A Hierarchical Attractor Network Model of perceptual versus intentional decision updates

The study investigates how people update voluntary actions when decisions must integrate internal intentions with external perceptual evidence and motor costs. While prior work on Changes of Mind (CoM) emphasized perceptual decision reversals in noisy environments, voluntary actions also involve higher-order distal intentions (what goal to pursue) and lower-level motor intentions (how to implement it). The authors propose two types of CoM in voluntary action: (1) perceptual CoM (reversals about external evidence) and (2) perceptual plus intentional CoM (reversals that additionally change the internally generated goal). They hypothesize that: (a) intention reversals are less likely when initial intentions are stronger; and (b) intention reversals are more likely when motor costs of pursuing the initial intention are higher. They also explore how CoM affects Sense of Agency (SoA). To explain mechanisms, they introduce a hierarchical, multimodal Attractor Network Model integrating intentions, sensory evidence, and motor costs, with top-down noise control from higher-order intentions.

Research on CoM has largely focused on perceptual decisions (e.g., random-dot motion tasks) showing ongoing evidence accumulation after action initiation and occasional trajectory reversals. Voluntary action theories distinguish distal intentions (what to do) from motor intentions (how to do it) and show flexible updates of lower-level motor plans (e.g., double-step paradigms). Less is known about when internally generated distal goals are maintained or abandoned. Prior modeling of CoM used bounded accumulators or attractor networks for unidimensional perceptual choices, sometimes incorporating static reward/cost biases. The present work extends this by modeling continuous, dynamic integration of multiple modalities (intentions, sensory evidence, action costs) within a hierarchical framework where higher-order intentions exert top-down noise reduction over sensorimotor processes, aligning with theories of hierarchical control in frontal cortex and neural noise reduction as a marker of endogenous action control.

Two behavioral experiments adapted the random-dot motion (RDM) task to combine an endogenous color intention (e.g., blue vs. green) with a perceptual decision (left vs. right motion). Participants moved a cursor via a pen tablet to one of four color-coded targets (two per color). The correct response matched both the perceived motion direction and the internally chosen color. Movement trajectories (125 Hz sampling) were used to classify in-flight decision updates. Participants: Exp. 1 included 17 right-handed adults (13 female, mean age 22.6). Exp. 2 included 16 right-handed adults (11 female, mean age 23.2). Ethics approval and consent obtained; performance-based payment provided. Stimuli and apparatus: Variable-coherence RDM presented in a central aperture (4.5°) at 60 Hz; dot density 15.6 dots deg−2 s−1. In Exp. 1, four 1.8° targets at 9.6° eccentricity. In Exp. 2, horizontal distance between cross-color neighbors varied by trial (far: 18° horizontal, 6° vertical; close: 6° horizontal, 18° vertical), with constant radial distance to control detectability. Cursor speed slower in Exp. 2 to magnify travel costs. Procedure: At trial start, participants freely chose a color (silent most trials; occasional prompts to verbalize). Targets appeared either 700–1000 ms before dot motion (early) or at motion onset (late). Participants initiated a reach; the motion stimulus turned off at movement initiation. Upon target hit, 25/50/75/100% of the dots were shown in the chosen target color; SoA judgments (0–100) were collected after every CoM trial and 33% of no-CoM trials; outcome percentage estimates were collected on a subset of remaining trials. Three trial types: test (70%; low coherence individually titrated to ~60% accuracy), easy (10%; 80% coherence), and conflict (20%; 80% coherence; both same-color targets on one side to induce color–motion conflict). Response deadlines and feedback promoted speedy, accurate responding. Training: Day-1 training to 70% accuracy at 35% coherence, followed by staircase to identify individual coherence for ~60% accuracy for test trials (mean 11.8%). CoM classification: Online trajectory-based rules. If cursor exceeded 10% of x and y displacement toward an initial target then ended at the diagonally opposite target of the same color (direction change but color maintained), the trial was a perceptual CoM (CoMp). If it ended at the horizontally adjacent target of the other color (direction and color changed), it was a perceptual + intentional CoM (CoMp+I). Vertical switches (same side, different color) were analyzed separately (interpreted as motor target-selection corrections). Double CoM and atypical trajectories were excluded. Conflict trials: High coherence with target arrangement causing about half of trials to mismatch color intention and motion direction; participants instructed to respond to motion. RT and error costs in conflict vs easy trials index the presence and strength of initial color intentions. Statistical analyses: Mixed-effects logistic regressions (lme4) modeled CoM frequency: (a) no-CoM vs CoM (overall CoM) and (b) CoMp vs CoMp+I (within-CoM trials). Participants as random intercepts; likelihood-ratio tests for fixed effects. ANOVAs/t-tests for RT/accuracy. SoA analyzed with linear mixed-effects models; CoM type and movement time (MT; z-standardized per participant) as predictors; outcome percentage manipulations tested linear effects. Correlations (Spearman) assessed relationships across participants (e.g., conflict RT cost vs CoMp+I proportion). Computational model: A hierarchical Attractor Network with 12 nodes modeled dynamic integration. Nodes: two intention nodes (I1/I2), two sensory nodes (S1/S2), four cost nodes (C1–C4) encoding distance-derived motor costs, and four action nodes (A1–A4) corresponding to target choices. Excitatory links from intentions and sensory nodes to actions; inhibitory cost inputs; lateral inhibition within modalities; stronger inhibition among actions of the same color; self-excitation in sensory nodes for evidence integration. Hierarchy implemented as top-down noise regulation: stronger intentions reduce noise in associated action nodes (inspired by Hierarchical Gaussian Filters). Firing rates (0–100 Hz) updated at 1 ms with mean-field dynamics (τ=100 ms) plus Gaussian noise; external inputs normalized (60 Hz base; scaled by coherence or intention strength). Movements initiated when an action node crossed 40 Hz and exceeded others by 10 Hz, with non-decision time of 380 ms (200 ms sensory + 180 ms motor), during which further updates can cause CoM and movement redirection. Costs updated online from current cursor position to targets (Euclidean distance). Parameters (connectivity weights, evidence strengths, noise control) fitted via CMA-ES to Exp. 1 test-trial behavior (RTs, accuracy, CoMp/CoMp+I/vertical rates), then tested on Exp. 2 geometry to predict motor-cost effects. Model comparisons using AIC evaluated necessity of cost nodes and hierarchical noise control.

Experiment 1 (n=17):

• Manipulation checks: Test trials (low coherence) produced lower accuracy and longer RTs than easy trials. Accuracy: test M=56.6% (SD 9.1%) vs easy M=93.4% (SD 7.0%), t(16)=20.13, p<0.001, d=4.88. RTs: test M=570.5 ms (SD 58.3) vs easy M=534.2 ms (SD 41.5), t(16)=3.99, p=0.001, d=0.97.
• Overall perceptual CoM were more frequent in test than easy trials: test M=7.64% (SD 6.74%) vs easy M=1.28% (SD 2.33%), OR=6.27, χ²(1)=45.69, p<0.001. Most CoM corrected initial perceptual errors: 60.9% (SD 16.5%), t(16)=2.72, p=0.015, d=0.66.
• Within CoM trials, participants more often maintained color intentions (CoMp) than changed them (CoMp+I) in test trials: CoMp M=5.9% (SD 5.5%) vs CoMp+I M=1.7% (SD 2.2%); intercept < 0, OR=0.20, z=−4.42, p<0.001. Trial difficulty (easy vs test) did not affect the CoMp vs CoMp+I split (p=0.890).
• Conflict trials (n=17) evidenced genuine prior color intentions: RTs slower than easy (M 549.7 vs 534.2 ms), t(16)=2.51, p=0.023; accuracy trend lower (90.5% vs 94.1%), t(16)=2.11, p=0.051. Corrective movements occurred more often than in easy trials matched for coherence (OR=2.97, χ²(1)=9.93, p=0.002), and horizontal corrections were preferred over diagonal (OR=5.71, z=3.46, p<0.001), indicating sensitivity to motor costs.
• Intentional strength effect: Greater conflict RT costs (conflict−easy) predicted fewer intention reversals among CoM (lower CoMp+I proportion): Spearman ρ(15)=−0.50, p=0.043.
• Vertical corrections occurred (M 3.24% of test trials) and were more frequent with late targets (OR=3.25, χ²(1)=38.76, p<0.001); they did not correlate with conflict RT costs, suggesting target selection errors rather than true intention changes.

Experiment 2 (n=16):

• Manipulated horizontal distance between different-color neighbor targets (far vs close) to vary relative motor costs of intention pursuit after a perceptual CoM. CoMp+I was more frequent when switching to the other color was motor-cheaper (close) than when costs were similar (far): close M=2.59% (SD 0.44%) vs far M=1.48% (SD 0.68%), OR=2.15, χ²(1)=15.47, p<0.001. Target distance did not affect whether a perceptual CoM occurred at all (no effect on no-CoM vs CoM; p=0.404). Interaction with early/late target onset was non-significant (p=0.097).

Sense of Agency (Exp. 1+2 combined, N=33):

• In no-CoM test trials, SoA increased linearly with outcome percentage (F(1,32)=164.91, p<0.001, ηp²=0.837), validating sensitivity to outcomes.
• With outcome fixed at 50% in CoM trials, SoA was lower in CoMp than no-CoM (b=−3.02 percentage points, 95% CI [−4.62, −1.42], p<0.001), while CoMp+I did not differ from no-CoM (p=0.366). However, after including movement time (MT) as a covariate, the CoM effect on SoA disappeared (χ²(2)=1.51, p=0.470); longer MTs predicted lower SoA (b=−1.81 per z-MT unit, 95% CI [−2.58, −1.14], p<0.001), indicating SoA reductions were mediated by prolonged movement duration.

Computational model:

• The hierarchical Attractor Network reproduced behavioral patterns: frequencies similar to Exp. 1 (CoMp ~6.33%, CoMp+I ~1.41%, vertical ~3.6%) and confirmed that most vertical switches corrected initial color errors (~77.6%), while most CoMp+I reflected genuine intention changes after correct initial color selection (~56.6%).
• Increasing modeled intentional strength reduced CoMp+I and increased conflict RT costs, mirroring individual differences in Exp. 1.
• Varying target distances reproduced Exp. 2’s increase in CoMp+I when intention pursuit was motor-costly.
• Model comparisons favored the full model over alternatives without hierarchical noise control (AIC higher; relative likelihood 0.139) or without cost nodes (relative likelihood 0.0005).

The findings show that voluntary action selection involves continuous integration of endogenous intentions, perceptual evidence, and evolving motor costs. When perceptual evidence changed after action initiation, participants typically maintained their initial goal despite incurring larger motor corrections, indicating that distal intentions occupy a higher priority in the action hierarchy. However, intention reversals occurred more frequently when intentions were weaker (indexed by smaller conflict RT costs) and when the motor cost of intention pursuit was higher (shorter path to the alternative-color target). Thus, both internal goal strength and external action costs dynamically shape whether agents persist with or abandon goals during ongoing actions. Sense of Agency reductions associated with CoM were attributable to longer movement times rather than intention changes per se, aligning with accounts that SoA relies on retrospective inferences from action fluency and outcomes rather than access to initial intentions. The hierarchical Attractor Network captured these patterns through top-down noise reduction by intentions and dynamic integration in action nodes, supporting an interactive hierarchy where action representations contribute to evolving decisions. This bridges decision-making and motor control perspectives, emphasizing parallel representation and competition among action affordances and the reciprocal influence of movement execution on subsequent decision updates via changing costs.

This work introduces a behavioral paradigm and a hierarchical, multimodal Attractor Network Model that together account for Changes of Mind in voluntary action. Empirically, intention reversals are less likely with stronger intentions and more likely when intention pursuit is motor-costly, while overall perceptual CoM depend on sensory uncertainty. SoA depends on action outcomes and movement duration, not directly on intention reversals. Computationally, the model offers a biologically plausible mechanism for dynamic, hierarchical integration of intentions, sensory evidence, and costs, with top-down noise regulation. Future research should capture purely endogenous intention updates independent of perceptual changes, incorporate within-trial variability in inputs (sensory fluctuations, intention strength, cost estimates), and test neural predictions linking action-node dynamics to fine-grained motor policies (e.g., speed/vigor) and prefrontal control of sensorimotor noise.

• The computational model drives CoM primarily via internal neural noise, omitting within-trial fluctuations in sensory evidence and potential variability in intention strength and cost estimates; inputs were treated as constant across a trial.
• SoA conclusions regarding intention reversals are limited by low trial counts (some participants lacked CoMp+I) and the null effect after covariate control.
• Vertical movement corrections likely reflect target selection errors; while analyses argue against widespread intention changes, some ambiguity remains.
• Experiments were conducted once each without independent replication, though Exp. 2 results were consistent with Exp. 1.
• The paradigm indexes average intentional strength across the task rather than trial-by-trial fluctuations.