Skip to main content
Back to timeline
arXivSource publication:

Quantum experts stop slicing brain regions by index: grouping by functional network lifts strong-connection retention from 43.9% to 62.4%

Synopsis

Quantum mixture-of-experts models for high-dimensional spatio-temporal data have split input features by index, cutting the dependencies that link them. Feature-Partitioned QMoE instead assigns 180 brain regions to four quantum experts by functionally coherent network, so each expert sees one neuroscience-defined circuit. On ADHD classification from ABCD resting-state fMRI, this grouping matches arbitrary splitting in test AUC while raising the share of the top 5% strongest functional connections kept inside experts from 43.9% to 62.4%, and it supports circuit-level attribution.

Source-provided article image: Scalable and Interpretable Quantum Mixture of Experts for High-Dimensional Spatio-Temporal Data
arXiv

Interpretation

A quantum mixture-of-experts model can now split high-dimensional input into functionally coherent groups defined by domain knowledge, sending each group to a dedicated quantum expert instead of cutting by index order. Earlier multi-chip quantum ensembles partitioned features with simple index-based splits that ignored intrinsic feature dependencies, losing information in structured data such as spatio-temporal signals. On ADHD classification from ABCD resting-state fMRI, the theory-driven and arbitrary-split QMoE share the same architecture and 34,885 parameters, reach comparable test AUC, and show no significant difference in a paired t-test across seeds.

The 180 brain regions are grouped by the Yeo-17 functional networks into default mode, executive control, salience and sensorimotor sets, each handled by one quantum expert whose output a gating network weights before the classifier head. Each quantum expert is a quantum time-series transformer that first averages the 363 timepoints, projects features into rotation angles, and feeds a 10-qubit variational circuit with entangling gates. Training uses ABCD resting-state fMRI parcellated with the HCP-MMP1 atlas, with the four experts receiving 41, 52, 44 and 43 ROIs, implemented in PyTorch and PennyLane and trained on A100 GPUs.

The partitioned structure yields three levels of interpretability: gate weights give a per-sample routing profile, per-expert gradient saliency gives region importance inside each circuit, and multi-level clustering gives data-driven subtypes. A monolithic model can only approximate these analyses by post-hoc grouping of its input gradient, without physically isolated experts; the partitioned architecture gives per-expert, confound-free attributions by construction. In the executive expert, associations for the ContC, DorsAttnB and ContA networks and the anterior default-mode sub-network DefaultA survive FDR correction, as do the individual regions FST and POS2; multi-level clustering of ADHD subjects yields three subtypes, the largest at 59.7%.

Perspective

The demonstrated setting is binary ADHD classification on ABCD resting-state fMRI with 180 brain regions and 363 timepoints, partitioned by Yeo-17 networks and four functional circuits defined from prior ADHD literature; the authors propose future work on data-driven expert partitioning when domain groupings are unavailable.

Gate weights stay nearly uniform across the four experts, with no significant difference between ADHD and control subjects, which the authors attribute to load-balancing regularization and read as reliance on expert specialization rather than class-dependent routing; salience-network associations are only nominal and do not survive FDR correction; cross-level concordance between circuit and network levels is ARI = 0.011.