Reduced categorical learning of faces in dyslexia

Abstract

The perception of phonological categories in dyslexia is less refined than in typically developing (TD) individuals. Traditionally, this characteristic was considered unique to phonology, yet many studies showed non-phonological perceptual difficulties. Importantly, measuring the dynamics of cortical adaptation, associated with category acquisition, revealed a broadly distributed faster decay of cortical adaptation. Taken together, these observations suggest that the acquisition of perceptual categories in dyslexia may be slower across modalities. To test this, we tested adult individuals with developmental dyslexia (IDDs) and TDs on learning of two unknown faces, yielding face-specific categorization. Initial accuracy was similar in the two groups, yet practice-induced increase in accuracy was significantly larger in TDs. Modeling the learning process (using Drift Diffusion Model) revealed that TDs' steeper learning results from a larger increase in their effective face-specific signal. We propose that IDDs' slower item-specific categorical learning of unknown faces indicates that slower categorical learning in dyslexia is a core, domain-general difficulty.

Keywords: Drift diffusion model; Dyslexia; Face categorization; Perceptual learning.

Fig. 1

Stimuli and procedure. (A) The seven stimuli used in the experiment: Two endpoint faces of “Daphna” and “Liat”, marked as 0% and 100% according to the percent of the Liat face in them, were used in the introduction phase. Five morphed faces, marked as 30%, 40%, 50%, 60% and 70%, were used in the experimental phase. (B) Two example test trials. In each trial, participants were shown a morphed face for 1.5 sec, and were asked to determine whether the face was of Liat or of Daphna. These names were shown alongside the face (written in Hebrew in the actual experiment). Between the trials, a fixation cross was shown. The inter-trial interval (ITI) was either 2, 3.5 or 5 sec, in separate blocks.

Fig. 2

An illustration of the DDM. A simulated path depicts the noisy accumulation of evidence over time, with a drift rate v, which denotes the rate of information accumulation. The upper boundary denotes the threshold for a correct response (e.g., “Liat” for the 60% Liat morph), and the lower boundary denotes the threshold for an incorrect response. Accuracy is based on the face most similar to the stimulus, hence trials with a 50% stimulus, for which there was no correct response, were not included in the DDM modeling. The distance between the two boundaries is denoted by the threshold separation parameter – a. The starting point of the evidence accumulation path is denoted by z, and in our implementation it was fixed at a/2 (no bias). The non-decision time t0 is a single parameter which represents the duration of all non-decisional processes, both before and after the evidence accumulation process. RT is the sum of the decision time and t0.

Fig. 3

Reduced learning of face categorization in IDDs compared with TDs. (A) The mean categorization d′ of each group in each session. (B) The d′ of each participant in each session. Left – TDs. Right – IDDs. Colored bold lines connect the means. (C) Learning across sessions – the difference between d′ in the second versus the first session at the level of single participants. TDs improved while IDDs did not. Horizontal filled lines denote the mean. Horizontal dashed lines denote the median. In all graphs, values were jittered horizontally for display purposes. Error bars denote the standard error of the mean.

Fig. 4

Bias of responses to the 50% stimulus did not change in either group between sessions. (A) The percent of “Liat” responses to the ambiguous 50% stimulus in the second versus first session of each participant. Regression lines are presented for TDs (in blue) and IDDs (in red) separately. r denotes the Pearson correlation. ∗∗∗ denotes p < .001 in a two-tailed test. (B) The percent “Liat” responses to the ambiguous 50% stimulus in each session at the level of single participants. Colored bold lines connect the means. (C) The differences between the bias of responses to the ambiguous 50% face in the second versus the first session. The response bias of each participant was defined as the absolute difference between their mean percent of “Liat” responses and chance level (50%). Positive values denote an increase in bias from the first to the second session. Across groups, there was no change in response bias between sessions, suggesting that the learning effect, quantified by d′, was not due to reduced noise (enhanced response consistency for a given stimulus). Horizontal lines denote the mean. In all graphs, values were jittered horizontally for display purposes. Error bars denote the standard error of the mean.

Fig. 5

RT was longer in dyslexia throughout the experiment, and decreased similarly in the second session in the two groups. (A) The mean RT of TDs (blue) and IDDs (red) in each session. (B) The difference between RT in the second versus the first session, at the level of single participants. The decrease in RT was similar in the two groups. Horizontal lines denote the mean. Error bars denote the standard error of the mean.

Fig. 6

Greater training-induced improvement in face categorization in TDs was modeled in the DDM as a larger change in the drift rate (v) from the first to the second session. (A) The mean estimates of the drift rate (v) by session across ITIs for TDs (blue) and IDDs (red). (B) The difference between the drift rate estimates in the second versus the first session, at the level of single participants. Horizontal filled lines denote the mean. Dashed lines denote the median. Error bars denote the standard error of the mean.

Fig. 7

Increasing ITI improved categorization to a similar extent in both groups (A–B), with a tendency for a larger cross-session improvement of TDs in each ITI (C). (A) The mean d′ in each ITI for TDs (blue) and IDDs (red). (B) The difference between d′ in the 5 sec versus the 2 sec ITIs, at the level of single participants. Positive values indicate improved categorization from the shortest to the longest ITI. Horizontal lines denote the mean. (C) The mean d′ in each session for TDs (blue) and IDDs (red) in the three ITI conditions: 2 sec (left), 3.5 sec (middle) and 5 sec (right) shows similar interactions in the three ITIs. Error bars denote the standard error of the mean.

Fig. 8

RT increased with ITI in both groups, and similarly decreased across sessions in each ITI. (A) The mean RT increases with longer ITIs in both TDs (blue) and IDDs (red). (B) The difference in RT in the 5 sec versus 2 sec ITIs, at the level of single participants. Horizontal lines denote the mean. (C) The mean RT of TDs (blue) and IDDs (red) in each session, in each ITI: 2 sec (left), 3.5 sec (middle) and 5 sec (right). The group × session × ITI effect was not significant [F(1.71, 88.82) = .5, ${\mathrm{\eta }}_{\mathrm{G}}^2$ = .001, p = .577], i.e., the effect of ITI did not interact with cross-session learning of the two groups. Error bars denote the standard error of the mean.

Fig. 9

The behavioral effects of increasing the ITI are explained as an increase in both the drift rate (v) and the non-decision time (t0) in both groups (a). Left – the mean parameter estimates in each ITI for TDs (blue) and IDDs (red). Right – The difference in parameter estimates in the 5 sec versus the 2 sec ITIs, at the level of single participants. Horizontal lines denote the mean. (A) and (B) the drift rate (v); (C) and (D) the non-decision time (t0). Error bars denote the standard error of the mean.