate edition of Ishihara's Tests for Color Deficiency (Kanehara Trading Inc. Tokyo, Japan). The study was approved by the Committee for the Protection of Human Subjects at Dartmouth College.
Experimental design
Stimuli
The search display (Fig. 1a) consisted of 10 rings positioned equidistantly along an imaginary circle with a radius of 3° visual angle (all subsequent degree measurements refer to degrees visual angle), such that two rings fell on the horizontal midline, with four rings above and four rings below the midline. Each ring was 1.4° in diameter with a 0.19° line thickness and contained a small gap (0.22°) at either its top or bottom. One ring in each display, the target ring, was differentiated by instead having a gap on its left or right side. A fixation stimulus consisting of an inner dot surrounded by four quarter-circle wedges that formed a crosshair shape (“bullseye and crosshairs”; Thaler et al., 2013) was presented in the center of the screen (inner dot diameter, 0.15°; outer wedges and crosshairs diameter, 0.76°) in a gray approximately equiluminant to the search stimuli.
Task design and measurement of N2pc component. a, Stimulus display, probability manipulation, and color wheel used in the experiment. Colors were always 120° apart on the color wheel, but the specific colors used varied between participants (see Materials and Methods). b, Grand average ERPs elicited by the search task averaged across all conditions. The red and blue waveforms correspond to activity recorded over the hemisphere contralateral and ipsilateral with respect to the target's location, respectively, with the shaded region indicating our main measure of interest, the N2pc difference, obtained by subtracting contralateral minus ipsilateral activity. c, Grand average N2pc difference waves elicited by each decile of trials sorted by response time. Colored tick marks indicate the quantified latency of the N2pc using the half area under the curve method for the correspondingly colored waveform. Faster trials produce more peaked N2pc waveforms with sharper onsets, whereas slower trials produce more gradual and sustained waveforms. d, Raincloud plots (Allen et al., 2021) depicting the response time distributions of all trials across participants included in each ERP depicted in c. Black-bordered circles indicate the mean response time of each subset of trials. e, Correlation between quantified N2pc latencies depicted in c and mean response times plotted in d for each decile of trials. The point marked with an asterisk corresponds to the slowest 10% of trials and illustrates how quantification of the N2pc becomes unreliable for slower and more variable trials.
To maximize our ability to measure the N2pc ERP component, whose amplitude is substantially reduced for stimuli in the upper visual field (Luck et al., 1997; Bacigalupo and Luck, 2019), the top two rings in the search display contained no gap and were presented in an equiluminant gray to the search item colors, serving only as placeholders to maintain the symmetry of the display.
Probability manipulation
To enable our primary manipulation of varying the target color probability, the stimulus rings could appear in three different colors across the experiment, with these colors being spaced 120° apart on an approximately equiluminant 360°CIE L*a*b color wheel (Suchow et al., 2013). The specific colors we used were counterbalanced across participants, such that the color space was sampled every 36°, resulting in 10 sets of colors used in total across the study.
On each trial, the rings could appear in two colors, with four of the rings appearing in one color and the remaining four in the other, balanced such that an equal number of rings of each color appeared in the left and right visual hemifields, with locations within each hemifield being chosen at random. Critically, the target ring (left-or-right gap ring) was more likely to appear in one color (the frequent target color) than in the other two “standard” colors.
This resulted in three trial types: “valid” trials (60% of total) where the target ring was one of the rings that was presented in the frequent target color, and the other half of rings were presented in one of the two standard colors; “neutral” trials (20% of total) where all rings were presented in one of the two standard colors; and “invalid” trials (20% of total) where the target ring was one of the items presented in one of the standard colors but was paired alongside rings presented in the frequent target color. Thus, when the target color appeared on the screen (on the valid and invalid trials), it was 75% likely that the target ring was one of the frequent target colored items.
Experimental procedure
Each trial began with the presentation of the fixation stimulus for a variable intertrial interval of 500–900 ms, after which the search display appeared for 600 ms. Participants were instructed to locate the left/right gap target ring and report the gap direction (left vs right) as quickly as possible without sacrificing accuracy and made their responses using the “left” and “down” arrow keys, respectively. Notably, given that each search display contained two differently colored sets of rings that were equally distributed across the display (Fig. 1a), locating the target could not be achieved based on a single feature but required this difficult gap discrimination task. Participants were instructed to maintain fixation and to use their peripheral vision to perform the search task. Two hundred milliseconds after the participant’s response (or 200 ms after the stimulus display disappeared, on trials where participants responded within 600 ms), the fixation stimulus dimmed slightly for 200 ms, signaling the end of the trial and that it was safe to blink prior to the next trial, which began automatically following the 500–900 ms intertrial interval. Participants completed 1,440 trials in the main task (864 valid trials, 288 invalid, and 288 neutral trials), with target color probabilities balanced every 240 trials. Participants were given a self-paced break after every mini-block of 30 trials and received feedback about their mean response time and accuracy across the previous mini-block during these breaks. We recorded participants’ brain activity via electroencephalography (EEG) to assess neural activity related to the visual search task (see below, Electrophysiological recordings).
Prior to the main experiment, participants completed a fixation training followed by a task practice. For the task practice, they performed four blocks of 30 trials each. During this practice, the search display appeared initially for 1,000 ms and linearly decreased in duration across the 120 practice trials such that it reached the real experimental display duration of 600 ms by the end of the fourth practice block. At this point, participants were informed that they could take an additional block of practice if their accuracy in the gap direction detection task was below 85% or if they were not yet comfortable performing the search task while maintaining fixation.
The fixation training paradigm was modeled after the method presented by Guzman-Martinez et al. (2009). Participants viewed a rapidly alternating flicker display that appears as a static gray patch when fixation is maintained, and visibly flickers when eye movements are made, until they were comfortable maintaining fixation, and then repeated this procedure viewing the same stimulus while brightly colored flashing rectangles appeared intermittently in the periphery, in order to practice maintaining central fixation in the presence of salient flashing stimuli. They then viewed two additional variants of the display, one where a single search ring would appear superimposed over the flicker stimulus and another with an entire search array superimposed in which they searched for the target stimulus, while the flicker or stability of the stimulus served as immediate feedback whether they were completing the task while maintaining fixation. Participants viewed each of these displays until they correctly identified 10 targets.
Electrophysiological recordings
The brain’s electrical activity was recorded continuously from 32 active Ag/Cl electrodes mounted in an elastic cap and arranged according to the international 10–20 system. Signals were amplified by an ActiCHamp amplifier (Brain Products GmbH), bandpass filtered from 0.01 to 112.5 Hz, and digitized at 500 Hz. An electrode attached to the right mastoid served as the reference during recording. Eye movements were monitored using horizontal electrooculogram (HEOG) activity recorded from two additional electrodes placed on the external ocular canthi, which were grounded by an electrode placed on the neck of the participant.
Statistical analysis
Behavior analysis
Our primary behavioral analysis focused on response times for correct trials. We first excluded trials with excessively short (<300 ms) or long (>3,000 ms) response times (0.2% of trials on average) and additionally excluded outlier trials for each participant, defined as those greater than three standard deviations from their mean response time (1.1% of trials on average). To assess the impact of the frequency manipulation on task performance, participants’ mean response times for correct trials within each validity condition were submitted to a repeated-measures ANOVA, with follow-up pairwise dependent t tests conducted to assess differences between individual experimental conditions. Though we did not necessarily expect differences in accuracy, for completeness, we performed a similar secondary analysis on participants’ average accuracy scores across conditions.
EEG analysis
EEG data were processed using the EEGLAB and ERPLAB toolboxes (Delorme and Makeig, 2004; Lopez-Calderon and Luck, 2014) in MATLAB (The MathWorks, Inc.) and custom-written scripts. Recordings were rereferenced to the average of the left and right mastoids and low-pass filtered at 25 Hz with a 12 dB/octave roll-off. Individual trials were segmented and time-locked to the onset of the search display and were baseline corrected using the window from −200 to 0 ms prestimulus.
Trials containing artifacts (e.g., blinks, eye movements, or excessive muscle/general noise) were rejected semiautomatically using peak-to-peak, square wave, and absolute amplitude cutoff functions as implemented in ERPLAB, respectively. On average 12.83% of trials were rejected due to artifacts among included participants. We further excluded trials where subjects incorrectly identified the target stimulus (6.4% of trials on average) for a total of 1,171.3 trials per participant left for analysis on average.
Quantification of ERP components
For each participant, average ERP waveforms were created using the artifact-free data. Then, to isolate the lateralized N2pc component, a neural marker of early attentional selection (Luck and Hillyard, 1994), left-target and right-target trials were averaged separately and then collapsed across visual hemifield (left, right) and hemisphere of recording (left, right) to obtain waveforms recorded contralaterally and ipsilaterally with respect to the target location. Next, we calculated the contralateral-minus-ipsilateral difference waveforms averaged across parietal-occipital electrodes PO7/8 and P7/8. For the N2pc component, we examined its latency as our primary dependent measure, based on a recent finding showing a latency shift of the N2pc following explicit spatial cues during a search task (Foster et al., 2020). We reasoned that if incidental learning of visual features operates like endogenous attention, we should also observe a similar latency effect.
Importantly, our visual search task differed from search tasks more commonly used to assess the N2pc, which necessitated a nonstandard measurement approach. In contrast to typical tasks where the target item is often defined by at least one unique feature (e.g., popout search, or a uniquely shaped target in a color additional singleton paradigm), our task required a difficult gap direction discrimination in order to locate the target, which we hypothesized would lead to more temporally variable attentional selection of the target across trials (and hence more variable response times and temporally extended N2pcs) compared with tasks where a unique or more salient feature can consistently guide attentional selection (see Fig. 1b for validation of this hypothesis, and Kuo et al., 2009 for a similar extended N2pc in a serial shape search task).
To ensure that our neural measure would capture the meaningful behavioral variability observed in the RT distributions that was present due to the serial natural of our search task (Figs. 1d, 2a), we quantified the latency of the N2pc using the fractional area under the curve method with a 50% area criterion (a highly reliable and statistically powerful measure of latency; Kiesel et al., 2008), using a rather long measurement window from 175 ms to 600 ms (the offset of our stimulus display). Because trials with extremely slow RTs may not contain a temporally consistent signal to give rise to robust and meaningfully interpretable ERP components (especially those that are relatively transient, like the N2pc), we additionally excluded the overall slowest one third of trials (within each validity condition for each participant) for our primary analyses of learning effects on the N2pc and LPC (cf. Dowdall et al., 2012 for similar exclusion of slow trials in analysis of the N2pc).
In addition to our primary analysis of N2pc latency, we also examined the amplitude of the N2pc, as recent work found amplitude changes resulting from spatial learning during search (Duncan, et al., 2023). Following prior work, we quantified the N2pc amplitude using the mean amplitude in a 100 ms window centered around the peak of the N2pc in the grand average waveform, collapsed across all conditions, resulting in a measurement window of 282–382 ms (cf. Duncan, et al., 2023).
In addition to the N2pc component, we also assessed the amplitude of the LPC (late positive complex) component, a later nonlateralized positivity whose amplitude has been implicated in postperceptual processes related to memory judgments and decision-making (Woodruff et al., 2006; Yang et al., 2019). Following previous work examining the LPC in a feature learning task, we calculated the mean LPC amplitude across electrodes P3/4, P7/8, PO3/4, and PO7/8 during the predefined time window of 400–600 ms (cf. Wang et al., 2023).
For our statistical analyses, we used a jackknife ANOVA procedure to assess latency and amplitude differences in the N2pc with appropriate corrections to our statistical tests (Ulrich and Miller, 2001) and performed a standard repeated-measures ANOVA on subjects’ mean amplitude across the window of interest for analysis of the LPC.
Validation of the N2pc latency measure
To validate our approach for measuring N2pc latency, we first examined the relationship between the N2pc and response times across all trials in our data set, similar to previous research (Jannati et al., 2013, Tay et al., 2019). We sorted each participant's trials, irrespective of validity condition, and excluding only trials with excessively long or slow response times as defined in our behavioral analysis, from fastest to slowest response time. We then plotted the grand average N2pc associated with each decile of trials sorted by response time (Fig. 1c), along with a raincloud plot showing the response time distribution of all trials across participants that contributed to each grand average waveform (Fig. 1d). This analysis resulted in a strong correlation (r = 0.88) between N2pc latency and response times, and the general relationship is visually apparent with fast trials in the earlier deciles (depicted in red) showing a more peaked waveform with an earlier upslope and downslope compared with the flatter and more temporally extended waveforms from slower and more variable trials from the later deciles (depicted in blue). Thus, the latency of the N2pc as assessed appears to be a strong neural marker of attentional selection that directly relates to response speed.
Multivariate decoding analysis
We additionally used multivariate ERP-based decoding methods as reported by Bae and Luck (2018) to test for the presence of condition differences beyond the N2pc and LPC within participants that may not be visible in the grand-averaged univariate signal. Specifically, we trained three separate support vector machine (SVM) classifiers to distinguish between valid versus neutral, invalid versus neutral, and valid versus invalid trial conditions using the data across all electrodes, excluding the HEOG channel. The same trials as used in our primary EEG analysis were used for decoding analysis, with these data low-pass filtered at 20 Hz and downsampled to 250 Hz and with decoding performed independently at each time point (every 4 ms). We performed 100 iterations of random assignment of each participant’s trials into six blocks, each of which were averaged to create individual ERP waveforms and then submitted to cross-validated training and testing, using a leave-one-out approach. Trial numbers were balanced across valid, invalid, and neutral conditions per participant based on the condition with the fewest available trials, to assure that differences in signal-to-noise ratio did not produce spurious above-chance decoding performance. This assignment scheme resulted in ∼30 trials per ERP used for training and testing per participant on average. Decoding accuracy was then averaged across participants and smoothed using a five-time point moving average for final analysis.
To statistically assess decoding performance, we used a cluster-based permutation method as described in Bae and Luck (2018). In brief, we first simulated the accuracy of a random decoder that had no sensitivity to identify the stimulus conditions by randomly and independently shuffling the true condition labels at each time point, across all blocks and iterations of the decoding procedure. These random data were smoothed with a moving average across five time points and were averaged across blocks and iterations to obtain a mean simulated decoding accuracy at each time point, with this procedure repeated once for each participant. One-sample t tests were used at each time point to compare this group-level simulated decoding accuracy to chance (50%) performance, and clusters of adjacent significant time points (p < 0.05) were identified and their t-values summed into a cluster-level t mass. We then constructed a null distribution of the cluster t masses that would be expected from a decoder operating at chance performance by repeating this procedure 10,000 times for each condition, allowing us to quantify p values to a resolution of p = 0.0001.
We then identified clusters of above-chance decoding accuracy in our observed data and calculated their cluster t masses in the same fashion. Identified clusters were deemed significant if their t masses fell above the 95th percentile of the null distribution constructed from the permutation procedure (average critical t mass across conditions = 16.432; one-tailed α = 0.05).
Results
Impact of feature probabilities on visual search response times
Overall accuracy in the search task was high at 93.6% (SD = 2.6%) and the mean response time across participants was 677.3 ms (SD = 116.6 ms). Our main hypothesis was that participants would incidentally learn the color regularities, leading to faster response times for targets appearing in the frequent color. As expected, a repeated-measures ANOVA showed reliable differences between conditions (F(2,62) = 34.88, p = 7.11 × 10−11). Planned pairwise comparisons showed that participants responded significantly faster to trials with targets presented in the more frequent color (valid trials; M = 663.06 ms, SD = 118.48 ms) compared with trials with distractors presented in the more frequent color (invalid trials; M = 714.82 ms, SD = 117.37 ms, p = 3.34 × 10−11, Cohen's d = 0.46) and trials in which the frequent color was absent (neutral trials; M = 684.77 ms, SD = 118.1 ms, p = 0.0025, Cohen’s d = 0.20), with these neutral trials additionally being significantly faster than invalid trials (p = 2.78 × 10−5, Cohen’s d = 0.25; Fig. 2a).
Effects of learning across all participants. a, Behavioral results. The raincloud plot depicts the distribution of response times for all trials across participants in the experiment separately for valid (red), neutral (light blue), and invalid (dark blue) conditions. The box plots below depict the distribution of subjects' mean RTs across conditions, with large circles indicating the overall mean RT across subjects. Asterisks indicate a significant difference between conditions. b, ERP results. N2pc and LPC waveforms plotted by condition. The colored tick marks below the N2pc indicate the quantified latency using the half area under the curve method for each condition, measured from 175 to 600 ms. Scalp maps depict the overall topography of the N2pc averaged across conditions in the windows from 200 to 400 ms and 400 to 600 ms for the N2pc and from 400 to 600 ms for the LPC.
Similarly, we observed a significant effect of the probability manipulation on response accuracy (F(2,62) = 15.37, p = 3.79 × 10−6), with pairwise comparisons indicating that invalid trials (M = 91.9%) were significantly less accurate than both valid (M = 94.3%, p = 2.29 × 10−6, Cohen's d = 0.60) and neutral trials (M = 93.3%, p = 0.0037, Cohen's d = 0.36), but with no significant difference in accuracy between valid and neutral trials (p = 0.09, Cohen’s d = 0.26). Thus, participants showed sensitivity to the color probability manipulation, resulting in faster and more accurate responses for valid relative to invalid trials.
EEG results: impact of learned feature probabilities on the N2pc and LPC
Incidental learning of the frequent target color led to earlier N2pc latencies for valid trials where the target appeared in the frequent color (Fig. 2b). The N2pc latencies (quantified using the fractional area latencies of the grand average N2pcs; Kiesel et al., 2008) were 352 ms for valid trials [jackknife standard error (jSE) = 13.49 ms], 368 ms for neutral trials (jSE = 14.26), and 386 ms for invalid trials (jSE = 9.23). A repeated-measures jackknife ANOVA (Ulrich and Miller, 2001) indicated a statistically significant difference between our conditions (F(2,62)corrected = 3.70, p = 0.030); subsequent pairwise comparisons revealed that this result was driven by a significant difference between our valid and invalid condition (p = 0.023) with differences between valid and neutral conditions (p = 0.413) and neutral and invalid trials (p = 0.328) failing to reach significance. This N2pc effect shows that a target was selected approximately 30 ms faster when it appeared in the frequent color relative to when only distractors appeared in the frequent color, indicating that early attentional selection was sped up by incidental learning.
In contrast to latency, a repeated-measures jackknife ANOVA indicated that the amplitude of the N2pc amplitude was unreliably modulated by incidental learning (F(2,62)corrected = 2.366, p = 0.102). Exploratory post hoc t tests suggested a greater amplitude N2pc for valid (M = −1.49 µv, jSE = 0.19 µv) compared with both invalid (M = −1.26 µv, jSE = 0.20 µv, puncorrected = 0.018) and neutral trials (M = −1.34 µv, jSE = 0.20 µv, puncorrected = 0.047), with no suggested difference between neutral and invalid trials (puncorrected = 0.723). These results suggest that amplitude modulations of the N2pc in response to incidental learning may be less sensitive or more variable across participants than latency effects.
The later nonlateralized LPC amplitudes significantly differed across conditions (F(2,62) = 9.27, p = 3.00 × 10−4), with valid trials (M = 3.39 µv, SD = 3.53 µv) showing a significantly more positive LPC than invalid (M = 2.54 µv, SD = 3.33 µv, p = 1.89 × 10−4) and neutral trials (M = 2.89 µv, SD = 3.44 µv, p = 0.038), but with no significant difference between invalid and neutral trials (p = 0.188). Thus, later memory-related processes also showed sensitivity to the overall probability structure of the task.
ERP measures track behavioral learning effect across participants
To further explore the relationship between the behavioral learning effects and the N2pc and LPC modulations, we next divided participants into “good” versus “poor” learners, using a median split according to the magnitude of their behavioral learning effect (quantified as the difference in mean response times between valid and invalid trials). Of note, this median-split was not done based on an underlying assumption that there would be two distinct subgroups of participants (i.e., participants who learned successfully vs not at all), as we clearly see a continuum in the magnitude of the learning effects across participants (Fig. 3b); the goal was merely to have enough participants to have sufficient power for this brain–behavior correlational analysis. We hypothesized that if our neural measures are a meaningful correlate of the behavioral learning effects, participants who—on average—showed a greater behavioral effect (“good learners”) would show a stronger difference in both N2pc latency and LPC amplitude relative to participants who—on average—showed smaller behavioral effects (“poor learners”). Following this split, good learners showed strong RT differences across conditions as expected (F(2,30) = 112.49, p = 1.146 × 10−14, average Cohen's d = 0.7 across pairwise comparisons) whereas poor learners showed only marginally significant differences between conditions (F(2,30) = 3.23, p = 0.054, average Cohen's d = 0.09 across pairwise comparisons). Importantly, on average, the two groups did not differ in either RT (good learners mean RT = 666.3 ms, SD = 93.9 ms; poor learners mean RT = 687.5 ms, SD = 137.6 ms; t(15) = 0.47, p = 0.64) or accuracy (good learners mean accuracy = 93.8%, SD = 2.7%; poor learners mean accuracy = 93.4%, SD = 2.6%; t(15) = 0.59, p = 0.56); thus, the critical difference between the subgroups was about the learning effects, not overall differences in performance.
Behavior, N2pc, and LPC waveforms plotted for good versus poor learners. a, Response time effects plotted separately for the good versus poor learners. b, Individual participants' learning effects, quantified as the difference between participants’ mean RTs for valid versus invalid trials. Light gray bars correspond to the good learners and dark gray to the poor learners. The red dashed line indicates the mean learning benefit across all participants. c, EEG results from median split. N2pc and LPC waveforms are depicted for each group separately, with colored tick marks below the N2pc waveforms indicating the quantified latency of the N2pc. Raincloud plots at the top depict the distribution of all trials across participants that are reflected in each ERP waveform, which were calculated using the fastest two-thirds of trials as in our primary analysis.
N2pc for good and poor learners
ERP waveforms are plotted separately for good and poor learners in Figure 3c. The fractional area latencies of the grand average N2pcs for good learners were 318 ms for valid trials (jSE = 15.30 ms), 340 ms for neutral trials (jSE = 19.21), and 378 ms for invalid trials (jSE = 11.58). A repeated-measures jackknife ANOVA indicated a statistically significant difference between our conditions (F(2,30)corrected = 7.43, p = 0.0024). Pairwise comparisons again indicated that valid trials had an earlier N2pc latency than invalid trials (p = 0.0017) with valid and neutral trials showing no significant difference (p = 0.352) and with neutral trials being marginally faster than invalid trials (p = 0.055). In strong contrast to the good learners, the fractional area latencies of the grand average N2pcs for poor learners did not differ reliably between conditions (F(2,30)corrected = 0.55, p = 0.582) and were 378 ms for valid trials (jSE = 15.61 ms), 388 ms for neutral trials (jSE = 17.04), and 392 ms for invalid trials (jSE = 7.69 ms).
No significant N2pc amplitude differences across conditions were found for either good (F(2,30)corrected = 0.57, p = 0.573) or poor learners (F(2,30)corrected = 2.1689, p = 0.132), although we note that valid trial N2pcs (good learners M = −1.29, jSE = 0.25; poor learners M = −1.69, jSE = 0.28) had numerically larger amplitudes than both neutral (good learners M = −1.17, jSE = 0.28; poor learners M = −1.51, jSE = 0.29) and invalid (good learners M = −1.20, jSE = 0.27; poor learners M = −1.31, jSE = 0.31) N2pcs for both groups, as was the case in primary analysis of all participants.
LPC for good and poor learners
We found a similar relationship between behavioral learning effects and LPC amplitudes: for good learners, amplitudes differed reliably across conditions (F(2,30) = 24.94, p = 4.17 × 10−7), with valid trials (M = 4.51 µv, SD = 2.93 µv) showing a significantly more positive LPC than invalid (M = 2.74 µv, SD = 2.97 µv, p = 2.22 × 10−7) and neutral trials (M = 3.62 µv, SD = 3.00 µv, p = 0.0033) and with neutral trials also significantly more positive than invalid trials (p = 0.0043). In contrast, there were no differences in LPC amplitudes across conditions for the group of poor learners (F(2,30) = 0.35, p = 0.71, mean LPC amplitudes = 2.26, 2.16, 2.33 µv, SDs = 3.81,3.79,3.74 µv for valid, neutral, and invalid trials, respectively).
These analyses highlight the behavioral significance of the two ERP markers: the group of participants that showed the strongest behavioral benefits due to the frequency manipulation also showed the strongest ERP effects, indicating that effective feature learning influences both early attentional selection and later memory-related processes.
Multivariate decoding reveals early sensitivity to frequent color for good learners
The ERP analysis, and in particular the split-group analysis, revealed that successful learning of the statistical structure of the task sped attentional selection as indicated by the N2pc latency shift. However, because the number of stimuli presented in each color was balanced across visual hemifields on each trial, the lateralized N2pc (which tracks hemispheric differences in visual processing) cannot reflect participants simply using the amount of color in a hemifield to shift attention to one side versus the other. Thus the N2pc occurring earlier is likely a consequence of enhanced processing of items in the learned color, which leads to faster identification of the target item’s location but may not reflect this enhanced processing directly. In this case, an earlier stage of visual processing (i.e., before the onset of the N2pc at ∼175 ms) would likely be modulated by the frequency manipulation. In our grand averaged ERPs, however, no lateralized or nonlateralized differences were visually apparent prior to the onset of the N2pc.
One possibility is that early enhancement of the target colored items may manifest idiosyncratically in the EEG signal across participants, such that no consistent difference remains when averaging across participants. To investigate this possibility, we turned to within-participant multivariate decoding methods to quantify whether information about the presence or absence of the frequent target color was present in the brain data. If the electrophysiological brain signal is sensitive to the frequent target color, we should be able to successfully decode the trial category (valid vs invalid vs neutral) on the basis of the frequent target color being present on valid and invalid trials and absent on neutral trials. Specifically, we hypothesized that binary condition decoders should successfully be able to discern valid from neutral and invalid from neutral trials with a greater degree of accuracy than they could discern valid from invalid trials (which both contain the frequent target color).
As shown in Figure 4, for the good learners (left), significant clusters of above-chance accuracy were observed both early and late in time for the valid versus neutral and invalid versus neutral comparisons. These clusters extended from ∼116 to 204 ms (tcluster = 75.79, p < 0.0001) and ∼448 to 600 ms (tcluster = 76.61, p < 0.0001) in the valid versus neutral comparison and from ∼116 to 220 ms (tcluster = 79.14, p < 0.0001) and ∼476 to 600 ms (tcluster = 68.44, p < 0.0001) in the invalid versus neutral comparison, with an additional midlatency cluster observed from ∼292 to 364 ms (tcluster = 52.18, p < 0.0001) for the valid versus neutral comparison. In contrast, when decoding between valid and invalid trials, where both trial types contain the frequent target color, only a single later period of above-chance decoding accuracy was observed from ∼348 to 600 ms (tcluster = 263.21, p < 0.0001).
Condition decoding performance for good versus poor learners. Mean accuracy of each binary condition decoder is plotted across time, with the x-axis corresponding to chance (50%) accuracy. The shaded region around each line reflects the standard error of decoding accuracy within each condition. Colored dots indicate the extent of clusters of significantly above-chance decoding performance, as determined via cluster-based permutation testing.
The poor learners showed a starkly different pattern of results, with only three significant and less extensive clusters observed across all decoding conditions, all occurring for the valid versus neutral comparison (early cluster from ∼160 and 196 ms, tcluster = 19.98, p = 0.0184, middle cluster ∼316 to 348 ms, tcluster = 20.72, p = 0.0145, late cluster ∼488 to 516 ms, tcluster = 18.55, p = 0.0275) and with no significant ability to decode between invalid and neutral or valid and invalid trials (Fig. 4, right).
These results indicate that the good learners displayed much greater sensitivity to the presence of the frequent target color early in processing compared with poor learners. While cluster-based significance tests cannot establish the precise latency of an observed effect (Sassenhagen and Draschkow, 2019), we note that the early period of significant decoding beginning at ∼120 ms and peaking at ∼140 ms is of particular interest, given that it occurs immediately prior to the earliest visible deflection of the N2pc waveform between ∼150 and 175 ms, suggesting that this greater sensitivity to frequent target colored items may be facilitating their processing and subsequent selection.
Discussion
Behavioral studies have shown that response times are faster for frequently relative to infrequently occurring target features during visual search (Sha et al., 2017; Addleman and Störmer, 2023); however, because response times reflect the accumulation of multiple processing steps, it is unclear which stages are influenced by these feature regularities. Using EEG, we found that incidental learning of relevant feature values is accompanied by early neural sensitivity to the presence of the frequent target color, followed by speeded target selection and differential later-stage neural amplitude modulations. Critically, by splitting participants into “good” and “poor” learners based on their average response time benefits, we found that these neural modulations across all three stages were related to the magnitude of behavioral effects, signifying their functional relevance. Overall, our results suggest that incidental learning of feature regularities can have effects across multiple processing stages, similar to what has been shown for tasks using explicit attention cues (Serences and Kastner, 2014).
Using multivariate decoding analysis, we found that the pattern of electrical activity across the scalp distinguished between search displays that contained the more frequent relative to infrequent target colors beginning at ∼120 ms post stimulus onset. This early electrical signature did not differentiate between trials in which the frequent color was a target (valid trials) or distractor (invalid trials), indicating that it tracked the general presence of the learned feature, but not whether that feature was associated with the search target (the left/right gap stimulus). Importantly, participants who showed strong learning effects behaviorally were also those that showed this neural differentiation most clearly. Thus, this early emerging sensitivity to the more relevant target color demarcates the first stage of processing impacted by learning and suggests that recently experienced statistical regularities can alter even the early flow of visual information processing. This is consistent with accounts proposing that learned features are upweighted at early sensory stages, possibly even proactively, which has been suggested in search paradigms with spatial regularities (Wang et al., 2019).
The early multivariate signal was followed by a speeded N2pc component, a well-established ERP marker associated with target selection and target–distractor individuation (Luck and Hillyard, 1994; Mazza and Caramazza, 2011), for targets that appeared in the more frequent compared with less frequent colors. Importantly, just like the multivariate signal, the N2pc latency shift was more pronounced for participants showing larger behavioral effects. Importantly, this latency shift cannot be explained by simple left versus right shifts of spatial attention, given our balanced task display always contained an equal number of items of each color in each visual hemifield. Yet this latency shift resembles a recently reported effect in a visual search task in which participants were endogenously cued to direct spatial attention to the likely target location prior to the search display onset using a symbolic cue (Foster et al., 2020). The fact that we find a similar latency shift in a task where participants are not provided with explicit spatial attention cues and instead learn incidentally about relevant nonspatial visual features suggests at least some overlap between the effects of incidentally learned and explicit goal-driven attention on search processes.
Another recent study in which participants learned the likely location of targets during visual search reported amplitude modulations of the N2pc (Duncan et al., 2023). While we observed a similar numerical trend, in our study, these amplitude modulations were not statistically reliable and did not track behavioral performance (see also Wang et al., 2023, who found unreliable N2pc amplitude effects in a feature learning task). It is unclear whether these differences in N2pc modulations (amplitude vs latency) relate to potentially distinct mechanisms between spatial and feature-based incidental learning. In our view, however, changes in the N2pc latency offer an interpretability advantage compared with amplitude measures, given that earlier attentional selection of a target should straightforwardly lead to earlier responses to that target. In contrast, a change in the amplitude of the selection process lacks clear mechanistic implications for behavior, especially considering the many factors that are known to impact N2pc amplitude (Luck et al., 1997, Poncet et al., 2016, Drisdelle and Jolicoeur, 2018). Our N2pc results relating its latency to response times across conditions (Fig. 1c–e) clearly reveal their precise relationship. The strong correlation suggests that RT differences between experimental conditions should frequently be accompanied by N2pc latency differences when attention is manipulated (as opposed to manipulations impacting decision and response-related factors) and highlights the utility of trial sorting in obtaining reliable measures of the N2pc.
The amplitude of a later component, the LPC, showed sensitivity to the probability structure of the task. It remains to be determined to what extent the LPC modulations reflect a distinct process related to target learning or are instead the reflection of downstream consequences of the earlier neural effects. Prior work has implicated the LPC in recognition memory, tracking details about prior stimulus encounters (Woodruff et al., 2006; Rugg and Curran, 2007), as well as making judgments about these memories (Yang et al., 2019) and people’s confidence in them (Brezis et al., 2017). This may suggest that, in addition to speeding selection of the target itself, memory representations are updated more effectively according to the learned probability structure. Our LPC effects could also be interpreted in the framework of a “template matching” process in decision-making, where targets sharing the incidentally learned feature may generate a stronger match signal (Yu et al., 2022a,b). It is also worth noting that the LPC modulations—across conditions and between participants—are strikingly similar to the later effects in our multivariate decoding analysis that showed differentiation between valid and invalid trials around the same time period as the LPC (∼400–600 ms). Our multivariate signal at this later time point is thus likely driven (at least in part) by the observed univariate differences in the LPC, with both potentially reflecting sensitivity to violations of the probability structure associated with the frequent target color.
The observed amplitude modulations of the LPC resemble the findings reported in a recent study that examined effects of feature learning during visual search (Wang et al., 2023). However, that study did not find reliable differences in the N2pc component (neither amplitude nor latency) and thus concluded that only later processing stages are affected by feature regularities (Wang et al., 2023). The discrepancy between our and their results may be due to several differences in experimental design, such as the complexity of the feature regularities employed (60% frequent target color trials in our paradigm vs a maximum of 33% for their most frequent color, and with less differentiation compared with other colors), the colors used (three approximately equiluminant vs eight highly distinct), and the search display itself (eight items in two colors vs 12 items in more unique colors, with the target item always being uniquely colored). It is possible that learning was less efficient or more variable (due to needing to learn a complex set of statistics as opposed to a single frequent feature) and/or less effective at guiding attentional selection to the learned feature (e.g., by needing to maintain a complex attentional set of multiple likely colors, rather than one) in their paradigm. Another possibility is that naturally occurring between-subject variability—like we observed in our data, where not all participants show the same sensitivity to the color probability structure—could obscure learning effects when not accounted for. Thus, in future research it may be critical to explore individual differences in the magnitude of the learning effects and their relation to the neural effects.
In our own analysis of between-subject differences for “good” versus “poor” learners, it is important to note that we divided participants into these subgroups in order to gain sufficient power to explore neural–behavioral relationships and not because we think there exists a theoretically meaningful binary distinction between “good” and “poor” learners, for example, due to some participants being aware versus unaware of the probability manipulation. While we did not systematically collect awareness ratings in this study, based on anecdotal evidence from debriefing subjects (where few participants guessed that color was in any way relevant to the purpose of the study) and other data in our lab in similar experiments (Addleman & Störmer, unpublished observations), there is no evidence to suggest that these subgroups differed in their awareness or that awareness is a primary driver of the incidental learning effects in our experiments.
In conclusion, we uncovered a series of neural markers that closely track the behavioral effects of learned feature prioritization. Our results are generally consistent with theoretical accounts that assume incidentally experienced selection history, in addition to explicit task goals and bottom-up saliency of a stimulus, jointly determine the weights in priority maps that integrate across both spatial and nonspatial feature information, and that form the basis for attentional guidance and selection (Bisley and Goldberg, 2010; Theeuwes, 2019). Overall, our results contribute to a growing literature characterizing the remarkable ability of the human brain to exploit statistical structures in the environment to support efficient information processing.
Data Availability
Experimental code and data are available at https://osf.io/bdwhc/.
Footnotes
- Received March 26, 2025.
- Revision received July 17, 2025.
- Accepted July 25, 2025.
-
The authors declare no competing financial interests.
- Correspondence should be addressed to Viola S. Störmer at Viola.S.Stoermer{at}dartmouth.edu.
- Copyright © 2025 the authors
