Quantum-Inspired Hybrid Neural Networks for Neural Decoding: A Controlled Ablation Study of Learnable Quantum Sidecar Integration

Diana Legziel Levy; Menachem Finkelstein; Peter Chin; Eilon Vaadia; Sarel Cohen

cs.LGAug 23, 2026

Quantum-Inspired Hybrid Neural Networks for Neural Decoding: A Controlled Ablation Study of Learnable Quantum Sidecar Integration

Authors: Diana Legziel Levy, Menachem Finkelstein, Peter Chin, Eilon Vaadia, Sarel Cohen

Organizations: Reichman University, Herzliya, Israel · Dartmouth College, USA · The Edmond and Lily Safra Center for Brain Sciences at The Hebrew University of Jerusalem, Israel

Abstract

We study parameterized quantum circuits (PQCs) integrated as residual sidecar modules within a ResNet-50 backbone for 31-class neural population decoding---imagined handwriting classification from multi-neuron spike rasters. Under strictly controlled conditions (fixed data splits, seeds, and optimizer), we compare four model variants: baseline, quantum sidecar with frozen input projection, quantum sidecar with backbone-gradient-trained projection, and a measurement-guided variant that aligns angle encodings with circuit measurement outcomes. The backbone-gradient variant improves accuracy in 3/4 seeds (+0.19% mean, 95% CI [-1.10%, +1.48%]) and consistently reduces Linear CKA similarity to baseline features ( $Δ=-0.025$ , 4/4 seeds), indicating genuine structural reorganization of representations. A nine-variant ablation identifies simple shallow architectures as the most effective and reproducible configuration. Measurement-guided training consistently improves representation geometry without reducing accuracy. All results use noiseless statevector simulation on 4 qubits, a regime chosen to reflect the practical constraints of current near-term superconducting hardware; no quantum computational advantage over classical methods is claimed.

Figures & tables

Panel of Figure 1 — Figure 1: Circuit topologies (1 of 2 layers shown). Chain: sequential nearest-neighbor CNOTs. Star: qubit 0 as hub with connections to all others.

Seed	Model	Top-1	Macro-F1	$\Delta$ Acc	Inter/Intra $R$	$\Delta R$	(vs Base)	(star pref.)
		Classification			Geometry ( $R\uparrow$ )		CKA $\downarrow$	$\boldsymbol{\pi}_{\textbf{star}}$
42	Baseline	85.88%	86.18%	—	2.919	—	1.000	—
	Quantum-Frozen	85.53%	85.91%	$-0.35\%$	3.001	$+0.082$	0.768	0.504
	Quantum-BackboneGrad	85.90%	86.31%	$\mathbf{+0.02\%}$	2.959	$+0.040$	0.747	0.562
	Geometry-loss	85.27%	86.06%	$-0.61\%$	3.809	$+0.890$	0.577	0.302
	Meas.-guided	85.67%	86.15%	$-0.21\%$	3.823	$\mathbf{+0.864}$	0.613	0.455

Table 1: Multi-seed comparison of five variants across four seeds.

R

= inter/intra cosine centroid ratio. CKA = Linear CKA between baseline and quantum features; lower indicates greater structural reorganization.

\pi_{\text{star}}

= learned star-topology weight at convergence. BackboneGrad : input projection trained via backbone pathway only (no gradient through quantum circuit; see Section 3 ). Geometry-loss : CE + feature geometry + angle geometry objectives (see Section 5.2 );

R

column shows the geometry improvement.

Variant	What changed	Top-1	$\Delta$	$R$
Quantum-BackboneGrad	Trainable feature-to-angle encoder; backbone-only gradient path; Z-only measurement; learned topology mixing.	86.32%	—	3.005
XYZ measurement	Full $\{X,Y,Z\}$ Pauli readout (12 observables) instead of Z-only (4 outputs).	85.74%	$-0.58\%$	2.902
Deep proj MLP	Shallow linear projection replaced by $1024\to 64\to 16\to 4$ MLP with PCA-initialized output.	85.95%	$-0.37\%$	3.034
Attention projection	Attention-style adaptive projection before angle generation.	85.77%	$-0.55\%$	2.943
Diverse projection	Orthogonality and diversity regularization to decorrelate angle subspaces.	86.08%	$-0.24\%$	3.009
Classical bottleneck	Same $1024\to 4\to 1024$ residual architecture with tanh nonlinearity; no PQC. Identical parameter count (18,442).	85.14%	$-1.18\%$	2.949

Table 2: Ablation and variant results (seed 777 unless noted ∗ ). Reference: Quantum-BackboneGrad 86.32%.

R

= inter/intra cosine centroid ratio.

\Delta

= Top-1 relative to reference.

Panel of Figure 2 — Figure 2: Left: per-class representation effects are topology-dependent, not frequency-driven. Right: the star-topology preference is a consistent emergent property of backbone-coupled training.

Explore similar work

Hybrid Quantum Neural Networks: Theory, Implementations, and Applications
Aug 2, 2026Léo Monbroussou, Maniraman Periyasamy, Viacheslav Kuzmin +4Hybrid Quantum-Classical ML Variational Quantum Circuits
Terra Quantum AG, Kornhausstrasse 25, St. Gallen, 9000, Switzerland · Département de Physique Appliquée, Université de Genève, 24 Quai Ernest-Ansermet, Geneva, 1211, Switzerland
Hybrid Variational Quantum-Classical Framework with Adaptive Weighting and Efficiency Assessment
Aug 30, 2026Dilli Hang RaiHybrid Quantum-Classical ML Quantum Neural Networks
Institute of Science and Technology Tribhuvan University, Kritipur, Nepal
Hybrid Quantum-Classical Neural Architecture Search
May 18, 2026Alberto Marchisio, Muhammad Kashif, Nouhaila Innan +1Quantum Machine Learning Hybrid Quantum-Classical ML
eBRAIN Lab, Division of Engineering, New York University Abu Dhabi (NYUAD), Abu Dhabi, United Arab Emirates · Center for Quantum and Topological Systems (CQTS), NYUAD Research Institute, NYUAD, Abu Dhabi, United Arab Emirates

Aug 2, 2026J/K move · Enter open · S save

Hybrid Quantum Neural Networks: Theory, Implementations, and Applications

Léo Monbroussou, Maniraman Periyasamy, Viacheslav Kuzmin, Pavel Sekatski+3

Terra Quantum AG, Kornhausstrasse 25, St. Gallen, 9000, Switzerland · Département de Physique Appliquée, Université de Genève, 24 Quai Ernest-Ansermet, Geneva, 1211, Switzerland

Hybrid Quantum-Classical ML Variational Quantum Circuits Quantum Neural Networks

Open paper AskPDF arXiv

Artificial intelligence has been transformed by deep neural networks, yet the search for new learning architectures continues. Quantum machine learning offers one such direction, and hybrid quantum neural networks, which combine classical neural-network components with quantum information processing units, have emerged as a practical framework for near-term quantum technologies. However, the rapid development of the field across diverse architectures, benchmarks and hardware assumptions makes it difficult to assess the utility of various proposals, identify where genuine advantages may arise, and determine how practitioners can use these models. While recent benchmarks caution that such gains have not yet been demonstrated at scale, theoretical work has identified tasks on which quantum models hold provable advantages, and hybrid approaches have delivered promising results on practical problems using deliberately compact quantum components and substantially fewer trainable parameters. Here, we review hybrid quantum neural networks for the machine-learning and quantum-machine-learning communities. We summarize their main theoretical and methodological foundations, survey some of the most promising architectures developed so far, and examine their implementation challenges and reported performance. By consolidating these perspectives, this review provides a structured view of the state of the field and helps identify promising paths for future research and application-driven development.

Abstract

Explore similar work

Hybrid Quantum Neural Networks: Theory, Implementations, and Applications

Hybrid Variational Quantum-Classical Framework with Adaptive Weighting and Efficiency Assessment

Hybrid Quantum-Classical Neural Architecture Search