Local response heterogeneity indexes experience-based neural differentiation in reading

Abstract

The ability to read requires learning letter-string representations whose neural codes would be expected to vary depending on the amount of experience that an individual has with reading them. Motivated by sparse coding theories (e.g., Rolls and Tovee, 1995; Olshausen and Field, 1996), recent work has demonstrated that better-learned relative to less well-learned neural representations are associated with more strongly differentiated, locally heterogeneous blood oxygenation level dependent (BOLD) responses (e.g., Jiang et al., 2013). Here we report a novel analysis method we call local heterogeneity regression (Local-Hreg) that quantifies the cross-voxel heterogeneity of BOLD responses, thereby providing a sensitive and methodologically flexible method for quantifying the local neural differentiation of neural representations. In a study of literate adults, we applied Local-Hreg to fMRI data obtained when participants read letter strings that varied in their frequency of occurrence in the written language. Consistent with previous research identifying the left ventral occipitotemporal cortex (vOTC) as a key site for orthographic representation in reading and spelling, we found that the cross-voxel heterogeneity of neural responses in this region varies according to the frequency with which the written letter strings have been experienced. This work provides a novel approach for examining the local differentiation of neural representations, and demonstrates that well-learned words have greater representational differentiation than less well-learned or unfamiliar words.

Keywords: Fusiform gyrus; Hcorr; Heterogeneity; Hreg; Orthographic; Reading; VWFA.

Copyright © 2018 Elsevier Inc. All rights reserved.

Figures

Fig. 1.

Functionally defined Orthographic Network based on the Standard GLM contrast (HF, MF, LF, PW, CS) > Fixation projected onto a smoothed inflated surface rendering using BrainNet viewer (Xia et al., 2013); voxel-wise corrected p

Fig. 2.

Results obtained from the Parametric Linear Analyses within the Orthographic Network (Fig. 1) projected onto left hemisphere ventral surfaces using BrainNet viewer (Xia et al., 2013). (A) Results of standard GLM HF > MF > LF > PW > CS linear relationship (voxel-wise p MF > LF > PW > CS linear relationship (voxel-wise p

Fig. 3.

Results from the Local-Hreg analysis of HF vs. LF and HF vs. PW conditions within the Orthographic Network, projected onto left hemisphere ventral surfaces using BrainNet viewer (Xia et al., 2013). (A) HF > LF words (MNI peak = —53, —45, —23; cluster size = 187 voxels; corrected p = 0.04). (B) HF words > PWs (MNI peak = —50, —44, —23; cluster size = 98 voxels; voxels; corrected p = 0.09).

Fig. 4.

Direct comparison of Local-Hreg with Hcorr and with standard GLM analysis. A-C results constrained inclusively by a left vOTC anatomical mask (green outline). Statistical maps are presented on left hemisphere only axial slices (z-plane listed to the left); voxel wise p LF (top) and HF > PW (bottom). (B) Local-Hreg clusters (Analysis 2) clusters for the following contrasts: HF> LF (top) and HF > PW (bottom). (C) Hcorr clusters (Analysis 3) for the contrasts: HF > LF (top) and HF > PW (bottom). (D) Scatter plot with trend-line of Local-Hreg and GLM β estimate extracted for each participant from the intersecting clusters in (A) and (B); each dot is a participant. Correlation coefficients and p-values are reported in the upper right. (E) Scatter plots with trend-lines of average Local-Hreg and Hcorr values extracted for each participant from the intersecting clusters in (B) and (C); each dot is a participant. Correlation coefficients and p-values are reported in the upper right. Overall, these data indicate that Local-Hreg is a robust measure of local heterogeneity, and Local-Hreg measures are related to Hcorr values, but not to the standard GLM β estimate values.