More variability and inaccuracy during initial category learning in ASC compared to NTASC participants are less accurate at the initial stage of category learning

All participants successfully performed the categorization training. The average accuracy across the whole training was above 65% for each participant. Only one ASC participant had an average accuracy below 50% and only completed the first blocks of training (block 1 and 2). This participant was excluded for further analyses comparing discrimination sensitivity before and after category learning (see Results Sect. 2).

Figure 2 displays the results of participants’ accuracy during the category learning. We observed a significant main effect of trial bin (F(1,2156.93) = 141.22, p < 0.001, η2 = 0.06, CI: [0.05,1.00]), block (F(2,2157.05) = 52.30, p < 0.001, η2 = 0.05, CI: [0.03,1.00]) and group (F(1,96.94) = 4.43, p = 0.04, η2 = 0.04, CI: [0.00,1.00]). No significant main effect of stimulus dimension (F(1,96.94) = 2.71, p = 0.10) was present. In addition, we observed a significant interaction effect of trial bin x block (F(2,2156.93) = 57.96, p < 0.001, η2 = 0.05, CI: [0.04,1.00]), group x block x stimulus dimension (F(2,2157.05) = 3.28, p = 0.04, η2 = 0.003, CI: [0.00,1.00]) and a marginal significant effect of group x trial bin x block x stimulus dimension (F(2,2156.93) = 2.32, p = 0.098, η2 = 0.002, CI: [0.00,1.00]). These results point to a general lower accuracy for participants with ASC but also indicate that participants’ accuracy increased across the different trials and blocks, pointing to a learning effect during training. Post-hoc testing of the later interactions effects, specifically, revealed that differences between groups (ASC and NT) were specifically present at the initial block of category learning (t(87.9)ASC−NT=-2.13, p = 0.04, d=-0.46, CI: [-0.88,-0.03]) and for the CR stimulus dimension (t(88.3)ASC−NT=-2.02, p = 0.046, d=-0.43, CI: [-0.85,-0.01]). This indicates that ASC participants were slower in learning the categories, especially for the CR dimension. No other significant interaction effects were present (all p > 0.2).

Fig. 2

Results of participants’ accuracy during the category learning. Participants’ accuracy generally increased across each block. The accuracy was lower at block 2 and 3 (compared to block 1) because of an increased categorization difficulty of these blocks due to a higher stimulus sampling around the trained category boundary. ASC participants were slower in learning the categories. Differences between group (ASC and NT) were prominent at the initial block of category learning. *Mean values are plotted with 95% confidence interval

For RTs, we found a significant main effect of trial bin (F(1,2139.76) = 124.67, p < 0.001, η2 = 0.06, CI: [0.04,1.00]) and block (F(2,2139.23) = 100.19, p < 0.001, η2 = 0.09, CI: [0.07,1.00]), in which RTs significantly decreased across trial bins and blocks for both groups. No main effect of group (F(1,82.60) = 0.51, p = 0.48) and stimulus dimension (F(1,82.60) = 0.22, p = 0.64) was found. We did find a significant interaction effect of trial bin x block (F(2,2139.30) = 88.20, p < 0.001, η2 = 0.08, CI: [0.06,1.00]), trial bin x stimulus dimension (F(2,2139.76) = 5.27 p = 0.02, η2 = 0.003, CI: [0.00,1.00]), group x block (F(2,2139.23) = 4.01 p = 0.02, η2 = 0.004, CI: [0.00,1.00]) and group x stimulus dimension (F(1,82.60) = 5.94 p = 0.02, η2 = 0.07, CI: [0.01,1.00]). Post-hoc testing showed that ASC participants responded significantly faster (RTs significantly decreased) after the first block (t(2141)block1−2=5.50, p < 0.0001, d = 0.24, CI: [0.15,0.32] and t(2140)block1−3=3.88, p = 0.0003, d = 0.17, CI: [0.08,0.25]) in line with their enhanced category learning after the initial block. In addition, post-hoc testing revealed that NT participants responded faster (lower RTs) for the CR dimension compared to ASC participants (t(72.0)ASC−NT =1.81, p = 0.07, d = 0.42, CI: [-0.04,0.89]). This is in line with ASC participants’ difficulty for categorization in the CR dimension (cf. accuracy results). No other significant interaction effects were present (all p > 0.1).

ASC participants show more heterogeneity in the initial stage of category learning

When fitting a logistic function to participants’ response (i.e., proportion of category ‘A’ or ‘B’ response) across the stimulus space, we can obtain two parameters for each participant and for each training block: the position of the category boundary (i.e., threshold) and the steepness of the category boundary (i.e., slope). Figure 3 displays the results of participants’ fitted category boundary during category learning. Pertaining to the position of the category boundary, we observed a significant main effect of stimulus dimension (F(1,69.73) = 14.08, p < 0.001, η2 = 0.17, CI: [0.05,1.00]). Post-hoc testing showed a slight bias in the location of the category boundary for the AR dimension (t(70.7)AR−CR=3.75, p = 0.0004, d = 0.89, CI: [0.40,1.38], see Supplementary Material Figure S5). No significant main effects of group (F(1,69.73) = 0.62, p = 0.43) or block were present (F(2,136.48) = 0.35, p = 0.7), nor any interaction effects (all p > 0.4). For the steepness of the category boundary, we observed a significant main effect of block (F(2,139.04) = 19.51, p < 0.001, η2 = 0.22, CI: [0.12,1.00]). No significant main effects of group (F(1,70.99) = 0.48, p = 0.49) or stimulus dimension were present (F(1,70.99) = 1, p = 0.32), nor any significant interaction effects (all p > 0.2). Bayesian statistical modelling confirmed these findings (Supplementary Material Figure S3). Using the Bayesian approach, we found an effect of block (estimate: 0.04, HDI: [0.03,0.05]) and level x block (estimate: 0.01, HDI: [0.00,0.01]).

However, when we tested the variability of obtained results, we could detect a higher variability in the slopes values of the ASC participants (Fvar(111, 110) = 0.63, p = 0.02). This effect was especially driven by the initial block (Fvar(37,37) = 0.31, p = 0.0005, see Fig. 3).

Fig. 3

Results of participants’ fitted category boundary during category learning. The obtained category boundary (trained at midpoint of the stimulus dimension) in the first block did not significantly change in subsequent blocks. Participants did learn to be more precise in their assignment to the two different categories. Precision of category boundary was more variable across the different ASC participants, as evidenced by the larger variability in slope, especially in the initial block. *Mean values are plotted with participants’ individual curves

These results indicate that participants were able to quickly distinguish the shapes along the stimulus continua in two different categories. The Bayesian modeling approach points to a significant effect of block. The psychometric approach further specifies that the obtained category boundary did not significantly change in subsequent blocks, but that participants did learn to be more precise in their assignment to the two different categories (leading to a significant change in the steepness of the slope). We additionally observed significantly higher variability of the slope of the category boundary for the ASC participants, especially in the initial block. In conclusion, these results indicate that more participants with ASC are less precise/consistent in the initial phase of explicit category learning.

Discrimination sensitivity only changes after explicit category learning in ASC compared to NTBehavioral categorical perception is already induced by implicit learning in NT

Figure 4 displays the results for the behavioral discrimination task, both pre- and post-training. For d-prime, we observed a significant main effect of comparison (i.e., overall increased discrimination sensitivity for pair across the category boundary, F(1,1682.02) = 16.05, p < 0.001, η2 = 0.01, CI: [0.00, 1.00]), assessment moment (i.e., overall increased discrimination sensitivity after training; F(1,1682.02) = 77.07, p < 0.001, η2 = 0.04, CI: [0.03,1.00]), block (i.e., overall increased discrimination sensitivity across the different blocks, F(3,1682.02) = 24.81, p < 0.001, η2 = 0.04, CI: [0.03,1.00]), and group (i.e., overall lower discrimination sensitivity for ASC, F(1,73.75) = 19.37, p < 0.001, η2 = 0.21, CI: [0.09,1.00]). We did not observe a main significant effect of stimulus dimension (F(1,73.75) = 1.63, p = 0.21).

Most importantly, we observed a significant interaction effect of comparison x assessment moment x group (F(1,1682.02) = 3.93, p = 0.047, η2 = 0.002, CI: [0.00,1.00]). Post-hoc testing revealed that categorical perception (i.e., increased sensitivity for pairs that cross the category boundary compared to pairs within the category) is both present before (t(1682)between−within=3.02, p = 0.003, d = 0.15, CI: [0.05,0.24]) and after (t(1682)between−within=2.07, p = 0.04, d = 0.10, CI: [0.01,0.20]) categorization training in the NT group, while for the ASC group behavioral categorical perception is absent before training (t(1682)between−within=-0.05, p = 0.96, d=-0.003, CI: [-0.10,0.09]) and only present after explicit category learning (t(1682)between−within=2.97, p = 0.003, d = 0.14, CI: [0.05,0.24]). This implies that NT participants already implicitly pick up the underlying characteristics of the stimulus dimension, without explicit categorization training, and that this already influences their perception. ASC participants, on the other hand, are not able to implicitly incorporate the underlying dimensions of the stimuli, without explicit training (see Fig. 4, left panel). Consequently, their behavioral discrimination sensitivity is only influenced after explicit category training (see Fig. 4, right panel). Bayesian statistical modelling confirmed the main findings (Supplementary Material Figure S4). We found a clear interaction effect of comparison x assessment moment x group x stimulus dimension (estimate: 0.41, HDI: [0.04,0.78]). This interaction effect, again, indicates that NT participants implicitly picked up the underlying categorical dimension (and therefore already showed a categorical perception effect) before training in comparison to ASC who only showed a categorical perception effect after training. This effect seems to be mainly driven by the AR stimuli. This is in line with the observed marginal interaction effect of assessment moment x group x stimulus dimension (F(1,1682.02) = 2.86, p = 0.09, η2 = 0.002, CI: [0.00,1.00]) (see Supplementary Material Figure S6).

Fig. 4

Results for the behavioral discrimination task. Stimulus pairs 1–3 and 5–7 reflect within category perceptual discrimination, while stimulus pair 3–5 reflects between category perceptual discrimination. Left panel: NT participants implicitly picked up the underlying categorical dimension (and therefore already showed a categorical perception effect) before training in comparison to ASC participants. Right panel: ASC participants only showed a categorical perception effect after training. Do (visually) note a systemic bias in discrimination sensitivity towards pair 1–3 (compared to pair 5–7). This bias seems to be mainly driven by the CR dimension (see Supplementary Material Figure S6). *Mean values (lines) are plotted with 95% confidence interval and participants’ individual values (dots)

We also observed a significant interaction of block with assessment moment (F(3,1682.02) = 3.79, p = 0.01, η2 = 0.007, CI: [0.00,1.00]) and stimulus dimension (F(3,1682.02) = 3.73, p = 0.01, η2 = 0.007, CI: [0.00,1.00]) and its marginal interaction with comparison (F(3,1682.02) = 2.22, p = 0.08, η2 = 0.004, CI: [0.00,1.00]) and group (F(3,1682.02) = 2.23, p = 0.08, η2 = 0.004, CI: [0.00,1.00]). Most importantly, we observed a significant interaction between assessment moment x block x group (F(3,1682.02) = 3.25, p = 0.02, η2 = 0.006, CI: [0.00,1.00]). Post-hoc testing showed that the d-prime significantly increased across the consequent blocks, and this specifically for the NT participants in the pre-training assessment (t(1682)block2−1=4.41, p = 0.001, d = 0.20, CI: [0.10,0.30] and t(1682)block3−2=2.53, p = 0.03, d = 0.12, CI: [0.03,0.22]). This is in line with the significantly decreased RTs for the NT participants before training across the different blocks (Supplementary Material Figure S2) and could reflect the implicit learning of the NT participants. No other interaction effects were present (all p > 0.1).

Neural discrimination sensitivity changes only after explicit category learning in ASC

Figure 5 displays the results for the neural FT-EEG sweep oddball amplitudes. Using this direct neural approach, we observed a significant main effect of sweep step (F(2,741.11) = 128.98, p < 0.001, η2 = 0.26, CI: [0.21,1.00]), which is clearly represented in the increasing baseline-subtracted oddball amplitude along the dimension. We observed a marginal main effect of assessment moment (F(1,732.37) = 3.64, p = 0.06) and no effect of group (F(1,65.58) = 0.44, p = 0.51).

Most interestingly, we observed a significant three-way interaction of sweep step x assessment moment x group (F(2,732.39) = 5.08, p = 0.006, η2 = 0.01, CI: [0.00,1.00]). Post-hoc testing revealed that the neural sensitivity at the category boundary (i.e., hallmark of categorical perception) significantly increased after explicit category learning as compared to pre-training for ASC participants (t(737)post−pre=2.69, p = 0.007, d = 0.20, CI: [0.05,0.34]). This is not the case for the NT participants, who had similar levels pre- and post-training (t(736)post−pre=0.01, p = 0.99), which is in line with the behavioral discrimination data (in which NT already showed the categorical perception effect before the training).

Fig. 5

Oddball results for the neural FT-EEG sweep (averaged across both ROIs). Left panel: Analyses revealed that the neural sensitivity at the category boundary (see dashed line, i.e., hallmark of categorical perception) did not significantly differ after explicit category learning (compared to before) for NT participants. Right panel: We did find a significant increase in neural sensitivity at the category boundary after explicit category learning (compared to before) for the ASC participants. The bottom panel shows the head topographies along the 7 sweep steps. *Error bars correspond to standard errors of the mean

Finally, we observed significant main effects of ROI (F(1,732.68) = 17.96, p < 0.001, η2 = 0.02, CI: [0.01,1.00]) and stimulus dimension (F(1,65.58) = 10.38, p = 0.002, η2 = 0.14, CI: [0.03,1.00]). This indicates a higher elicited baseline-subtracted oddball amplitude along the FT-EEG sweep for the right OT (t(737)left−right=-4.24, p < 0.001, d=-0.31, CI: [-0.46,-0.17]) and AR dimension (t(69.8)AR−CR=3.22, p = 0.002, d = 0.77, CI: [0.28,1.25]). We also observed a significant interaction effect of ROI x sweep step (F(2,732.67) = 3.86, p = 0.02, η2 = 0.01, CI: [0.00,1.00]), stimulus dimension x sweep step (F(2,741.11) = 8.46, p < 0.001, η2 = 0.02, CI: [0.01,1.00]), ROI x stimulus dimension (F(1,732.68) = 6.79, p = 0.009, η2 = 0.01), assessment moment x group x stimulus dimension (F(1,732.37) = 4.88, p = 0.03, η2 = 0.007, CI: [0.00,1.00]), sweep step x ROI x stimulus dimension (F(2,732.67) = 5.44, p = 0.005, η2 = 0.01, CI: [0.00,1.00]), and a marginal interaction effect of sweep step x group x stimulus dimension (F(2,741.11) = 2.90, p = 0.06, η2 = 0.008, CI: [0.00,1.00]). Post-hoc testing (on interaction effects with group) revealed that increased discrimination sensitivity after training was mainly due to the AR dimension for the ASC participants (t(737)post−pre=2.11, p = 0.03, d = 0.16, CI: [0.01,0.30]) and that ASC participants displayed a reduced neural category tuning for the CR dimension (i.e., an enhanced neural discrimination sensitivity towards the start of the CR dimension; t(737)level2−4=-1.21, p = 0.45, d=-0.09, CI: [-0.23,0.06], see Supplementary Material Figure S6). This is in line with behavioral results which showed increased category learning of ASC participants for the AR dimension during training (compared to the CR dimension) and a bias in behavioral discrimination sensitivity (for both groups) towards the start of the CR dimension (i.e., pair 1–3, see Fig. 4). No other interaction effects were present (all p > 0.1).

Finally, the lack of group differences in terms of orthogonal color change detection task performance and base-rate synchronization amplitude during FT-EEG assessments (Supplementary Material Figure S1) confirms that there are no systematic changes in processing demands (e.g., effort or attention) or brain synchronization and that potential differences in oddball activity between groups are thus due to effectively perceived stimulus differences.

Correlations with participants’ characteristicsLearning differences are more pronounced across AQ characteristics

When we correlated accuracy of the three different categorization training blocks (averaged across the different trial bins) with questionnaire scores of the participants, we found a significant negative correlation with AQ characteristics for block 1 (r=-0.26, p = 0.02). This correlation indicates that participants with higher AQ scores performed worse in the categorization training. For the estimated threshold and slope for the fitted logistic function across the blocks, we only found a marginally significant negative correlation with AQ characteristics in block 1 for the slope values (r=-0.19, p = 0.09). This was mainly driven by the CR dimension (r=-0.30, p = 0.07). No significant correlations were found for the GSQ scores.

Differences in behavioral discrimination sensitivity are driven by AQ and GSQ characteristics

When we correlated the behavioral categorical perception effect (i.e., difference in discrimination sensitivity of pairs between versus within the category) with participants’ characteristics, we found a significant correlation with AQ scores for the pre-training assessment (r=-0.27, p = 0.02). This negative correlation seemed to be specifically driven by the AR dimension (r=-0.44, p = 0.005). This correlation indicates that participants with lower AQ scores are better able to implicitly pick up the underlying categorical dimension (and therefore already show a categorical perception effect for the dimension) before training in comparison to participants with higher AQ scores. This indicates that our main finding, that NT but not ASC participants already implicitly picked up the underlying dimensions of the stimuli without explicit training, could mainly be driven by their lower AQ characteristics. In Fig. 6, we can see that this effect is mainly driven by the NT participants (AR: r=-0.36, p = 0.02 in Fig. 6A and AR pre-training: r=-0.50, p = 0.03 in Fig. 6B). There was also a marginal correlation effect present post-training for the ASC participants (AR: r=-0.42, p = 0.07, Fig. 6B and both dimensions: r=-0.28, p = 0.09). This suggests that ASC participants with higher AQ traits showed a decreased categorical perception effect after training. We also found a significant negative correlation of the behavioral categorical perception effect with GSQ scores for the pre-training assessment (r=-0.29, p = 0.01). This was also mainly driven by the AR dimension (r=-0.37, p = 0.02).

Fig. 6

Correlations of behavioral categorical perception and participants characteristics. (a) We found a significant negative correlation of the behavioral categorical perception effect with AQ scores for the AR dimension. (b) This was specifically driven by the results before training for the NT participants. We also found a marginal significant negative correlation effect after training, specifically for the ASC participants

When we examined the overall behavioral discrimination sensitivity in general (i.e., not taking into account the position along the category dimension and the pre- versus post-training assessment), we could find a significant negative correlation with both AQ- and GSQ-scores (AQ: r=-0.44, p < 0.0001 and GSQ: r=-0.26, p = 0.02). Moreover, for the GSQ scores this correlation was mainly driven by the hypo-sensitivity sub-score (r=-0.27, p = 0.02). This is in line with overall reduced d-prime values for the ASC participants and could suggest that this effect is mainly driven by their hypo-sensitivity characteristics.

Finally, when we correlated the neural categorical perception effect (i.e., sensitivity at the category boundary before and after training) and overall neural sensitivity with AQ-, and GSQ-scores across the participants, we found no significant correlation with participant characteristics.