Fig 1.
Distances from a decision boundary through activation space can be used to predict reaction times (RT).
(A) A hypothetical 2D activation space for human IT representing animate and inanimate object exemplars. Activation patterns for individual exemplars are projected onto a discriminant axis, which differentiates patterns based on animacy. A decision boundary placed along the axis allows for classification of animate and inanimate exemplars. Gaussian distributions along the discriminant axis reflect “decision noise”. Exemplar representations closer to the boundary produce more ambiguous evidence compared to exemplar representations far from the boundary. An implication of classic signal detection theory is that RTs will correlate negatively with distance from the boundary. (B) A hypothetical emergent activation space for animate vs. inanimate object exemplars as would be revealed using MEG decoding methods. Stimulus onset is the time the stimulus is presented. The decoding onset (dashed line) is the first time point that a classifier trained to discriminate between animate and inanimate examplars performs significantly above chance. Peak decoding (gray box) is the optimal time point to read out information about stimulus category. Clusters depict the hypothetical 2D activation spaces at notable points in the decoding time-course.
Fig 2.
Trial structure for the two experimental tasks. On each trial a letter (vowel or consonant) was superimposed on the fixation circle in the centre of each object exemplar. When performing the categorization task subjects judged whether the exemplar stimulus was animate or inanimate, while during the distracted viewing task subjects judged whether the letter was a vowel or consonant. After subjects responded the fixation circle briefly flashed green (correct response) or red (incorrect response or no response) to provide trial-by-trial feedback on performance.
Fig 3.
Mean classifier performance (d’) for both the categorization task (red) and distracted viewing task data (orange) plotted over time. Shaded regions above and below the mean lines indicate +/- 1 SEM across subjects. Color-coded asterisks indicate time points at which classifier performance was significantly above chance, using a Wilcoxon signed rank test (* = false discovery rate (FDR) adjusted p < 0.05). Decoding onset for both data sets (60ms) is indicated by a dashed vertical line. The period of peak decoding is indicated by a gray box extending 120–240 ms post-stimulus onset. The bar along the x-axis indicates the stimulus duration.
Fig 4.
Time-series correlation between representational distances from the animacy boundary and categorization task reaction times.
The time-varying rank-order correlation (Spearman’s ρ) between the average object exemplar representational distance and average reaction time across subjects for the categorization task (red), distracted viewing task (orange), and cross-over between the tasks (blue), in which representational distances from the distracted viewing task were correlated with the median RTs for the categorisation task. Color-coded asterisks indicate time points at which a correlation between distance and RT achieves significance (* = FDR adjusted p < 0.05). The decoding onset is indicated by dashed vertical line (60 ms). The period of peak decoding is indicated by the gray shaded region extending from 120–240 ms post-stimulus onset. The bar along the x-axis indicates the stimulus duration.
Fig 5.
Matrix of correlations between decoding time-courses and representational distance-RT correlation time-courses.
The “heat” of each tile reflects the strength of the rank-order correlation (Spearman’s ρ) between a decoding time-course and a representational distance-RT correlation time-course. Asterisks indicate significant correlations (* = p < .01; ** = p < .001).
Fig 6.
Time varying representational distance for individual exemplars separated by object category.
Time varying (A) categorization task (red) and (B) cross-over (blue) rank-order correlations (Spearman’s ρ) for animate and inanimate exemplar stimuli. The color-coded asterisks in the top row of plots indicate time points at which there is a significant correlation between distances and RTs (* = FDR adjusted p < 0.05). The bottom row of plots displays the distances from animacy decision boundaries at each time point, computed for both the categorization task and distractor task data sets. Each time-varying line depicts the representational distance for an individual exemplar stimulus, at each 20 ms time point (either animate or inanimate). The color of each line is based on the rank-order of the median RT for each exemplar (rank is always within category). In all plots, decoding onset is indicated by dashed vertical line (60 ms). The period of peak decoding is indicated by a gray box extending 120–240 ms post-stimulus onset. The bar along the x-axis indicates the stimulus duration.
Fig 7.
Time-averaged (0–600 ms) correlations between representational distance and RTs for both animate and inanimate object exemplars.
Mean time-averaged categorization (red) and cross-over (blue) correlations (Spearman’s ρ) between representational distances and RTs. Asterisks indicate significant comparisons (Wilcoxon Signed Rank Test; * = p < .01; ** = p < .001). Error bars indicate +/- 1 SEM across time-points.
Fig 8.
Correlations between sensory peak latencies and amplitudes and categorization task RTs.
Scatter plots for the rank-order (A) categorization and (B) cross-over correlations (Spearman’s ρ) between peak latency and peak amplitude and median normalized RTs for the categorization task. (A) Red circles indicate animate exemplar data points, while red rings indicate inanimate data points. (B) Blue circles indicate animate exemplar data points, while blue rings indicate inanimate data points. Peak amplitude scale is in units of 10−14 T. For comparison also plotted are the correlations between representational distance and RTs at 120 ms. Asterisks indicate significant correlations (* = p < .01; ** = p < .001).