Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Log inSign up
Challenges
Datasets
Workspaces
Papers
Discussions
Leaderboard
Blog
Job Board
Q3AS

© 2026 Aqora Quantum S.A.S.

TermsPrivacyLegal Notice

arXiv:2610.22621

Noise-shaped Sampling and Annealing for Scalable On-the-edge Training of Binary and Ternary AI Models

Zihao Chen, Zhili Xiao, Ruijie Zhu, Yipeng Wang, Jason K. Eshraghian, Shantanu Chakrabartty

Published Oct 15, 2026• 0 citations
arXivHTMLPDF

Abstract

The training loss of binary and ternary neural networks, expanded as a multilinear polynomial over the discrete weights, is structurally equivalent to the energy of a hierarchical high-order Ising model. We exploit this equivalence to apply Ising solvers to discrete-weight network training, while accounting for a fundamental difference in objectives: an Ising solver seeks the ground state of a fixed energy function, whereas training seeks solutions that generalize to unseen data. Our key contribution is $ΣΔ$-Sampler, an optimizer that reconciles these objectives by embedding a stochastic sampling-and-annealing step as a noisy quantizer inside a first-order $ΣΔ$ noise-shaping loop. The loop pushes noise from weight quantization, high-order frustration, and mini-batch sampling out of the low-frequency adaptation band, so that discrete weight updates follow the smooth descent trajectories embedded in the rugged loss landscape. We validate the framework on a planted quadratic unconstrained binary optimization (QUBO) benchmark with a known minimizer; on CIFAR-10, where a $13$M-parameter ternary model comes within $1\%$ of full-precision accuracy; on ImageNet-1k, where a ternary ResNet-18 comes within $0.4\%$ Top-5 accuracy of a ternary baseline; on class-incremental continual learning, where prior-task accuracy is preserved; and on MatMul-free language models up to $2.7$B parameters, where it outperforms a straight-through-estimator baseline under matched continued pre-training. Compared with standard quantization-aware training, $ΣΔ$-Sampler reduces the training memory footprint by over $50\%$, delivers an order-of-magnitude speed-up, and reduces weight-update memory traffic by four orders of magnitude, making it attractive for on-device training and fine-tuning of large binary and ternary models.

DOI
10.48550/arxiv.2610.22621

Related papers

Training squared-hinge support vector machines by an explicit QUBO–Ising construction with quantum-annealing execution

Zonglin Yang, Jie Zhou, Xiaohui Yang +2 more

Quantum annealers natively minimize quadratic unconstrained binary optimization (QUBO) problems, yet faithfully compiling continuous convex objectives into discrete binary forms with formal guarantees remains challenging. We present a complete, algebraically verifiable pipeline for training a linear squared-hinge support vector machine on quantum annealing hardware. The construction comprises four stages with rigorous justification: (i) an exact epigraph reformulation eliminating the hinge nonlinearity, (ii) equality conversion via surplus variables with a quadratic penalty whose exactness on the finite binary domain is formally established, (iii) closed-form QUBO coefficients and a provably energy-preserving Ising mapping, and (iv) a moment-based three-component decoder that reconstructs continuous parameters from noisy annealer samples using empirical first- and second-order statistics. We execute this pipeline end-to-end on D-Wave Advantage systems and evaluate under a rigorous protocol with 30 stratified splits, bootstrap confidence intervals, paired tests, and effect sizes. The feature-wise solver achieves 87-89% test accuracy on the Iris benchmark, competitive with classical baselines at this scale. We contribute a fully auditable reduction from convex SVM training to Ising optimization rather than claiming quantum advantage, and explicitly characterize limitations from discretization, embedding overhead, and feature-wise decomposition.

Scientific Reports•May 31, 2026•0 citations

Quadratic Unconstrained Binary Optimisation for Training and Regularisation of Binary Neural Networks

Jonas Christoffer Villumsen, Yusuke Sugita

Advances in artificial intelligence (AI) and deep learning have raised concerns about its increasing energy consumption, while demand for deploying AI in mobile devices and machines at the edge is growing. Binary neural networks (BNNs) have recently gained attention as energy and memory efficient models suitable for resource constrained environments; however, training BNNs exactly is computationally challenging because of its discrete characteristics. Recent work proposing a framework for training BNNs based on quadratic unconstrained binary optimisation (QUBO) and progress in the design of Ising machines for solving QUBO problems suggest a potential path to efficiently optimising discrete neural networks. In this work, we extend existing QUBO models for training BNNs to accommodate arbitrary network topologies and propose two novel methods for regularisation. The first method maximises neuron margins biasing the training process toward parameter configurations that yield larger pre-activation magnitudes. The second method employs a dropout-inspired iterative scheme in which reduced subnetworks are trained and used to adjust linear penalties on network parameters. We apply the proposed QUBO formulation to a small binary image classification problem and conduct computational experiments on a GPU-based Ising machine. The numerical results indicate that the proposed regularisation terms modify training behaviour and yield improvements in classification accuracy on data not present in the training set.

arXiv:2601.00449•Jan 1, 2026•0 citations

Training Multi-layer Neural Networks on Ising Machine

Xujie Song, Tong Liu, Shengbo Eben Li +3 more

As a dedicated quantum device, Ising machines could solve large-scale binary optimization problems in milliseconds. There is emerging interest in utilizing Ising machines to train feedforward neural networks due to the prosperity of generative artificial intelligence. However, existing methods can only train single-layer feedforward networks because of the complex nonlinear network topology. This paper proposes an Ising learning algorithm to train quantized neural network (QNN), by incorporating two essential techinques, namely binary representation of topological network and order reduction of loss function. As far as we know, this is the first algorithm to train multi-layer feedforward networks on Ising machines, providing an alternative to gradient-based backpropagation. Firstly, training QNN is formulated as a quadratic constrained binary optimization (QCBO) problem by representing neuron connection and activation function as equality constraints. All quantized variables are encoded by binary bits based on binary encoding protocol. Secondly, QCBO is converted to a quadratic unconstrained binary optimization (QUBO) problem, that can be efficiently solved on Ising machines. The conversion leverages both penalty function and Rosenberg order reduction, who together eliminate equality constraints and reduce high-order loss function into a quadratic one. With some assumptions, theoretical analysis shows the space complexity of our algorithm is $\mathcal{O}(H^2L + HLN\log H)$, quantifying the required number of Ising spins. Finally, the algorithm effectiveness is validated with a simulated Ising machine on MNIST dataset. After annealing 700 ms, the classification accuracy achieves 98.3%. Among 100 runs, the success probability of finding the optimal solution is 72%. Along with the increasing number of spins on Ising machine, our algorithm has the potential to train deeper neural networks.

arXiv:2311.03408•Nov 6, 2023•1 citation

Layer-wise QUBO-Based Training of CNN Classifiers for Quantum Annealing

Mostafa Atallah, Rebekah Herrman

Variational quantum circuits for image classification suffer from barren plateaus, while quantum kernel methods scale quadratically with dataset size. We propose an iterative framework based on Quadratic Unconstrained Binary Optimization (QUBO) for training the classifier head of convolutional neural networks (CNNs) via quantum annealing, entirely avoiding gradient-based circuit optimization. Following the Extreme Learning Machine paradigm, convolutional filters are randomly initialized and frozen, and only the fully connected layer is optimized. At each iteration, a convex quadratic surrogate derived from the feature Gram matrix replaces the non-quadratic cross-entropy loss, yielding an iteration-stable curvature proxy. A per-output decomposition splits the $C$-class problem into $C$ independent QUBOs, each with $(d+1)K$ binary variables, where $d$ is the feature dimension and $K$ is the bit precision, so that problem size depends on the image resolution and bit precision, not on the number of training samples. We evaluate the method on six image-classification benchmarks (sklearn digits, MNIST, Fashion-MNIST, CIFAR-10, EMNIST, KMNIST). A precision study shows that accuracy improves monotonically with bit resolution, with 10 bits representing a practical minimum for effective optimization; the 15-bit formulation remains within the qubit and coupler limits of current D-Wave Advantage hardware. The 20-bit formulation matches or exceeds classical stochastic gradient descent on MNIST, Fashion-MNIST, and EMNIST, while remaining competitive on CIFAR-10 and KMNIST. All experiments use simulated annealing, establishing a baseline for direct deployment on quantum annealing hardware.

quant-pharXiv:2603.02958•Mar 4, 2026•0 citations

Projected Stochastic Gradient Descent with Quantum Annealed Binary Gradients

Maximilian Krahn, Sasdelli, Michele, Fengyi Yang +4 more

We present, QP-SBGD, a novel layer-wise stochastic optimiser tailored towards training neural networks with binary weights, known as binary neural networks (BNNs), on quantum hardware. BNNs reduce the computational requirements and energy consumption of deep learning models with minimal loss in accuracy. However, training them in practice remains to be an open challenge. Most known BNN-optimisers either rely on projected updates or binarise weights post-training. Instead, QP-SBGD approximately maps the gradient onto binary variables, by solving a quadratic constrained binary optimisation. Under practically reasonable assumptions, we show that this update rule converges with a rate of $\mathcal{O}(1 / \sqrt{T})$. Moreover, we show how the $\mathcal{NP}$-hard projection can be effectively executed on an adiabatic quantum annealer, harnessing recent advancements in quantum computation. We also introduce a projected version of this update rule and prove that if a fixed point exists in the binary variable space, the modified updates will converge to it. Last but not least, our algorithm is implemented layer-wise, making it suitable to train larger networks on resource-limited quantum hardware. Through extensive evaluations, we show that QP-SBGD outperforms or is on par with competitive and well-established baselines such as BinaryConnect, signSGD and ProxQuant when optimising the Rosenbrock function, training BNNs as well as binary graph neural networks.

arXiv:2310.15128•Oct 23, 2023•0 citations

Black-Box Optimization of QUBO-Based Machine Learning Models Using Tensor Train Sampling

Hadi Salloum, Osama Orabi, Salman Ahmdi +4 more

The Probabilistic Optimizer with TEnsor Sampling (PROTES) is a gradient-free optimization method for multidimensional arrays and discretized multivariate functions based on the tensor train (TT) format. This novel tensor-network methodology demonstrates significant promise for complex Quadratic Unconstrained Binary Optimization (QUBO) problems. We establish the first application of PROTES to machine learning model training via QUBO formulations and provide the inaugural comparative analysis against D-Wave quantum annealing. Our results reveal a critical performance dichotomy: For smallscale, structurally complex QUBO instances, PROTES achieves superior solution accuracy with enhanced computational stability compared to quantum annealing, while maintaining competitive efficiency. Conversely, Quantum Annealing retains scalability advantages for larger problem dimensions. These findings delineate precise operational domains for tensor-network versus quantum optimization paradigms, establishing PROTES as the preferred solver for compact, high-complexity QUBO problems. This research provides fundamental insights for strategic solver selection and advances the engineering of hybrid quantumclassical optimization frameworks.

•Jul 19, 2025•0 citations
The training loss of binary and ternary neural networks, expanded as a multilinear polynomial over the discrete weights, is structurally equivalent to the energy of a hierarchical high-order Ising model. We exploit this equivalence to apply Ising solvers to discrete-weight network training, while accounting for a fundamental difference in objectives: an Ising solver seeks the ground state of a fixed energy function, whereas training seeks solutions that generalize to unseen data. Our key contribution is ΣΔΣΔΣΔ-Sampler, an optimizer that reconciles these objectives by embedding a stochastic sampling-and-annealing step as a noisy quantizer inside a first-order ΣΔΣΔΣΔ noise-shaping loop. The loop pushes noise from weight quantization, high-order frustration, and mini-batch sampling out of the low-frequency adaptation band, so that discrete weight updates follow the smooth descent trajectories embedded in the rugged loss landscape. We validate the framework on a planted quadratic unconstrained binary optimization (QUBO) benchmark with a known minimizer; on CIFAR-10, where a 131313M-parameter ternary model comes within 1%1\%1% of full-precision accuracy; on ImageNet-1k, where a ternary ResNet-18 comes within 0.4%0.4\%0.4% Top-5 accuracy of a ternary baseline; on class-incremental continual learning, where prior-task accuracy is preserved; and on MatMul-free language models up to 2.72.72.7B parameters, where it outperforms a straight-through-estimator baseline under matched continued pre-training. Compared with standard quantization-aware training, ΣΔΣΔΣΔ-Sampler reduces the training memory footprint by over 50%50\%50%, delivers an order-of-magnitude speed-up, and reduces weight-update memory traffic by four orders of magnitude, making it attractive for on-device training and fine-tuning of large binary and ternary models.
Quantum annealers natively minimize quadratic unconstrained binary optimization (QUBO) problems, yet faithfully compiling continuous convex objectives into discrete binary forms with formal guarantees remains challenging. We present a complete, algebraically verifiable pipeline for training a linear squared-hinge support vector machine on quantum annealing hardware. The construction comprises four stages with rigorous justification: (i) an exact epigraph reformulation eliminating the hinge nonlinearity, (ii) equality conversion via surplus variables with a quadratic penalty whose exactness on the finite binary domain is formally established, (iii) closed-form QUBO coefficients and a provably energy-preserving Ising mapping, and (iv) a moment-based three-component decoder that reconstructs continuous parameters from noisy annealer samples using empirical first- and second-order statistics. We execute this pipeline end-to-end on D-Wave Advantage systems and evaluate under a rigorous protocol with 30 stratified splits, bootstrap confidence intervals, paired tests, and effect sizes. The feature-wise solver achieves 87-89% test accuracy on the Iris benchmark, competitive with classical baselines at this scale. We contribute a fully auditable reduction from convex SVM training to Ising optimization rather than claiming quantum advantage, and explicitly characterize limitations from discretization, embedding overhead, and feature-wise decomposition.
Advances in artificial intelligence (AI) and deep learning have raised concerns about its increasing energy consumption, while demand for deploying AI in mobile devices and machines at the edge is growing. Binary neural networks (BNNs) have recently gained attention as energy and memory efficient models suitable for resource constrained environments; however, training BNNs exactly is computationally challenging because of its discrete characteristics. Recent work proposing a framework for training BNNs based on quadratic unconstrained binary optimisation (QUBO) and progress in the design of Ising machines for solving QUBO problems suggest a potential path to efficiently optimising discrete neural networks. In this work, we extend existing QUBO models for training BNNs to accommodate arbitrary network topologies and propose two novel methods for regularisation. The first method maximises neuron margins biasing the training process toward parameter configurations that yield larger pre-activation magnitudes. The second method employs a dropout-inspired iterative scheme in which reduced subnetworks are trained and used to adjust linear penalties on network parameters. We apply the proposed QUBO formulation to a small binary image classification problem and conduct computational experiments on a GPU-based Ising machine. The numerical results indicate that the proposed regularisation terms modify training behaviour and yield improvements in classification accuracy on data not present in the training set.
As a dedicated quantum device, Ising machines could solve large-scale binary optimization problems in milliseconds. There is emerging interest in utilizing Ising machines to train feedforward neural networks due to the prosperity of generative artificial intelligence. However, existing methods can only train single-layer feedforward networks because of the complex nonlinear network topology. This paper proposes an Ising learning algorithm to train quantized neural network (QNN), by incorporating two essential techinques, namely binary representation of topological network and order reduction of loss function. As far as we know, this is the first algorithm to train multi-layer feedforward networks on Ising machines, providing an alternative to gradient-based backpropagation. Firstly, training QNN is formulated as a quadratic constrained binary optimization (QCBO) problem by representing neuron connection and activation function as equality constraints. All quantized variables are encoded by binary bits based on binary encoding protocol. Secondly, QCBO is converted to a quadratic unconstrained binary optimization (QUBO) problem, that can be efficiently solved on Ising machines. The conversion leverages both penalty function and Rosenberg order reduction, who together eliminate equality constraints and reduce high-order loss function into a quadratic one. With some assumptions, theoretical analysis shows the space complexity of our algorithm is O(H2L+HLNlog⁡H)\mathcal{O}(H^2L + HLN\log H)O(H2L+HLNlogH), quantifying the required number of Ising spins. Finally, the algorithm effectiveness is validated with a simulated Ising machine on MNIST dataset. After annealing 700 ms, the classification accuracy achieves 98.3%. Among 100 runs, the success probability of finding the optimal solution is 72%. Along with the increasing number of spins on Ising machine, our algorithm has the potential to train deeper neural networks.
Variational quantum circuits for image classification suffer from barren plateaus, while quantum kernel methods scale quadratically with dataset size. We propose an iterative framework based on Quadratic Unconstrained Binary Optimization (QUBO) for training the classifier head of convolutional neural networks (CNNs) via quantum annealing, entirely avoiding gradient-based circuit optimization. Following the Extreme Learning Machine paradigm, convolutional filters are randomly initialized and frozen, and only the fully connected layer is optimized. At each iteration, a convex quadratic surrogate derived from the feature Gram matrix replaces the non-quadratic cross-entropy loss, yielding an iteration-stable curvature proxy. A per-output decomposition splits the CCC-class problem into CCC independent QUBOs, each with (d+1)K(d+1)K(d+1)K binary variables, where ddd is the feature dimension and KKK is the bit precision, so that problem size depends on the image resolution and bit precision, not on the number of training samples. We evaluate the method on six image-classification benchmarks (sklearn digits, MNIST, Fashion-MNIST, CIFAR-10, EMNIST, KMNIST). A precision study shows that accuracy improves monotonically with bit resolution, with 10 bits representing a practical minimum for effective optimization; the 15-bit formulation remains within the qubit and coupler limits of current D-Wave Advantage hardware. The 20-bit formulation matches or exceeds classical stochastic gradient descent on MNIST, Fashion-MNIST, and EMNIST, while remaining competitive on CIFAR-10 and KMNIST. All experiments use simulated annealing, establishing a baseline for direct deployment on quantum annealing hardware.
We present, QP-SBGD, a novel layer-wise stochastic optimiser tailored towards training neural networks with binary weights, known as binary neural networks (BNNs), on quantum hardware. BNNs reduce the computational requirements and energy consumption of deep learning models with minimal loss in accuracy. However, training them in practice remains to be an open challenge. Most known BNN-optimisers either rely on projected updates or binarise weights post-training. Instead, QP-SBGD approximately maps the gradient onto binary variables, by solving a quadratic constrained binary optimisation. Under practically reasonable assumptions, we show that this update rule converges with a rate of O(1/T)\mathcal{O}(1 / \sqrt{T})O(1/T​). Moreover, we show how the NP\mathcal{NP}NP-hard projection can be effectively executed on an adiabatic quantum annealer, harnessing recent advancements in quantum computation. We also introduce a projected version of this update rule and prove that if a fixed point exists in the binary variable space, the modified updates will converge to it. Last but not least, our algorithm is implemented layer-wise, making it suitable to train larger networks on resource-limited quantum hardware. Through extensive evaluations, we show that QP-SBGD outperforms or is on par with competitive and well-established baselines such as BinaryConnect, signSGD and ProxQuant when optimising the Rosenbrock function, training BNNs as well as binary graph neural networks.
The Probabilistic Optimizer with TEnsor Sampling (PROTES) is a gradient-free optimization method for multidimensional arrays and discretized multivariate functions based on the tensor train (TT) format. This novel tensor-network methodology demonstrates significant promise for complex Quadratic Unconstrained Binary Optimization (QUBO) problems. We establish the first application of PROTES to machine learning model training via QUBO formulations and provide the inaugural comparative analysis against D-Wave quantum annealing. Our results reveal a critical performance dichotomy: For smallscale, structurally complex QUBO instances, PROTES achieves superior solution accuracy with enhanced computational stability compared to quantum annealing, while maintaining competitive efficiency. Conversely, Quantum Annealing retains scalability advantages for larger problem dimensions. These findings delineate precise operational domains for tensor-network versus quantum optimization paradigms, establishing PROTES as the preferred solver for compact, high-complexity QUBO problems. This research provides fundamental insights for strategic solver selection and advances the engineering of hybrid quantumclassical optimization frameworks.