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

Proceedings of the AAAI Conference on Artificial Intelligence

On the Challenges of Physical Implementations of RBMs

Vincent Dumoulin, Ian Goodfellow, Aaron C. Courville, Yoshua Bengio

Published Jun 21, 2014• 47 citations
Publisher

Abstract

Restricted Boltzmann machines (RBMs) are powerful machine learning models, but learning and some kinds of inference in the model require sampling-based approximations, which, in classical digital computers, are implemented using expensive MCMC. Physical computation offers the opportunity to reduce the costof sampling by building physical systems whose natural dynamics correspond to drawing samples from the desired RBM distribution. Such a system avoids the burn-in and mixing cost of a Markov chain. However, hardware implementations of this variety usually entail limitations such as low-precision and limited range of the parameters and restrictions on the size and topology of the RBM. We conduct software simulations to determine how harmful each of these restrictions is. Our simulations are based on the D-Wave Two computer, but the issues we investigate arise in most forms of physical computation.Our findings suggest that designers of new physical computing hardware and algorithms for physical computers should focus their efforts on overcoming the limitations imposed by the topology restrictions of currently existing physical computers.

DOI
10.1609/aaai.v28i1.8924

Related papers

On the Challenges of Physical Implementations of RBMs

Vincent Dumoulin, Ian Goodfellow, Aaron C. Courville +1 more

Restricted Boltzmann machines (RBMs) are powerful machine learning models, but learning and some kinds of inference in the model require sampling-based approximations, which, in classical digital computers, are implemented using expensive MCMC. Physical computation offers the opportunity to reduce the cost of sampling by building physical systems whose natural dynamics correspond to drawing samples from the desired RBM distribution. Such a system avoids the burn-in and mixing cost of a Markov chain. However, hardware implementations of this variety usually entail limitations such as low-precision and limited range of the parameters and restrictions on the size and topology of the RBM. We conduct software simulations to determine how harmful each of these restrictions is. Our simulations are designed to reproduce aspects of the D-Wave quantum computer, but the issues we investigate arise in most forms of physical computation.

arXiv:1312.5258•National Conference on Artificial Intelligence•Dec 18, 2013•37 citations

Investigation of D-Wave quantum annealing for training Restricted Boltzmann Machines and mitigating catastrophic forgetting

Abdelmoula El-Yazizi, Yaroslav Koshka

Abstract Modest statistical differences between the sampling performances of the D-Wave quantum annealer (QA) and the classical Markov Chain Monte Carlo (MCMC), when applied to Restricted Boltzmann Machines (RBMs), are explored to explain, and possibly address, the absence of significant and consistent improvements in RBM trainability when the D-Wave sampling was used in previous investigations. A novel hybrid sampling approach, combining the classical and the QA contributions, is investigated as a promising way to benefit from the modest differences between the two sampling methods. No improvements in the RBM training are achieved in this work, thereby suggesting that the differences between the QA-based and MCMC sampling, mainly found in the medium-to-low probability regions of the distribution, which are less important for the quality of the sample, are insufficient to benefit the training. Difficulties in achieving sufficiently high quality of embedding RBMs into the lattice of the newer generation of D-Wave hardware could be further complicating the task. On the other hand, the ability to generate samples of sufficient variety from lower-probability parts of the distribution has a potential to benefit other machine learning applications, such as the mitigation of catastrophic forgetting (CF) during incremental learning. The feasibility of using QA-generated patterns of desirable classes for CF mitigation by the generative replay is demonstrated in this work for the first time. While the efficiency of the CF mitigation using the D-Wave QA was comparable to that of the classical mitigation, both the speed of generating a large number of distinct desirable patterns and the potential for further improvement make this approach promising for a variety of challenging machine learning applications.

arXiv:2508.15697•Dec 15, 2025•1 citation

Hardware Acceleration of Frustrated Lattice Systems using Convolutional Restricted Boltzmann Machine

Pratik Prabhanjan Brahma, J. H. Han, Tamzid Razzaque +2 more

Geometric frustration gives rise to emergent quantum phenomena and exotic phases of matter. While Monte Carlo methods are traditionally used to simulate such systems, their sampling efficiency is limited by the complexity of interactions and ground-state properties. Restricted Boltzmann Machines (RBMs), a class of probabilistic neural networks, offer improved sampling by incorporating machine learning techniques. However, fully-connected bipartite RBMs are inefficient for representing physical lattices with sparse interactions. To address this, we implement Convolutional Restricted Boltzmann Machines (CRBMs) that leverage translational symmetry inherent to lattices. Using the classical Shastry-Sutherland (SS) Ising lattice, we demonstrate (i) CRBM formulation that captures SS interactions, and (ii) digital hardware accelerator to enhance sampling performance. We simulate lattices with up to 324 spins, recovering all known phases of the SS Ising model, including the long range ordered fractional plateau. Our hardware characterizes spin behavior at critical points and within spin liquid phases. This implementation achieves a speedup of 3 to 5 orders of magnitude (33 ns to 120 ms) over GPU-based implementations. Moreover, the time-to-solution is within two orders of magnitude of quantum annealers, while offering superior scalability, room-temperature operation and reprogrammability. This work paves a pathway for scalable digital hardware that embeds physical symmetries to enable large scale simulations of material systems.

arXiv:2511.20911•Nov 25, 2025•0 citations

Determination of the Lowest-Energy States for the Model Distribution of Trained Restricted Boltzmann Machines Using a 1000 Qubit D-Wave 2X Quantum Computer

Yaroslav Koshka, Dilina Perera, Spencer Hall +1 more

The possibility of using a quantum computer D-Wave 2X with more than 1000 qubits to determine the global minimum of the energy landscape of trained restricted Boltzmann machines is investigated. In order to overcome the problem of limited interconnectivity in the D-Wave architecture, the proposed RBM embedding combines multiple qubits to represent a particular RBM unit. The results for the lowest-energy (the ground state) and some of the higher-energy states found by the D-Wave 2X were compared with those of the classical simulated annealing (SA) algorithm. In many cases, the D-Wave machine successfully found the same RBM lowest-energy state as that found by SA. In some examples, the D-Wave machine returned a state corresponding to one of the higher-energy local minima found by SA. The inherently nonperfect embedding of the RBM into the Chimera lattice explored in this work (i.e., multiple qubits combined into a single RBM unit were found not to be guaranteed to be all aligned) and the existence of small, persistent biases in the D-Wave hardware may cause a discrepancy between the D-Wave and the SA results. In some of the investigated cases, introduction of a small bias field into the energy function or optimization of the chain-strength parameter in the D-Wave embedding successfully addressed difficulties of the particular RBM embedding. With further development of the D-Wave hardware, the approach will be suitable for much larger numbers of RBM units.

Neural Computation•May 31, 2017•15 citations

Software techniques for training restricted Boltzmann machines on size-constrained quantum annealing hardware

Ilmo Salmenperä, Jukka K. Nurminen

Restricted Boltzmann machines are common machine learning models that can utilize quantum annealing devices in their training processes as quantum samplers. While this approach has shown promise as an alternative to classical sampling methods, the limitations of quantum annealing hardware, such as the number of qubits and the lack of connectivity between the qubits, still pose a barrier to wide-scale adoption. We propose the use of multiple software techniques such as dropout method, passive labeling, and parallelization techniques for addressing these hardware limitations. The study found that using these techniques along with quantum sampling showed comparable results to its classical counterparts in certain contexts, while in others the increased complexity of the sampling process hindered the performance of the trained models. This means that further research into the behavior of quantum sampling needs to be done to apply quantum annealing to training tasks of more complicated RBM models.

Frontiers in Computer Science•Oct 16, 2023•3 citations

2000 Qubit D-Wave Quantum Computer Replacing MCMC for RBM Image Reconstruction and Classification

Yaroslav Koshka, M. A. Novotny

Restricted Boltzmann Machines trained by the classical Contrastive Divergence algorithm were embedded into a next generation D-Wave quantum computer having more than 2000 qubits. The embedding was used to sample missing visible units and labels of incomplete test images. The previously introduced RBM embedding that combined qubits into individual RBM units in order to increase the connectivity between units was used. In comparison to the 1000 qubit machine investigated earlier, the same embedding approach applied to the 2000 qubit D-Wave enabled a connectivity between RBM units sufficient for RBM training on a classical computer with a satisfactory classification error. After a varied number of CD training iterations, the visible units corresponding to a particular incomplete test image were fixed in the D-Wave embedding. The resulting state returned by the D-Wave represents the most energetically favorable combination of the remaining qubits for the given fixed incomplete input image. The classification errors compared favorably to those obtained by the classical MCMC sampling. Main factors influencing the quality of the embedding were investigated. Regardless of the computational cost incentives for using the D-Wave for this operation, the reconstruction and classification errors determined in this way can be used at least as a verification of the validity and quality of the particular embedding of the given RBM architecture. This is essential for successfully employing a QC in much more computationally demanding parts of the RBM training. opportunities for using QC not only for recognition but especially for RBM training, in particular for obtaining a representative sample from the RBM model distribution, are discussed.

International Joint Conference on Neural Networks (IJCNN)•Jul 1, 2018•5 citations
Restricted Boltzmann machines (RBMs) are powerful machine learning models, but learning and some kinds of inference in the model require sampling-based approximations, which, in classical digital computers, are implemented using expensive MCMC. Physical computation offers the opportunity to reduce the costof sampling by building physical systems whose natural dynamics correspond to drawing samples from the desired RBM distribution. Such a system avoids the burn-in and mixing cost of a Markov chain. However, hardware implementations of this variety usually entail limitations such as low-precision and limited range of the parameters and restrictions on the size and topology of the RBM. We conduct software simulations to determine how harmful each of these restrictions is. Our simulations are based on the D-Wave Two computer, but the issues we investigate arise in most forms of physical computation.Our findings suggest that designers of new physical computing hardware and algorithms for physical computers should focus their efforts on overcoming the limitations imposed by the topology restrictions of currently existing physical computers.
Restricted Boltzmann machines (RBMs) are powerful machine learning models, but learning and some kinds of inference in the model require sampling-based approximations, which, in classical digital computers, are implemented using expensive MCMC. Physical computation offers the opportunity to reduce the cost of sampling by building physical systems whose natural dynamics correspond to drawing samples from the desired RBM distribution. Such a system avoids the burn-in and mixing cost of a Markov chain. However, hardware implementations of this variety usually entail limitations such as low-precision and limited range of the parameters and restrictions on the size and topology of the RBM. We conduct software simulations to determine how harmful each of these restrictions is. Our simulations are designed to reproduce aspects of the D-Wave quantum computer, but the issues we investigate arise in most forms of physical computation.
Abstract Modest statistical differences between the sampling performances of the D-Wave quantum annealer (QA) and the classical Markov Chain Monte Carlo (MCMC), when applied to Restricted Boltzmann Machines (RBMs), are explored to explain, and possibly address, the absence of significant and consistent improvements in RBM trainability when the D-Wave sampling was used in previous investigations. A novel hybrid sampling approach, combining the classical and the QA contributions, is investigated as a promising way to benefit from the modest differences between the two sampling methods. No improvements in the RBM training are achieved in this work, thereby suggesting that the differences between the QA-based and MCMC sampling, mainly found in the medium-to-low probability regions of the distribution, which are less important for the quality of the sample, are insufficient to benefit the training. Difficulties in achieving sufficiently high quality of embedding RBMs into the lattice of the newer generation of D-Wave hardware could be further complicating the task. On the other hand, the ability to generate samples of sufficient variety from lower-probability parts of the distribution has a potential to benefit other machine learning applications, such as the mitigation of catastrophic forgetting (CF) during incremental learning. The feasibility of using QA-generated patterns of desirable classes for CF mitigation by the generative replay is demonstrated in this work for the first time. While the efficiency of the CF mitigation using the D-Wave QA was comparable to that of the classical mitigation, both the speed of generating a large number of distinct desirable patterns and the potential for further improvement make this approach promising for a variety of challenging machine learning applications.
Geometric frustration gives rise to emergent quantum phenomena and exotic phases of matter. While Monte Carlo methods are traditionally used to simulate such systems, their sampling efficiency is limited by the complexity of interactions and ground-state properties. Restricted Boltzmann Machines (RBMs), a class of probabilistic neural networks, offer improved sampling by incorporating machine learning techniques. However, fully-connected bipartite RBMs are inefficient for representing physical lattices with sparse interactions. To address this, we implement Convolutional Restricted Boltzmann Machines (CRBMs) that leverage translational symmetry inherent to lattices. Using the classical Shastry-Sutherland (SS) Ising lattice, we demonstrate (i) CRBM formulation that captures SS interactions, and (ii) digital hardware accelerator to enhance sampling performance. We simulate lattices with up to 324 spins, recovering all known phases of the SS Ising model, including the long range ordered fractional plateau. Our hardware characterizes spin behavior at critical points and within spin liquid phases. This implementation achieves a speedup of 3 to 5 orders of magnitude (33 ns to 120 ms) over GPU-based implementations. Moreover, the time-to-solution is within two orders of magnitude of quantum annealers, while offering superior scalability, room-temperature operation and reprogrammability. This work paves a pathway for scalable digital hardware that embeds physical symmetries to enable large scale simulations of material systems.
The possibility of using a quantum computer D-Wave 2X with more than 1000 qubits to determine the global minimum of the energy landscape of trained restricted Boltzmann machines is investigated. In order to overcome the problem of limited interconnectivity in the D-Wave architecture, the proposed RBM embedding combines multiple qubits to represent a particular RBM unit. The results for the lowest-energy (the ground state) and some of the higher-energy states found by the D-Wave 2X were compared with those of the classical simulated annealing (SA) algorithm. In many cases, the D-Wave machine successfully found the same RBM lowest-energy state as that found by SA. In some examples, the D-Wave machine returned a state corresponding to one of the higher-energy local minima found by SA. The inherently nonperfect embedding of the RBM into the Chimera lattice explored in this work (i.e., multiple qubits combined into a single RBM unit were found not to be guaranteed to be all aligned) and the existence of small, persistent biases in the D-Wave hardware may cause a discrepancy between the D-Wave and the SA results. In some of the investigated cases, introduction of a small bias field into the energy function or optimization of the chain-strength parameter in the D-Wave embedding successfully addressed difficulties of the particular RBM embedding. With further development of the D-Wave hardware, the approach will be suitable for much larger numbers of RBM units.
Restricted Boltzmann machines are common machine learning models that can utilize quantum annealing devices in their training processes as quantum samplers. While this approach has shown promise as an alternative to classical sampling methods, the limitations of quantum annealing hardware, such as the number of qubits and the lack of connectivity between the qubits, still pose a barrier to wide-scale adoption. We propose the use of multiple software techniques such as dropout method, passive labeling, and parallelization techniques for addressing these hardware limitations. The study found that using these techniques along with quantum sampling showed comparable results to its classical counterparts in certain contexts, while in others the increased complexity of the sampling process hindered the performance of the trained models. This means that further research into the behavior of quantum sampling needs to be done to apply quantum annealing to training tasks of more complicated RBM models.
Restricted Boltzmann Machines trained by the classical Contrastive Divergence algorithm were embedded into a next generation D-Wave quantum computer having more than 2000 qubits. The embedding was used to sample missing visible units and labels of incomplete test images. The previously introduced RBM embedding that combined qubits into individual RBM units in order to increase the connectivity between units was used. In comparison to the 1000 qubit machine investigated earlier, the same embedding approach applied to the 2000 qubit D-Wave enabled a connectivity between RBM units sufficient for RBM training on a classical computer with a satisfactory classification error. After a varied number of CD training iterations, the visible units corresponding to a particular incomplete test image were fixed in the D-Wave embedding. The resulting state returned by the D-Wave represents the most energetically favorable combination of the remaining qubits for the given fixed incomplete input image. The classification errors compared favorably to those obtained by the classical MCMC sampling. Main factors influencing the quality of the embedding were investigated. Regardless of the computational cost incentives for using the D-Wave for this operation, the reconstruction and classification errors determined in this way can be used at least as a verification of the validity and quality of the particular embedding of the given RBM architecture. This is essential for successfully employing a QC in much more computationally demanding parts of the RBM training. opportunities for using QC not only for recognition but especially for RBM training, in particular for obtaining a representative sample from the RBM model distribution, are discussed.