A Hierarchical Attractor Network Model of perceptual versus intentional decision updates

Abstract

Changes of Mind are a striking example of our ability to flexibly reverse decisions and change our own actions. Previous studies largely focused on Changes of Mind in decisions about perceptual information. Here we report reversals of decisions that require integrating multiple classes of information: 1) Perceptual evidence, 2) higher-order, voluntary intentions, and 3) motor costs. In an adapted version of the random-dot motion task, participants moved to a target that matched both the external (exogenous) evidence about dot-motion direction and a preceding internally-generated (endogenous) intention about which colour to paint the dots. Movement trajectories revealed whether and when participants changed their mind about the dot-motion direction, or additionally changed their mind about which colour to choose. Our results show that decision reversals about colour intentions are less frequent in participants with stronger intentions (Exp. 1) and when motor costs of intention pursuit are lower (Exp. 2). We further show that these findings can be explained by a hierarchical, multimodal Attractor Network Model that continuously integrates higher-order voluntary intentions with perceptual evidence and motor costs. Our model thus provides a unifying framework in which voluntary actions emerge from a dynamic combination of internal action tendencies and external environmental factors, each of which can be subject to Change of Mind.

Fig. 1. Behavioural task.

Participants generated an endogenous colour intention (1) that had to be integrated with the sensory input of the dot-motion stimulus (2). Responses were indicated by moving the cursor to the target that matched both the colour intention and dot-motion direction (3). Continuous movement trajectories were measured during response execution allowing for online classification of perceptual Changes of Mind (CoM_P) and perceptual + intentional Changes of Mind (CoM_P+I). Once participants reached the target, 25/50/75/100% of the dots were painted in the colour of the hit target (4). On some trials, participants were asked to provide Sense of Agency (SoA) judgements (5a) or to estimate the percentage of dots that matched their initial colour intention (5b).

Fig. 2. Changes of Mind (CoM) in Exp. 1 (n = 17 participants).

A Percentage of trials classified as perceptual CoM (CoM_P; blue) and perceptual + intentional CoM (CoM_P+I; green) in test and easy trials. B Percentage of conflict trials with diagonal (blue) and horizontal (green) movement corrections of partial errors that were induced by mismatches between colour intentions and dot-motion direction. In both A and B, data are presented as mean values ± 1 SEM (**p < 0.001). Dots represent data points from individual participants. Statistical significance was obtained using likelihood ratio tests to compare logistic mixed-effects regression models with vs. without the fixed effects of interest (see “Methods”). No correction for multiple comparisons was performed since all comparisons are orthogonal and were planned prior to data collection. C Correlation across participants between RT costs in conflict trials and frequency of CoM_P+I (relative to overall percentage of all CoM), ρ(15) = −0.50, p = 0.043, two-tailed. Source data are provided as a Source Data file.

Fig. 3. Experiment 2: manipulation of horizontal target distance.

A Target locations in Exp. 1 and 2. B Motor costs for each type of Change of Mind (CoM) as measured by the distance from the diagonal vs. horizontal target as a function of travelled distance (assuming straight movement trajectories towards targets). In the far-target condition, costs associated with each target were roughly equal, whereas in the close-target condition, the target of different colour was closer, hence rendering intention pursuit relatively more costly. C Effect of target distance on frequency of CoM_P (blue) and CoM_P+I (green) in Exp. 2 (n = 16 participants). Data are presented as mean values ± 1 SEM (^**p < 0.001). Dots represent data points from individual participants. (likelihood ratio test for logistic mixed-effects models: **p < 0.001; no correction for multiple comparisons).

Fig. 4. Effect of Changes of Mind (CoM) on Sense of Agency (SoA) ratings in Exp. 1 and 2 (n = 33 participants).

A Mean SoA ratings for each type of CoM (grey = no CoM; blue = perceptual CoM; green = perceptual + intentional CoM) in test trials. In both (A) and (B), data are presented as mean values ± 1 SEM. Dots represent data points from individual participants. Post hoc tests in a linear mixed-effects regression model revealed that SoA ratings were lower in CoM_P than no-CoM trials (**p < 0.001; Bonferroni-corrected α-level of 0.05/3 = 0.017). B) Predicted SoA ratings (marginal effects) for a mixed-effects model with CoM (no-CoM/CoM_P/CoM_P+I) and movement times (MT) as predictors, based on collapsed data from Exp. 1 and 2 (M ± 95% CI).

Fig. 5. Hierarchical Attractor Network Model of Changes of Mind in voluntary action.

The network consists of 12 neural nodes that encode different pieces of information. Nodes are connected through excitatory (black) or inhibitory (red) connections. The action nodes A₁–A₄ compete against each other to determine which one of the four choice targets is selected. This competition takes into account information about (1) endogenous intentions (blue/green represented by nodes I₁ and I₂), (2) sensory information (left/right encoded by sensory nodes S₁ and S₂) and action costs (C₁–C₄) that depend on the distance d to each target location. Intention nodes are represented on a hierarchically higher level than sensorimotor nodes, allowing for top-down regulation of the degree of variability in firing rates of the action nodes. All firing rates are updated continuously and can change dynamically. Hence, CoM can occur when one action node crosses the threshold for movement execution first, but later on, another action wins the competition. Different types of CoM can be dissociated based on which action the network switches to when a decision reversal occurs (e.g., perceptual CoM: switch from A₁ to A₂; perceptual + intentional CoM: switch from A₁ to A₄).

Fig. 6. Changes of Mind (CoM) in attractor network model.

A Average firing rates of action nodes A₁ to A₄ (M ± 1 SD; n = 30,000 simulated trials) in trials without CoM (no-CoM), perceptual CoM (CoM_P), perceptual + intentional CoM (CoM_P+I), and vertical CoM. Firing rates were locked to time of first threshold crossing (first choice). For illustration purposes, only trials where the final choice was left-blue were included, except for CoM_P+I where a change with respect to the colour intention resulted in a final left-green choice. B Single-trial simulations showing firing rates of action nodes (upper row) and the resulting movement trajectories (bottom row) for no-CoM, CoM_P, CoM_P+I and vertical CoM. Dotted trajectories indicate completion of movements after the non-decision time of 380 ms (i.e., after the time period during which CoM can occur). C Effect of intentional strength on CoM_P (blue) and CoM_P+I (green). Stronger intentions result in lower frequency of intention reversals. D Stronger intentions increase RT costs in conflict trials (relative to RTs in easy trials). E Effect of target distance on CoM_P (blue) and CoM_P+I (green). Higher motor cost of intention pursuit increases frequency of intention reversals. Grey dashed lines in C and D indicate strength of intention in the model that was optimised to the results obtained in Exp. 1. In C–E, data are presented as the mean ± 1 SD from n = 30 model simulations with 1000 trials each. Black data points in E represent behavioural results from Exp. 2 (n = 16; M ± 1 SEM).