Recurrent processes support a cascade of hierarchical decisions

  1. Laura Gwilliams  Is a corresponding author
  2. Jean-Remi King
  1. Department of Psychology, New York University, United States
  2. NYU Abu Dhabi Institute, United Arab Emirates
  3. Frankfurt Institute for Advanced Studies, Germany
  4. Laboratoire des Systèmes Perceptifs (CNRS UMR 8248), Département d’Études Cognitives, École Normale Supérieure, PSL University, France
6 figures, 2 videos and 2 additional files

Figures

Experimental protocol and behavioral results.

Experiment 1: eight human subjects provided perceptual judgments on variably ambiguous digits briefly flashed at the center of a computer screen (A). Reports were made by clicking on a disk, where …

Figure 2 with 12 supplements
Spatio-temporal hierarchy.

(A) Mass-univariate statistics. Each row plots the average-across-subjects beta coefficients obtained from regression between single-trial evoked activity and each of the five features orthogonally …

Figure 2—figure supplement 1
Exhaustive set of results for decoding different features of the stimulus, time-locked to stimulus onset.

Shaded area indicates significant decoding as confirmed with a one-sample temporal cluster test.

Figure 2—figure supplement 2
Exhaustive set of results for decoding different features of the stimulus, time-locked to motor onset.

Shaded area indicates significant decoding as confirmed with a one-sample temporal cluster test.

Figure 2—figure supplement 3
Decoding letter/digit contrast for symbols not presented in the active condition, where subjects viewed the symbols passively and did not have to make a motor response.
Figure 2—figure supplement 4
Univariate significance mask for ‘Stimulus Position’.

Localized sources that are contained in a significant (p<0.05) cluster and the specified time-points are highlighted in black.

Figure 2—figure supplement 5
Univariate significance mask for ‘Stimulus Identity’.

Localized sources that are contained in a significant (p<0.05) cluster and the specified time-points are highlighted in black.

Figure 2—figure supplement 6
Univariate significance mask for ‘Uncertainty’.

Localized sources that are contained in a significant (p<0.05) cluster and the specified time-points are highlighted in black.

Figure 2—figure supplement 7
Univariate significance mask for ‘Motor Side’.

Localized sources that are contained in a significant (p<0.05) cluster and the specified time-points are highlighted in black.

Figure 2—figure supplement 8
Univariate significance across the four significant features.

Localized sources that are contained in a significant (p<0.05) cluster. Colormap indicates for how many time-points the vertex was included in a cluster.

Figure 2—figure supplement 9
Significance of multivariate tests across the five significant features.

Localized sources that are contained in a significant (p<0.05) cluster in the spatial decoding analysis.

Figure 2—figure supplement 10
Non-corrected log-transformed (base 10) p-values for the mass univariate tests, plotted for each of the five features.

Note that each feature is has a different color-map threshold.

Figure 2—figure supplement 11
Average decoding timecourses for each of the five features.

Timing is locked to stimulus onset (above) and motor response onset (below). Unlike the analysis of Ambiguity (blue line) in the main test, here we median-split the ambiguity variable to fit a …

Figure 2—animation 1
Violin plot of decoding accuracy for the five features of interest over time.

Each dot represents a different subject. The shaded area represents the distribution density of the decoding performance scores across subjects. The dashed black line indicates chance decoding …

Source and temporal generalization predictions for various neural architectures.

(A) Four increasingly complex neural architectures compatible with the spatial and temporal decoders of Figure 2. For each model (rows), the five layers (L1, L2 … L5) generates new representations. …

Temporal generalization results.

(A) Temporal generalization for each of the five features orthogonally varying in our study. Color indicate decoding score (white = chance). Contours indicate significant decoding clusters across …

Figure 5 with 1 supplement
Correlation between TG peaks and reaction times.

(A, B) Recurrent processing at a given processing stage is hypothesized to take a variable amount of time to generate adequate representations. (C) According to this hypothesis, the rise of the …

Figure 5—figure supplement 1
Schematic of the processing delay analysis pipeline.

The top ‘empirical’ figure shows the temporal generalization matrix for decoding ambiguity split into the trials that were responded to the fastest (top 25%) and slowest (bottom 25%). The cartoon …

Figure 6 with 2 supplements
Motor and perceptual decisions.

(A) Hypothesis space for when responses become categorical: during sensory, perceptual or motor processing. (B, top) Time course of decoding the perceptual decision. (B, bottom) Classifier …

Figure 6—figure supplement 1
Decoding the distinction between the unambiguous end-points and stimuli at different step distances away.

The purple line corresponds to decoding the difference between unambiguous end-points and one step. A temporal cluster test revealed no significant above-chance decodability. The blue line …

Figure 6—figure supplement 2
Decoding performance for the letter/digit contrast in active trials requiring a button press (blue line), and passively presented non-ambiguous trials requiring no response (red line).

For the active trials, the classifier is fit with a k-fold cross-validation – training on active trials and testing on active trials. For the passive trials, the classifier is trained on all of the …

Videos

Video 1
Source-localized evoked response averaged over all trials and subjects.

Activity is plot in noise-normalized dSPM units, and shown on an inflated cortical surface (center) as well as a two-dimensional ‘glass brain’ that shows activity averaged over the transverse plane …

Video 2
Temporal decoding results.

For each regressor of interest, the trajectory of normalized decoding accuracy is plot over time. The beta coefficients from the univariate spatio-temporal analysis are plot on the inflated brains, …

Additional files

Supplementary file 1

Summary table showing the timing and significance of the results across the five features across the three statistical analyses (temporal decoding, spatial decoding and mass univariate).

pfc corresponds to the p-value after Bonferroni correction across the five features. Average t-value corresponds to the average t-value in the cluster.

https://cdn.elifesciences.org/articles/56603/elife-56603-supp1-v2.docx
Transparent reporting form
https://cdn.elifesciences.org/articles/56603/elife-56603-transrepform-v2.docx

Download links