betaFeedback

WorksPhysics and Astronomy

Scalable and Interpretable Quantum Mixture of Experts for High-Dimensional Spatio-Temporal Data

Abstract

Quantum Mixture of Experts (QMoE) architectures offer a promising path toward scalable quantum machine learning, yet existing approaches remain largely confined to low-dimensional benchmarks and often rely on naive input partitioning that disrupts feature dependencies. We address these gaps by introducing Feature-Partitioned QMoE, which partitions high-dimensional spatio-temporal inputs into functionally coherent feature groups and assigns each group to a dedicated quantum expert. We further propose a hierarchical interpretability framework that leverages the modular structure of QMoE to produce circuit-level gating analysis, per-expert gradient saliency, and heterogeneity-based subtype discovery. Applied to ADHD classification from resting-state fMRI (N=4,458, ABCD Study, 180 brain ROIs, 363 timepoints), the proposed Theory-Driven QMoE achieves performance comparable to matched-capacity classical and quantum baselines while enabling circuit-attributable interpretation. The framework identifies FDR-corrected ADHD-associated biomarkers in executive-control and anterior default-mode networks and reveals three neurobiologically meaningful ADHD subtypes. These results position Feature-Partitioned QMoE as a scalable and interpretable methodology for applying quantum machine learning to high-dimensional biomedical data.

Introduction

Quantum Machine Learning (QML) leverages superposition and entanglement to model complex problems in high-dimensional Hilbert spaces, with applications spanning chemistry, finance, high-energy physics, healthcare, and materials science [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13]. However, practical deployment remains constrained by NISQ hardware limitations, including limited qubit counts, short coherence times, and sparse qubit connectivity [2].

Multi-chip ensemble variational quantum circuits (VQCs) have been proposed to improve scalability and generalization by distributing computation across multiple quantum circuits, achieving strong performance on image classification, EEG analysis, and reinforcement learning tasks [14, 15]. Nevertheless, existing approaches scale computational cost with the number of experts and typically partition features using simple index-based splits that ignore intrinsic feature dependencies, leading to information loss in structured data such as spatio-temporal signals [14].

Quantum Mixture of Experts (QMoE) addresses these limitations by combining sparse expert routing with specialized quantum experts, activating only the most relevant circuits for each input while better preserving complex feature relationships [16, 17, 18]. This provides a scalable framework for modeling high-dimensional datasets such as resting-state fMRI.

The main contributions are summarized as follows: (1) We propose Feature-Partitioned QMoE, a quantum mixture-of-experts architecture demonstrated on high-dimensional biomedical time series with 180 ROIs, 363 timepoints, and N=4,458N = 4{,}458 subjects; (2) We introduce a circuit-aligned feature partitioning strategy that assigns each quantum expert to a neuroscience-defined functional circuit; (3) We show that theory-driven partitioning preserves substantially more functional connections within experts (62.4% vs. 43.9%) while maintaining comparable predictive performance; (4) We develop an architecture-agnostic hierarchical interpretability framework based on gating analysis, per-expert saliency, and heterogeneity-based subtype discovery.

Background

Variational Quantum Circuit

A Variational Quantum Circuit (VQC) maps a classical input vector x⃗\vec{x} into a high-dimensional Hilbert space via an encoding circuit U(x⃗)U(\vec{x}), producing the state U(x⃗)∣0⟩⊗nU(\vec{x})|0\rangle^{\otimes n}. This state is transformed by a parameterized ansatz W(Θ)W(\Theta), defined as a product of LL trainable layers:

∣ψ(x⃗,Θ)⟩=W(Θ)U(x⃗)∣0⟩⊗n=(∏j=L1Vj(θ⃗j))U(x⃗)∣0⟩⊗n(1)|\psi(\vec{x},\Theta)\rangle= W(\Theta)U(\vec{x})|0\rangle^{\otimes n} = \left(\prod_{j=L}^{1} V_j(\vec{\theta}_j)\right) U(\vec{x})|0\rangle^{\otimes n} \tag*{(1)}

Information is extracted via the expectation values of Hermitian observables H^k\hat{H}_k:

⟨H^k⟩=⟨ψ(x⃗,Θ)∣H^k∣ψ(x⃗,Θ)⟩(2)\langle\hat{H}_k\rangle= \langle\psi(\vec{x},\Theta)|\hat{H}_k|\psi(\vec{x},\Theta)\rangle \tag*{(2)}

The VQC implements a parametric function f⃗(x⃗;Θ)=(⟨H^1⟩,…,⟨H^n⟩)\vec{f}(\vec{x};\Theta) = (\langle\hat{H}_1\rangle,\ldots,\langle\hat{H}_n\rangle), which is optimized by minimizing a loss function L(x⃗,y;Θ)\mathcal{L}(\vec{x}, y; \Theta)—such as the mean squared error ∑i∥f⃗(x⃗i;Θ)−y⃗i∥2\sum_i \lVert\vec{f}(\vec{x}_i; \Theta) - \vec{y}_i \rVert^2—through a hybrid quantum-classical loop.

Quantum Time-series Transformer

The Quantum Time-series Transformer (QTSTransformer) replaces the quadratic complexity of classical self-attention with quantum primitives based on the linear combination of unitaries (LCU) and quantum singular value transformation (QSVT) [19]. Sequence embeddings are represented as a weighted superposition,

M=∑jbjUj,(3)M = \sum_j b_j U_j, \tag*{(3)}

where UjU_j denotes a quantum-encoded sequence segment. QSVT then applies a polynomial transformation

Pc⃗(M)=cdMd+⋯+c1M+c0I,(4)P_{\vec{c}}(M) = c_d M^d + \cdots+ c_1 M + c_0 I, \tag*{(4)}

enabling nonlinear processing analogous to classical attention mechanisms [20, 21]. This approach achieves O(polylog⁡T)O(\operatorname{polylog} T) complexity for sequence length TT, making it attractive for long-sequence modeling [19].

Quantum Mixture of Experts

Recent work has extended the Mixture of Experts (MoE) paradigm to quantum machine learning to improve model capacity and scalability. Proposed approaches include native QMoE architectures with quantum routing [16], mixtures of trainable quantum unitaries [22], infinite-width Mixtures of Quantum Experts (MoQE) [23], multimodal quantum MoE frameworks [24], and hybrid classical–quantum expert ensembles [25].

Feature-Partitioned Quantum Mixture of Experts

We present a QMoE architecture that partitions input features into functionally coherent clusters, processes each cluster through a dedicated VQC, and combines expert outputs via a learned gating network. The architecture is designed for high-dimensional datasets where input features exhibit structured dependencies that can be exploited through informed partitioning. Figure 1 provides an overview of the proposed theory-driven partitioning strategy.

Architecture overview of the Theory-Driven QMoE

Figure 1. Architecture overview of the Theory-Driven QMoE. Theory-Driven QMoE partitions 180 fMRI ROIs into four neuroscience-validated functional circuits (DMN, Executive, Salience, and SensoriMotor) based on the Yeo 17-network atlas.

Feature Partitioning

Consider an input X∈RT×C\mathbf{X} \in\mathbb{R}^{T \times C} with TT timesteps and CC features. We partition the features into KK disjoint subsets S1,…,SKS_1, \ldots, S_K with ⋃kSk={1,…,C}\bigcup_k S_k = \{1, \ldots, C\}, yielding per-expert inputs Xk=X[:,Sk]∈RT×∣Sk∣\mathbf{X}_k = \mathbf{X}[:, S_k] \in\mathbb{R}^{T \times|S_k|}.

Theory-Driven Partitioning

When domain knowledge provides a validated grouping of features into functionally coherent categories, this knowledge directly defines the partition via a mapping G:{1,…,C}→{1,…,K}\mathcal{G}: \{1, \ldots, C\} \to\{1, \ldots, K\}:

Sk={c∣G(c)=k}(5)S_k = \{c \mid\mathcal{G}(c) = k\} \tag*{(5)}

Examples include mapping spatial regions to known functional systems, grouping sensor channels by modality, or assigning variables by established causal structure. The key assumption is that features within each cluster share dependencies relevant to the prediction task.

Quantum Expert Architecture

Each expert kk is a hybrid classical-quantum model that maps Xk\mathbf{X}_k to a fixed-dimensional output hk∈RH\mathbf{h}_k \in\mathbb{R}^{H}. A classical encoder maps the input to rotation angles, a VQC applies parameterized single-qubit rotations and two-qubit entangling gates to an nqn_q-qubit register, and observable expectation values are mapped to hk\mathbf{h}_k via a classical output layer:

hk=Wout[⟨ψk∣Oj∣ψk⟩]j=1NO+bout,\mathbf{h}_k = \mathbf{W}_{\mathrm{out}} \left[\langle\psi_k \vert O_j \vert\psi_k \rangle\right]_{j=1}^{N_O} + \mathbf{b}_{\mathrm{out}},
∣ψk⟩=W(Θk)U(Ek(Xk))∣0⟩⊗nq(6)\lvert\psi_k \rangle= W(\Theta_k) U(\mathcal{E}_k(\mathbf{X}_k)) \lvert0 \rangle^{\otimes n_q} \tag*{(6)}

where Ek\mathcal{E}_k is the classical encoder, U(⋅)U(\cdot) is the data-encoding unitary, W(Θk)W(\Theta_k) is the trainable ansatz parameterized by $\Theta_k】【。 (cf. Eq. (1)), and {Oj}j=1NO\{O_j\}_{j=1}^{N_O} are measurement observables (e.g., single-qubit Pauli operators). The framework is agnostic to the specific ansatz. In the feature-partitioned setting, the entangling gates within the VQC are particularly relevant because they couple features that share genuine dependencies within each cluster rather than creating spurious correlations across unrelated features. For spatio-temporal inputs, temporal integration can be performed via strategies such as pooling, sequential processing, or a linear combination of unitaries [20]. The specific ansatz and temporal strategy used in our experiments are described in the Experiments section IV.

Gating Network and Prediction

A two-layer gating network with Gaussian exploration noise [18] computes soft routing weights from the temporal mean of the input:

g=softmax⁡(W2ReLU⁡(W1xˉ+b1)+b2+ϵ)∈RK,xˉ=1T∑t=1TX(t).(7)\mathbf{g} = \operatorname{softmax}(\mathbf{W}_2\operatorname{ReLU}(\mathbf{W}_1\bar{\mathbf{x}}+\mathbf{b}_1)+\mathbf{b}_2+\boldsymbol{\epsilon}) \in\mathbb{R}^{K}, \qquad \bar{\mathbf{x}} = \frac{1}{T}\sum_{t=1}^{T}\mathbf{X}^{(t)}. \tag*{(7)}

The final prediction is y^=h(∑kgkhk)\hat{y}=h\left(\sum_k g_kh_k\right), where hh is the classifier head. Training minimizes the task loss with a load-balancing regularizer [18, 26], L=Ltask+αLbalance\mathcal{L}=\mathcal{L}_{\mathrm{task}}+\alpha\mathcal{L}_{\mathrm{balance}}, which encourages uniform expert utilization.

Hierarchical Interpretability Framework

Because each expert processes only its assigned feature subset, the architecture enables interpretability at three nested levels that are native to the partitioned model. A monolithic single-expert model can only approximate these analyses by post-hoc grouping of its input gradient, without physical expert isolation; the feature-partitioned architecture instead yields per-expert, confound-free attributions by construction.

Level 1: Cluster-Level Gating Analysis

The gating weights g∈RK\mathbf{g}\in\mathbb{R}^{K} provide a per-sample routing profile. Comparing mean gate weights between classes via statistical testing (e.g., Welch’s tt-test) identifies which feature clusters are differentially engaged:

Δgk=E[gk∣y=1]−E[gk∣y=0](8)\Delta g_k=\mathbb{E}[g_k\mid y=1]-\mathbb{E}[g_k\mid y=0] \tag*{(8)}

Level 2: Per-Expert Gradient Saliency

Per-expert saliency is obtained by reducing the expert Jacobian ∂hk/∂Xk∈RH×T×∣Sk∣\partial h_k/\partial\mathbf{X}_k\in\mathbb{R}^{H\times T\times\lvert S_k\rvert} to a per-feature vector via averaging over the hidden (HH) and time (TT) axes:

sk[c]=1HT∑h=1H∑t=1T∂hk,h∂Xk,t,c∈R∣Sk∣,c∈Sk.(9)s_k[c]=\frac{1}{HT}\sum_{h=1}^{H}\sum_{t=1}^{T}\frac{\partial h_{k,h}}{\partial X_{k,t,c}}\in\mathbb{R}^{\lvert S_k\rvert},\qquad c\in S_k. \tag*{(9)}

This isolates what each expert learned within its cluster. Unlike model-level saliency (∂y^/∂X)(\partial\hat{y}/\partial\mathbf{X}), which conflates all experts and routing, per-expert saliency is computed with respect to expert kk’s input only, providing two complementary views: absolute saliency ∣sk∣\lvert s_k\rvert (magnitude of influence) and signed saliency sks_k (direction toward positive or negative class). Aggregating over known feature sub-groupings within each cluster reveals intra-cluster feature hierarchies.

Level 3: Heterogeneity Analysis via Multi-Level Clustering

The architecture provides three nested representation spaces for data-driven subtype discovery within a class of interest:

Level 1: g(i)∈RK\mathbf{g}^{(i)}\in\mathbb{R}^{K} (routing profile)

Level 2: [h1(i),…,hK(i)]∈RKH[\mathbf{h}_1^{(i)},\ldots,\mathbf{h}_K^{(i)}]\in\mathbb{R}^{KH} (expert representations)

Level 3: s(i)=[s1(i),…,sK(i)]∈RC\mathbf{s}^{(i)}=[\mathbf{s}_1^{(i)},\ldots,\mathbf{s}_K^{(i)}]\in\mathbb{R}^{C} (feature saliency)

where the Level-3 vector is the concatenation of the per-expert saliencies in Eq. (9) (since ∑k∣Sk∣=C\sum_k\lvert S_k\rvert=C). Subjects are clustered at each level and characterized by their dominant profiles. Cross-level concordance, measured via the Adjusted Rand Index (ARI) [27, 28], assesses whether subtypes are consistent across representational scales. This yields subtypes defined by how the model processes each subject through its learned, task-optimized representations, rather than by post-hoc analysis of raw data. Crucially, the domain interpretability of these subtypes depends on the partitioning strategy: theory-driven partitioning produces cluster-attributed, domain-interpretable subtypes.

Experiments

Dataset and Rationale

We evaluate Feature-Partitioned QMoE on resting-state fMRI from the Adolescent Brain Cognitive Development (ABCD) Study [29] for binary ADHD classification.

Resting-state fMRI provides a high-dimensional spatio-temporal benchmark with structured functional dependencies between brain regions, making it well suited for evaluating feature partitioning. ADHD also has well-established functional circuits [30, 31, 32, 33], enabling a neuroscience-informed partition into the Default Mode, Executive Control, Salience, and SensoriMotor networks. Finally, the large feature dimension (C=180C=180 ROIs) relative to the capacity of individual quantum circuits (nq=10n_q=10 qubits) naturally motivates a multi-expert architecture.

Data Preprocessing

The ABCD resting-state fMRI data was parcellated using the HCP-MMP1 atlas [34]. Following prior work [19, 35], bilateral homologous regions were averaged to obtain C=180C=180 cortical ROIs. Each scan consisted of T=363T=363 timepoints (TR =0.8=0.8 s). ADHD diagnosis was defined using the Child Behavior Checklist, with attention and ADHD TT-scores >65>65 classified as ADHD-positive; healthy controls had no diagnosed mental health disorders [36]. Subjects with fewer than 300 timepoints were excluded, and shorter scans were zero-padded to 363 timepoints. Per-ROI zz-score normalization was computed from the training set and values were clipped at ±10\pm10 standard deviations. The final dataset comprised N=4,458N=4{,}458 subjects, split into 70%/15%/15% training/validation/test sets using stratified random sampling (Ntrain=3,120N_{\mathrm{train}}=3{,}120, Nval=669N_{\mathrm{val}}=669, Ntest=669N_{\mathrm{test}}=669).

Quantum Expert Implementation

Each quantum expert is a QTSTransformer [19]. Given its assigned ROI subset Xk∈R363×∣Sk∣\mathbf{X}_k \in\mathbb{R}^{363 \times|S_k|}, temporal average pooling (p=10p = 10) reduces the sequence to T~=36\widetilde{T} = 36 timesteps, followed by sinusoidal positional encoding [37]. A learned linear projection with sigmoid activation maps the features to rotation angles, θ(t)=2π⋅σ(WX~k(t)+b)\theta^{(t)} = 2\pi\cdot\sigma(\mathbf{W}\widetilde{\mathbf{X}}_k^{(t)} + \mathbf{b}).

The variational quantum circuit uses nq=10n_q = 10 qubits and L=2L = 2 ansatz layers (Eq. (1)) with alternating RYRY rotations and CRXCRX entangling gates. Temporal integration is performed using a classically simulated linear combination of unitaries [20], followed by a degree-D=3D = 3 polynomial transformation inspired by quantum singular value transformation [21]. Pauli-X/Y/ZX/Y/Z measurements produce a 3nq=303n_q = 30-dimensional representation, which is projected to the expert embedding hk∈R64\mathbf{h}_k \in\mathbb{R}^{64}.

QMoE Model Configurations

We train a Theory-Driven QMoE with four experts (K=4K = 4) over 180 brain ROIs. The ROIs are partitioned into four neuroscience-validated functional circuits based on the Yeo 17-network atlas [38] and prior ADHD literature [30, 31]: (1) Default Mode Network (DMN): 41 ROIs; (2) Executive Control: 52 ROIs; (3) Salience: 44 ROIs; (4) SensoriMotor: 43 ROIs. This theory-driven partition preserves known functional organization and provides interpretable expert assignments.

Baseline Models

We compare against representative classical and quantum baselines trained using the same protocol (learning rate 10−310^{-3}, batch size 32, cosine annealing, and early stopping with patience 20):

  • Classical Single Expert: A Transformer processing all 180 ROIs jointly (134,849 parameters).

  • Classical 4-Expert MoE: Four classical Transformer experts with theory-driven ROI partitioning and learned gating (516,485 parameters).

  • Quantum Single Expert: A single QTSTransformer processing all 180 ROIs (16,936 parameters).

  • Quantum Arbitrary-Split MoE: Four QTSTransformer experts with arbitrary ROI partitions matched only in size to the theory-driven split (34,885 parameters).

The proposed Theory-Driven QMoE contains 34,885 trainable parameters, matching the arbitrary-split baseline and closely matched by the compact classical MoE (31,917 parameters); this parameter-matched control lets us compare quantum and classical experts at comparable capacity rather than against the much larger conventional MoE (516,485 parameters). All models were implemented in PyTorch 2.5.0 with PennyLane 0.42.3 [39] and trained on NVIDIA A100 (80 GB HBM) GPUs on the NERSC Perlmutter supercomputer.

Results

Table I summarizes the performance of all models on the ABCD ADHD binary classification task.

Theory-Driven Partitioning Preserves Functional Dependencies

We compare two QMoE models with identical architecture and parameter count (34,885 parameters) but different ROI assignments: theory-driven Yeo-17 partitioning and arbitrary index-based splitting. Both achieve comparable test AUC (0.5631 ± 0.0177 vs. 0.5576 ± 0.0151; paired tt-test across seeds, p=0.61p = 0.61), indicating that the neuroscience-informed partition incurs no accuracy penalty while enabling circuit-level interpretation.

On the group-average training functional-connectivity matrix, the theory-driven partition retains 62.4% of the top-5% strongest functional connections within experts, compared with 43.9% for arbitrary splitting (chance: 24.8%; identical expert sizes). This demonstrates that theory-driven partitioning better preserves functional dependencies while supporting hierarchical interpretability.

Hierarchical Interpretability

We apply the three-level interpretability framework (Section III-D) to the Theory-Driven QMoE (10 qubits, degree 3, 4 experts). We focus on this model because its experts process neuroscience-defined circuits, making per-expert saliency directly circuit-attributable. Figure 2 presents four key visualizations.

Figure 2. Hierarchical interpretability of the Theory-Driven QMoE (10 qubits, 4 experts). Red/blue denote positive/negative association with ADHD. (a) Per-expert network saliency. Top: absolute saliency; bottom: signed saliency with Welch’s tt-test markers (∗p<0.05\ast p < 0.05, ∗∗p<0.01\ast\ast p < 0.01, ∗∗∗p<0.001\ast\ast\ast p < 0.001; nominal/uncorrected—the FDR-surviving associations are those reported in the text). (b) Expert input projection weights grouped by Yeo-17 sub-networks; black borders indicate the highest-weighted network for each expert. (c) ROI-level clustering (K=3K = 3) of ADHD subjects. Cluster 2 (N=24N = 24, 8.3%) exhibits a prominent limbic temporal pole signature and is also identified by the classical 4-expert model. (d) Cross-level Adjusted Rand Index (ARI) between Circuit (gate weights), Network (expert outputs), and ROI (saliency) representations; Network–ROI ARI = 0.220 indicates partial subtype consistency across scales.

Cluster-Level Gating Analysis

Gate weights remain nearly uniform across the four experts (~0.250 ± 0.003), with no significant differences between ADHD and control subjects (all p>0.27p > 0.27). This reflects the load-balancing regularization (α=0.1\alpha= 0.1), indicating that classification relies on expert specialization rather than class-dependent routing.

Per-Expert Gradient Saliency

Per-expert saliency (∂hk/∂Xk\partial h_k/\partial X_k) is evaluated at the network level via Welch’s tt-test between ADHD and control subjects, with Benjamini–Hochberg false-discovery-rate (FDR) correction applied within each expert’s per-ROI family. Four network-level associations survive correction (q<0.05q < 0.05), all led by the Executive expert: ContC (q=0.0004q = 0.0004, positively associated with ADHD), DorsAttnB (q=0.003q = 0.003, positively associated), and ContA (q=0.032q = 0.032, negatively associated), together with the anterior default-mode sub-network DefaultA (q=0.027q = 0.027, positively associated). Two individual ROIs survive ROI-level correction: FST (q=0.007q = 0.007, +ADHD) and POS2 (q=0.027q = 0.027, −ADHD). Area 8C, the top-saliency ROI within the FDR-surviving ContC network, is elevated in ADHD and also emerges in the classical 4-expert model, but does not itself survive ROI-level correction; the robust signal is carried at the network level by ContC.

Figure 2a shows distinct intra-circuit hierarchies. The DMN expert prioritizes the anterior sub-network DefaultA over the posterior DefaultC, consistent with the anterior–posterior DMN dissociation in ADHD [30]. The Salience expert identifies Limbic_OFC and Limbic_TempPole as negatively associated with ADHD, whereas SalVentAttnA/B are positively associated; these Salience associations are nominal and do not survive FDR correction. Figure 2b shows clear expert specialization, with each expert assigning the highest weights to its designated Yeo-17 subnetworks, supporting the effectiveness of the theory-driven partition.

ModelPartitioningParamsTest AUC
Classical SENone (all 180 ROIs)134,8490.5997 ±\pm 0.0174
Classical MoETheory-driven (Yeo-17)516,4850.5686 ±\pm 0.0246
Compact Classical MoETheory-driven (Yeo-17)31,9170.5873 ±\pm 0.0025
Quantum Arb MoEArbitrary (contiguous)34,8850.5576 ±\pm 0.0151
Theory QMoE (ours)Theory-driven (Yeo-17)34,8850.5631 ±\pm 0.0177

Table I. ADHD classification results on the ABCD dataset (N=4,458N = 4,458).

Heterogeneity Analysis (ADHD Subtype Discovery)

Multi-level clustering of subjects with ADHD (N=288N = 288) identifies three neurobiological subtypes at the ROI level (Figure 2c). The largest subtype (Cluster 1, 59.7%) is characterized by elevated SalVentAttnA activation, the second (Cluster 0, 31.9%) by increased Limbic_OFC importance, and a rare subtype (Cluster 2, 8.3%, N=24N = 24) by strong Limbic_TempPole involvement, suggesting distinct attentional, reward-processing, and memory/emotional dysfunction profiles, respectively. The rare subtype is also recovered by the classical 4-expert model, indicating that it reflects genuine neurobiological structure rather than a model artifact.

Cross-level concordance (Figure 2d) shows very low Circuit–Network agreement (ARI = 0.011), consistent with near-uniform gate weights, but moderate Network–ROI agreement (ARI = 0.220), indicating that ADHD heterogeneity is primarily captured at the network and ROI levels rather than by coarse circuit routing.

Conclusion

We presented Feature-Partitioned QMoE, a quantum mixture-of-experts architecture that addresses three limitations of existing multi-chip quantum ensembles: naïve feature partitioning, limited scalability to high-dimensional data, and the lack of systematic interpretability.

More broadly, Feature-Partitioned QMoE provides a scalable and interpretable framework for high-dimensional scientific data. Although demonstrated using neuroscience-defined functional circuits, the proposed feature-partitioning strategy naturally extends to other structured domains with meaningful feature groupings, such as gene regulatory networks and multimodal sensor data. Future work will investigate data-driven expert partitioning when domain-specific feature groupings are unavailable. The proposed interpretability framework—combining gating analysis, per-expert saliency, and heterogeneity clustering—generalizes to any MoE architecture with disjoint feature subsets, providing a systematic approach for extracting scientific insights from quantum models.

Acknowledgment

The authors extend their gratitude to the members of the Connectome Lab for their invaluable support and critical contributions to the development of this study. This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. 2021R1C1C1006503, RS-2023-00266787, RS-2023-00265406, RS-2024-00421268, RS-2024-00432301, RS-2024-00435727, RS-2025-25457329, RS-2021-NR061370, NRF-2021M3E5D2A01022515, and NRF-2021S1A3A2A02090597), by the Researchers Program through Seoul National University (No. 200-20250071, 200-20250049, 200-20250116, 200-20250115, 200-20250113, 0670-20250039, 200-20260009). Additional support was provided by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korea government (MSIT) [No. RS-2021-II211343 Artificial Intelligence Graduate School Program, Seoul National University] and by the Global Research Support Program in the Digital Field (RS-2024-00421268). This work was also supported by the Artificial Intelligence Industrial Convergence Cluster Development Project funded by the Ministry of Science and ICT and Gwangju Metropolitan City, by the Korea Brain Research Institute (KBRI) basic research program (25-BR-05-01), by the Korea Health Industry Development Institute (KHIDI) and the Ministry of Health and Welfare, Republic of Korea (HR22C1605), and by the Korea Basic Science Institute (National Research Facilities and Equipment Center) grant funded by the Ministry of Education (RS-2024-00435727). We acknowledge the National Supercomputing Center for providing supercomputing resources and technical support (KSC-2023-CRE-0568, KSC-2024-CRE-0198, KSC-2025-CRE-0340). An award for computer time was provided by the U.S. Department of Energy’s (DOE) ASCR Leadership Computing Challenge (ALCC). This research used resources of the National Energy Research Scientific Computing Center (NERSC), a DOE Office of Science User Facility, under ALCC award m4750-2024, and supporting resources at the Argonne and Oak Ridge Leadership Computing Facilities, U.S. DOE Office of Science user facilities at Argonne National Laboratory and Oak Ridge National Laboratory.

References

  1. [1]J. Biamonte, P. Wittek, N. Pancotti, P. Rebentrost, N. Wiebe, and S. Lloyd, “Quantum machine learning,” Nature, vol. 549, no. 7671, pp. 195–202, 2017.DOI
  2. [2]M. Cerezo, C. Verdon, H.-Y. Huang, L. Cinicio, and P. J. Coles, “Challenges and opportunities in quantum machine learning,” Nature Computational Science, vol. 2, no. 9, pp. 567–576, 2022.DOI
  3. [3]Y. Gujju, A. Matsuo, and R. Raymond, “Quantum machine learning on near-term quantum devices: Current state of supervised and unsupervised techniques for real-world applications,” Phys. Rev. Appl., vol. 21, p. 067001, Jun 2024.arxiv.org/abs/2307.00908
  4. [4]D. Peral-García, J. Cruz-Benito, and F. J. García-Peñalvo, “Systematic literature review: Quantum machine learning and its applications,” Computer Science Review, vol. 51, p. 100619, 2024.
  5. [5]J. Wen, Z. Wang, C. Chen, J. Xiao, H. Li, Q. Zhan, H. Fan, S. Wei, and G. Long, “A full circuit-based quantum algorithm for excited-states in quantum chemistry,” Quantum, vol. 8, p. 1219, Jan. 2024.
  6. [6]S. McArdle, S. Endo, A. Aspuru-Guzik, S. C. Benjamin, and X. Yuan, “Quantum computational chemistry,” Rev. Mod. Phys., vol. 92, p. 015003, Mar 2020.DOI
  7. [7]D. Herman, C. Googin, X. Liu, Y. Sun, A. Galda, I. Safro, M. Pistoia, and Y. Alexeev, “Quantum computing for finance,” Nature Reviews Physics, vol. 5, no. 8, pp. 450–465, 2023.DOI
  8. [8]D. J. Egger, C. Gambella, J. Marecek, S. McFaddin, M. Mevissen, R. Raymond, A. Simonetto, S. Woerner, and E. Yndurain, “Quantum computing for finance: State-of-the-art and future prospects,” IEEE Transactions on Quantum Engineering, vol. 1, pp. 1–24, 2020.DOI
  9. [9]S. Y.-C. Chen, T.-C. Wei, C. Zhang, H. Yu, and S. Yoo, “Quantum convolutional neural networks for high energy physics data analysis,” Phys. Rev. Res., vol. 4, p. 013231, Mar 2022.
  10. [10]A. Di Meglio, K. Jansen, I. Tavernelli, C. Alexandrou, S. Anurachalam, C. W. Bauer, et al., “Quantum computing for high-energy physics: State of the art and challenges,” PRX Quantum, vol. 5, no. 3, p. 037001, 2024.
  11. [11]S. Mensa, E. Sahin, F. Tacchino, P. K. Barkoutsos, and I. Tavernelli, “Quantum machine learning framework for virtual screening in drug discovery: a prospective quantum advantage,” Machine Learning: Science and Technology, vol. 4, no. 1, p. 015023, 2023.
  12. [12]U. Ullah, A. G. O. Jurado, I. D. Gonzalez, and B. Garcia-Zapirain, “A fully connected quantum convolutional neural network for classifying ischemic cardiopathy,” IEEE Access, vol. 10, pp. 134592–134605, 2022.DOI
  13. [13]I. Cong, S. Choi, and M. D. Lukin, “Quantum convolutional neural networks,” Nature Physics, vol. 15, no. 12, pp. 1273–1278, 2019.DOI
  14. [14]J. J. Park, J. Cha, S. Y.-C. Chen, H.-H. Tseng, and S. Yoo, “Addressing the current challenges of quantum machine learning through multi-chip ensembles,” 2025.arxiv.org/abs/2505.08782
  15. [15]J. J. Park, H.-H. Tseng, S. Yoo, S. Y.-C. Chen, and J. Cha, “It’s-A-Me, Quantum Mario: Scalable quantum reinforcement learning with multi-chip ensembles,” in 2025 IEEE International Conference on Quantum Artificial Intelligence (QAI), pp. 366–372, 2025.arxiv.org/abs/2509.00713
  16. [16]H.-Q. Nguyen, X.-B. Nguyen, S. Pandey, S. U. Khan, I. Safro, and K. Luu, “QMoE: A quantum mixture of experts framework for scalable quantum neural networks,” in 2025 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 02, pp. 223–228, 2025.arxiv.org/abs/2507.05190
  17. [17]R. A. Jacobs, M. I. Jordan, S. J. Nowlan, and G. E. Hinton, “Adaptive mixtures of local experts,” Neural Computation, vol. 3, pp. 79–87, 03 1991.DOI
  18. [18]N. Shazeer, A. Mirhoseini, K. Maziarz, A. Davis, Q. Le, G. Hinton, and J. Dean, “Outrageously large neural networks: The sparsely-gated mixture-of-experts layer,” in International Conference on Learning Representations, 2017.arxiv.org/abs/1701.06538
  19. [19]J. J. Park, J. Seo, S. Bae, S. Y.-C. Chen, H.-H. Tseng, J. Cha, and S. Yoo, “Resting-state fMRI analysis using quantum time-series transformer,” in 2025 IEEE International Conference on Quantum Computing and Engineering (QCE), vol. 01, pp. 2352–2363, 2025.arxiv.org/abs/2509.00711
  20. [20]A. M. Childs and N. Wiebe, “Hamiltonian simulation using linear combinations of unitary operations,” Quantum Information & Computation, vol. 12, no. 11–12, pp. 901–924, 2012.
  21. [21]A. Gilyén, Y. Su, G. H. Low, and N. Wiebe, “Quantum singular value transformation and beyond: exponential improvements for quantum matrix arithmetics,” in Proceedings of the 51st Annual ACM SIGACT Symposium on Theory of Computing, STOC 2019, (New York, NY, USA), p. 193–204, Association for Computing Machinery, 2019.DOI
  22. [22]B. Coyle, S. Raj, N. Mathur, A. E. Cherrat, N. Jain, S. Kazdaghli, and I. Kerenidis, “Training-efficient density quantum machine learning,” npj Quantum Information, vol. 11, no. 1, p. 172, 2025.arxiv.org/abs/2405.20237
  23. [23]P. A. X. Tognini, L. Banchi, and G. D. Palma, “Solving MNIST with a globally trained mixture of quantum experts,” 2025.arxiv.org/abs/2505.14789
  24. [24]S. Dutta, N. Innan, S. B. Yahia, M. Shafique, and D. E. Bernal Neira, “MQFL-FHE: Multimodal quantum federated learning framework with fully homomorphic encryption,” in 2025 International Joint Conference on Neural Networks (IJCNN), pp. 1–10, 2025.DOI
  25. [25]R. Chaves, K. Kumar, B. Chagas, R. Linerud, B. Sorem, J. Mancilla, and B. Bell, “A mixture-of-experts framework for practical hybrid-quantum models in credit card fraud detection,” 2026.arxiv.org/abs/2603.06473
  26. [26]W. Fedus, B. Zoph, and N. Shazeer, “Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity,” Journal of Machine Learning Research, vol. 23, no. 120, pp. 1–39, 2022.
  27. [27]L. Hubert and P. Arabie, “Comparing partitions,” Journal of Classification, vol. 2, no. 1, p. 193–218, 1985.DOI
  28. [28]D. Steinley, “Properties of the hubert-arable adjusted rand index,” Psychological Methods, vol. 9, no. 3, p. 386–396, 2004.DOI
  29. [29]B. Casey, T. Cannonier, M. I. Conley, A. O. Cohen, D. M. Barch, M. M. Hetzel, et al., “The Adolescent Brain Cognitive Development (ABCD) Study: Imaging acquisition across 21 sites,” Developmental Cognitive Neuroscience, vol. 32, pp. 43–54, 2018. The Adolescent Brain Cognitive Development (ABCD) Consortium: Rationale, Aims, and Assessment Strategy.
  30. [30]F. X. Castellanos and E. Proal, “Large-scale brain systems in ADHD: beyond the prefrontal–striatal model,” Trends in Cognitive Sciences, vol. 16, no. 1, pp. 17–26, 2012. Special Issue: Cognition in Neuropsychiatric Disorders.DOI
  31. [31]S. Cortese, C. Kelly, C. Chabernaud, E. Proal, A. D. Martino, M. P. Milham, and F. X. Castellanos, “Toward systems neuroscience of ADHD: A meta-analysis of 55 fMRI studies,” American Journal of Psychiatry, vol. 169, no. 10, pp. 1038–1055, 2012.DOI
  32. [32]J. T. Nigg, S. L. Karalunas, E. Feczko, and D. A. Fair, “Toward a revised nosology for attention-deficit/hyperactivity disorder heterogeneity,” Biological Psychiatry: Cognitive Neuroscience and Neuroimaging, vol. 5, no. 8, pp. 726–737, 2020. Understanding the Nature and Treatment of Psychopathology: Letting the Data Guide the Way.DOI
  33. [33]S. Koirala, G. Grimsrud, M. A. Mooney, B. Larsen, E. Feczko, J. T. Elison, S. M. Nelson, J. T. Nigg, B. Tervo-Clemmens, and D. A. Fair, “Neurobiology of attention-deficit hyperactivity disorder: historical challenges and emerging frontiers,” Nature Reviews Neuroscience, vol. 25, no. 12, p. 759–775, 2024.DOI
  34. [34]M. F. Glasser, T. S. Coalson, E. C. Robinson, C. D. Hacker, J. Harwell, E. Yacoub, K. Ugurbil, J. Andersson, C. F. Beckmann, M. Jenkinson, S. M. Smith, and D. C. Van Essen, “A multi-modal parcellation of human cerebral cortex,” Nature, vol. 536, no. 7615, pp. 171–178, 2016.DOI
  35. [35]S. Bae, J. Kwon, S. Yoo, and J. Cha, “Spatiotemporal learning of brain dynamics from fMRI using frequency-specific multi-band attention for cognitive and psychiatric applications,” 2025.
  36. [36]M. M. Cordova, D. M. Antovich, P. Ryabinin, C. Neighbor, M. A. Mooney, N. F. Dieckmann, O. Miranda-Dominguez, B. J. Nagel, D. A. Fair, and J. T. Nigg, “Attention-deficit/hyperactivity disorder: Restricted phenotypes prevalence, comorbidity, and polygenic risk sensitivity in the ABCD baseline cohort,” Journal of the American Academy of Child & Adolescent Psychiatry, vol. 61, no. 10, pp. 1273–1284, 2022.
  37. [37]A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. u. Kaiser, and I. Polosukhin, “Attention is all you need,” in Advances in Neural Information Processing Systems (I. Guyon, U. V. Luxburg, S. Bengio, H. Wallach, R. Fergus, S. Vishwanathan, and R. Garnett, eds.), vol. 30, Curran Associates, Inc., 2017.
  38. [38]B. T. Thomas Yeo, F. M. Krienen, J. Sepulcre, M. R. Sabuncu, D. Lashkari, M. Hollinshead, J. L. Roffman, J. W. Smoller, L. Zöllei, J. R. Polimeni, B. Fischl, H. Liu, and R. L. Buckner, “The organization of the human cerebral cortex estimated by intrinsic functional connectivity,” Journal of Neurophysiology, vol. 106, no. 3, pp. 1125–1165, 2011. PMID: 21653723.
  39. [39]V. Bergholm, J. Izaac, M. Schuld, C. Gogolin, S. Ahmed, V. Ajith, et al., “PennyLane: Automatic differentiation of hybrid quantum-classical computations,” 2022.

Paper details

vlnc:W_str9phr2
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY-SA 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Physics and Astronomy

Not recorded for this paper

View statement
Looking for related papers…
Contents