Pith. sign in

REVIEW 3 major objections 5 minor 22 references

Implementing Large Quantum Boltzmann Machines as Generative AI Models for Dataset Balancing

T0 review · 3 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read The paper claims that a 120-by-120 quantum restricted Boltzmann machine embedded on the Pegasus annealer generated 1.68 million synthetic attack samples and improved intrusion-detection classifiers beyond SMOTE and RandomOversampler.

desk verdict The 120x120 Pegasus embedding is a real engineering contribution, but the balancing evaluation is invalidated by test-set leakage in the reported arithmetic. read the letter →

arxiv 2502.03086 v2 pith:RRPLD6LT submitted 2025-02-05 cs.ET cs.AIcs.LGcs.NEquant-ph

Salvatore Sinno , Markus Bertl , Arati Sahoo , Bhavika Bhalgamiya , Thomas Gro{ss} , Nicholas Chancellor This is my paper · ORCID
classification cs.ETcs.AIcs.LGcs.NEquant-ph
keywords QuantumComputingMachineLearning(QML)RestrictedBoltzmannMachines(QRBM)(RBM)DatasetBalancingUnbalancedDataGenerativeModelsIntrusionDetectionSystems
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that a quantum restricted Boltzmann machine (QRBM) can be made large enough to act as a practical generative model for fixing class imbalance in intrusion detection data. The authors embedded a QRBM with 120 visible and 120 hidden units on the Pegasus quantum annealer using a custom embedding algorithm, trained it on binary-encoded attack samples, and used repeated quantum annealing to synthesize over 1.6 million attack records. Adding these to the original training data produced a balanced set of more than 4.2 million records, and classifiers trained on that set improved precision, recall, and F1 scores relative to the same classifiers trained on data balanced with SMOTE or RandomOversampler. The balancing step itself took 0.33 seconds. If this holds, it would mean current quantum annealing hardware can contribute to a routine but consequential machine-learning preprocessing task, producing synthetic minority-class data faster than classical baselines while improving detection performance.

What carries the argument

The load-bearing object is the paper's custom embedding algorithm (Algorithm 1), which maps the bipartite QRBM onto the Pegasus qubit graph by segmenting visible and hidden nodes according to periodicity parameters and assigning physical qubits in repeating vertical and horizontal bands, with ferromagnetic intra-layer couplings set to -1. It is what lets the authors fit a 120-visible, 120-hidden model on hardware where the vendor's default embedding tool takes hundreds of seconds and produces long chains; the paper reports no chains longer than six qubits and runtimes near 0.014 seconds. The second mechanism is a quantum-annealing training loop (Algorithm 2), adapted from the paper's reference [9], that replaces the Gibbs-sampling model expectation of contrastive divergence with samples drawn from the annealer.

What would settle it

For the claimed 120x120 model, retrieve the programmed coupling matrix from the QPU after embedding and compare bit-for-bit with the J matrix that Algorithm 1 returns; any rerouted or missing edge means the hardware is not sampling from the trained QRBM. A second check is to train a small QRBM with known weights, anneal many samples, and compare visible-unit marginals with Gibbs-sampling marginals of the same model; mismatch beyond shot noise would refute the sampling claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a 120-by-120 restricted Boltzmann machine, embedded on Pegasus by the authors' own mapping routine, samples well enough to serve as a generative model for intrusion-detection data. The RBM energy is converted to an Ising Hamiltonian, the weights are trained with an adapted contrastive-divergence procedure in which the model expectation comes from quantum annealing, and new samples are generated by repeated annealing calls. After preprocessing CICIDS2017 into 48 features and 120-bit vectors, the QRBM produced 1,683,771 synthetic attack samples, yielding a balanced dataset of 4,208,618 records. On this balanced data, Random Forest, Decision Tree, and KNN reached F1 scores near 90-96 percent, marginally above the same classifiers trained on SMOTE- or RandomOversampler-balanced data, and the balancing time was 0.33 seconds.

Load-bearing premise

The load-bearing premise is that the custom mapping places the 120-by-120 model onto the physical qubits faithfully, so that the couplings and chains actually encode the trained RBM's weights and the annealer samples from that model rather than from a distorted graph.

Editorial extensions

If this is right

  • If the central claim holds, a quantum annealer can embed and sample from a 120-by-120 restricted Boltzmann machine in a real preprocessing workflow, not just in proof-of-concept demonstrations.
  • Dataset balancing for intrusion detection can be completed in well under a second, with downstream classifiers matching or exceeding those trained on SMOTE- or RandomOversampler-balanced data.
  • The custom embedding removes the default embedding tool as the bottleneck, making QRBMs up to 172x120 reachable on current hardware.
  • Binarizing network-flow features to a fixed-width bit vector and generating minority-class samples by annealing appears to preserve enough statistical structure to improve precision, recall, and F1 across several classifier families.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A comparison the paper does not run: a classical RBM of the same 120x120 architecture trained by Gibbs sampling on the same binary features. If it matched the QRBM's F1 gains, the advantage would be the generative model rather than the quantum sampler; if not, the annealer is doing measurable work.
  • The embedding routine's parametric structure suggests it transfers to the newer Zephyr topology by changing periodicity parameters, which would give a direct scaling test: how large a bipartite RBM can be embedded as qubit degree grows.
  • The reported 0.33 seconds covers only the balancing step; a full cost comparison with GAN-type generators would need training time, preprocessing, and readout included.
  • Because annealing is stochastic, an obvious extension is to measure variance of downstream F1 across repeated independent annealing runs; if variance is small, QRBM balancing is safe to use in automated pipelines.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes a custom minor-embedding algorithm to place a restricted Boltzmann machine with 120 visible and 120 hidden units on D-Wave's Pegasus architecture, trains it on binarized CICIDS2017 attack samples, and generates roughly 1.68 million synthetic attack records. The resulting balanced dataset is compared against SMOTE and RandomOversampler across seven classifiers, with precision, recall, and F1 reported in Tables VI-IX. The authors claim that this constitutes one of the largest QRBM implementations to date and that QRBM-based balancing significantly improves IDS classifier performance while completing balancing in sub-second time.

Significance. If the evaluation were sound, this would be a notable demonstration of scaling QRBMs to nontrivial sizes on current quantum hardware and applying them to a concrete imbalanced-data problem. The paper's strengths include a concrete, described embedding algorithm (Algorithm 1), the use of real QPU sampling, and a practical dataset. However, the central quantitative claims currently rest on a contaminated evaluation and on an unvalidated embedding, so the significance is conditional on substantial correction.

major comments (3)
  1. [Section IV.D, Table IV] The arithmetic in Section IV.D indicates that the synthetic samples were added to the full dataset, not to the training partition. The paper reports 1,683,771 synthetic attack samples 'creating a perfectly balanced dataset of 4,208,618 samples.' The only way to obtain 4,208,618 from the numbers in Table IV is 2,524,847 (total) + 1,683,771 (synthetic attack). If the synthetic samples had been combined only with the training partition (1,767,393 records), the total would be 3,451,164, and the training attack count would be 294,377 + 1,683,771 = 1,978,148, which is not balanced against 1,473,016 benign training samples. Thus the pipeline appears to balance the entire dataset before evaluation, meaning the test partition (including its 126,161 attack samples) is used to generate QRBM samples and the test set contains synthetic attack samples derived from test examples. This contamination invalidates the classifier metrics in Table IX and the comparison with SMOTE and RandomOversampler.
  2. [Section IV.B, Algorithm 1, Table III] The embedding algorithm is validated only by wall-clock time and by the number of chains longer than six qubits. There is no check that the implemented Ising parameters (h,J) match the trained RBM weights, no report of chain-breakage rates during sampling, and no comparison between hardware samples and samples from the trained classical RBM. Consequently, the claim that the deployed distribution is the trained 120x120 QRBM distribution is not established. A concrete test, such as comparing the empirical visible-unit marginals or the energy statistics of hardware samples against classical Gibbs samples, is needed.
  3. [Tables VI-IX] All reported metrics in Tables VI-IX appear to come from single runs with no error bars, no repeated seeds, and no statistical significance testing. Differences of 0.1-0.5 percentage points (e.g., Random Forest F1 of 95.96 for RandomOversampler versus 96.1 for QRBM; KNN F1 of 90.5 versus 90.91) are not established as meaningful. The paper's claim of 'significantly improving' detection rates therefore is not supported by the presented evidence.
minor comments (5)
  1. [Section IV.C, Section IV.D, Discussion] The timing figures are inconsistent: Section IV.C reports a default annealing time of 20 µs and 124 µs readout, while the Discussion reports QPU sampling time of 134.92 µs and readout time of 94 µs; Section IV.D says each annealing process lasts only 20 µs. These numbers need reconciliation before the millisecond/sub-second timing claims can be assessed.
  2. [Table V] The header 'QRMB' in Table V is a typo; it should read 'QRBM.'
  3. [Section VII] The sentence 'We therefore contribute by demonstrate the practical feasibility...' has a grammatical error; it should be 'by demonstrating.'
  4. [Algorithm 2] In Algorithm 2, lines 9-11 use the unclear denominator 'm1' and the notation V^T H/m versus V'^T H'/m1; either the denominators should be m throughout or the mini-batch size should be defined separately. As written, the update rule is ambiguous.
  5. [Related Work] The Related Work section claims the algorithm enables minor embedding 'up to a 172x120 RBM,' but Table III reports results only for 60x60 and 120x120. The 172x120 claim is not supported by any experimental result in the paper.

Circularity Check

1 steps flagged · score 6.0 of 10

Central QRBM-vs-classical comparison is partially circular: the reported 4,208,618-record balanced dataset is generated from the full deduplicated corpus, so the Table IX test metrics include QRBM samples derived from test records.

  1. fitted input called prediction [Section IV.D (Balancing the Dataset using the QRBM); Table IV train/test split; Section IV methodology description.]
    "Repeating this process 1,700 times resulted in 1,683,771 synthetic attack samples, creating a perfectly balanced dataset of 4,208,618 samples."

    Table IV fixes the total deduplicated dataset at 2,524,847 records (training 1,767,393 + test 757,454; benign 2,104,309 + attack 420,538). The only arithmetic consistent with 4,208,618 is 2,524,847 + 1,683,771 = 2 x 2,104,309, i.e., synthetic attack samples were appended to the entire dataset, not to the training partition alone (which would give 1,767,393 + 1,683,771 = 3,451,164). Because the QRBM was trained on binarized attack data from this full corpus, it saw the 126,161 test attack records, and the classifier metrics in Table IX are computed on a test set that contains attack records and QRBM-generated samples derived from exactly those records.

full rationale

The non-leakage parts of the derivation are self-contained. Algorithm 2 (QRBM training) is adapted from external reference [9], and Algorithm 1 is specified in full in the appendix and benchmarked against minorminer on runtime and chain length; no fitted parameter is renamed as a prediction there. The self-citation [8] (Dattani, Szalay, Chancellor) supplies background on the Pegasus connectivity graph but is not the sole load-bearing justification for the embedding claim, because the algorithm and its outputs are described in the paper. The embedding correctness issue (no verification that implemented J/h match the intended RBM weights) is a validity risk, not circularity. The central comparative claim, however, reduces to an evaluation on data generated from the evaluation set: the reported 4,208,618-sample ‘perfectly balanced dataset’ equals the total pre-split corpus plus the synthetic samples, so the test partition used for Table IX was contaminated by construction. This makes the QRBM performance comparison with SMOTE and RandomOversampler partially circular and not evaluable as stated.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim rests on a binarization and feature-selection pipeline, an unvalidated embedding, and an assumed relationship between D-Wave samples and the RBM distribution. These are domain assumptions and ad hoc choices rather than free parameters in the physics sense, but they are load-bearing for the reported results.

free parameters (5)
  • 120-bit binarization width N = 120
    Equation 8 selects N based on feature ranges; N=120 determines the QRBM visible-layer size and is a hand-chosen capacity parameter.
  • Feature correlation threshold = 0.9 (absolute)
    Features with absolute correlation >= 0.9 were removed; the threshold is chosen by the authors, not justified in the text.
  • Embedding periodicity parameters = not reported
    Algorithm 1 depends on periodicity_v, periodicity_h, and n_periodicity, which are not specified in the paper.
  • QRBM training hyperparameters = not reported
    Learning rate, batch size, epochs, and convergence criteria in Algorithm 2 are omitted; they affect the trained weights and thus the synthetic samples.
  • Number of synthetic samples generated = 1,683,771
    The number of generated samples (1,700 rounds x 1,000) is a deliberate choice that determines the final dataset size and downstream metrics.
assumptions (5)
  • domain assumption The D-Wave annealer samples from a distribution close to the Boltzmann distribution of the Ising model (Eq. 7).
    Section III.C assumes quantum annealing efficiently samples low-energy configurations; in practice, noise and finite temperature bias the samples, and the paper does not calibrate this.
  • domain assumption CICIDS2017 labels are ground truth.
    The paper uses the dataset's class labels without independent verification; standard practice for benchmark datasets, but still an assumption.
  • ad hoc to paper The 120-bit binarized representation preserves the information needed to synthesize realistic attack samples.
    The encoding via Equation 8 is a custom choice; no analysis shows that the binarized features are sufficient for the generative task.
  • ad hoc to paper Algorithm 1 produces a valid minor embedding that preserves the bipartite RBM structure.
    The embedding is not proven correct; validation is limited to chain-length counts and speed comparisons in Table III.
  • standard math Contrastive divergence training with a quantum-sampled negative phase converges to a useful model.
    CD is a standard approximation; the paper adapts it from [9] without modification.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Implementing Large Quantum Boltzmann Machines as Generative AI Models for Dataset Balancing." pith.science (2026). https://pith.science/paper/RRPLD6LT

@misc{pith2026250203086,
  author       = {Pith},
  title        = {Pith review of: Implementing Large Quantum Boltzmann Machines as Generative AI Models for Dataset Balancing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RRPLD6LT}},
  note         = {Machine review of arXiv:2502.03086}
}
read the original abstract

This study explores the implementation of large Quantum Restricted Boltzmann Machines (QRBMs), a key advancement in Quantum Machine Learning (QML), as generative models on D-Wave's Pegasus quantum hardware to address dataset imbalance in Intrusion Detection Systems (IDS). By leveraging Pegasus's enhanced connectivity and computational capabilities, a QRBM with 120 visible and 120 hidden units was successfully embedded, surpassing the limitations of default embedding tools. The QRBM synthesized over 1.6 million attack samples, achieving a balanced dataset of over 4.2 million records. Comparative evaluations with traditional balancing methods, such as SMOTE and RandomOversampler, revealed that QRBMs produced higher-quality synthetic samples, significantly improving detection rates, precision, recall, and F1 score across diverse classifiers. The study underscores the scalability and efficiency of QRBMs, completing balancing tasks in milliseconds. These findings highlight the transformative potential of QML and QRBMs as next-generation tools in data preprocessing, offering robust solutions for complex computational challenges in modern information systems.

Figures

Figures reproduced from arXiv: 2502.03086 by the authors.

Figure 1
Figure 1. Flow chart for balancing a dataset using QRBM. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

22 extracted references · 19 canonical work pages

  1. [1]

    To- ward constructing a balanced intrusion detection dataset based on cicids2017

    Amer Abdulmajeed Abdulrahman and Mahmood Khalel Ibrahem. To- ward constructing a balanced intrusion detection dataset based on cicids2017. Samarra Journal of Pure and Applied Science , 2(3):132– 142, 2020

    work page 2020

  2. [2]

    Application of quan- tum annealing to training of deep neural networks

    Shin Adachi and Maxwell P Henderson. Application of quan- tum annealing to training of deep neural networks. arXiv preprint arXiv:1510.06356, 2015

    arXiv 2015

  3. [3]

    Quantum boltzmann machine

    Mohammad H Amin, Evgeny Andriyash, Jason Rolfe, Bohdan Kulchyt- skyy, and Roger Melko. Quantum boltzmann machine. Physical Review X, 8(2):021050, 2018

    work page 2018

  4. [4]

    Estimation of effective temperatures in quantum anneal- ers for sampling applications: A case study with restricted boltzmann machines

    Marcello Benedetti, John Realpe-Gomez, Rupak Biswas, and Alejandro Perdomo-Ortiz. Estimation of effective temperatures in quantum anneal- ers for sampling applications: A case study with restricted boltzmann machines. Physical Review A , 94(2):022308, 2016

    work page 2016

  5. [5]

    The ising model: teaching an old problem new tricks

    Zhengbing Bian, Fabian Chudak, William G Macready, and Geordie Rose. The ising model: teaching an old problem new tricks. D-wave systems, 2:1–32, 2010

    work page 2010

  6. [6]

    Explaining the gibbs sampler

    George Casella and Edward I George. Explaining the gibbs sampler. The American Statistician , 46(3):167–174, 1992

    work page 1992

  7. [7]

    Smote: synthetic minority over-sampling technique

    Nitesh V Chawla, Kevin W Bowyer, Lawrence O Hall, and W Philip Kegelmeyer. Smote: synthetic minority over-sampling technique. Jour- nal of artificial intelligence research , 16:321–357, 2002

    work page 2002

  8. [8]

    Pegasus: The second connectivity graph for large-scale quantum annealing hardware

    Nike Dattani, Szilard Szalay, and Nicholas Chancellor. Pegasus: The second connectivity graph for large-scale quantum annealing hardware. arXiv preprint arXiv:1901.07636 , 2019

    arXiv 1901

Show all 22 references
  1. [9]

    Quantum-assisted training of restricted boltzmann machines

    Akash Dixit, Rahul Sharma, and Priya Singh. Quantum-assisted training of restricted boltzmann machines. Quantum Information Processing , 22(5):135–150, 2023

    2023

  2. [10]

    On the effectiveness of preprocessing methods when dealing with different levels of class imbalance

    Vicente Garc ´ıa, Jos ´e Salvador S ´anchez, and Ram ´on Alberto Mollineda. On the effectiveness of preprocessing methods when dealing with different levels of class imbalance. Knowledge-Based Systems, 25(1):13– 21, 2012

    2012

  3. [11]

    Balancing approaches towards ml for ids: a survey for the cse-cic ids dataset

    Subiksha Srinivasa Gopalan, Dharshini Ravikumar, Dino Linekar, Ali Raza, and Maheen Hasib. Balancing approaches towards ml for ids: a survey for the cse-cic ids dataset. In 2020 International Conference on Communications, Signal Processing, and their Applications (ICCSPA) , pa...

    2020

  4. [12]

    Learning from imbalanced data

    Haibo He and Edwardo A Garcia. Learning from imbalanced data. IEEE Transactions on knowledge and data engineering , 21(9):1263– 1284, 2008

    2008

  5. [13]

    Training products of experts by minimizing contrastive divergence

    Geoffrey E Hinton. Training products of experts by minimizing contrastive divergence. Neural computation, 14(8):1771–1800, 2002

    2002

  6. [14]

    A practical guide to training restricted boltzmann machines

    Geoffrey E Hinton. A practical guide to training restricted boltzmann machines. Momentum, 9:1–10, 2012

    2012

  7. [15]

    Reducing the dimen- sionality of data with neural networks

    Geoffrey E Hinton and Ruslan R Salakhutdinov. Reducing the dimen- sionality of data with neural networks. Science, 313(5786):504–507, 2006

    2006

  8. [16]

    Pegasus topology, 2020

    D-Wave Systems Inc. Pegasus topology, 2020

    2020

  9. [17]

    Characterization of tor traffic using time based features

    Arash Habibi Lashkari, Gilberto Draper-Gil, Md Mamun, and Ali A Ghorbani. Characterization of tor traffic using time based features. Proceedings of the 3rd International Conference on Information Systems Security and Privacy , pages 253–262, 2017

    2017

  10. [18]

    Data-balancing algorithm based on generative adversarial network for robust network intrusion detection

    I-Hsien Liu, Cheng-En Hsieh, Wei-Min Lin, Jung-Shian Li, and Chu-Fen Li. Data-balancing algorithm based on generative adversarial network for robust network intrusion detection. Journal of Robotics, Networking and Artificial Life , 9(3):303–308, 2022

    2022

  11. [19]

    Machine learning with oversampling and undersampling techniques: overview study and experimental results

    Roweida Mohammed, Jumanah Rawashdeh, and Malak Abdullah. Machine learning with oversampling and undersampling techniques: overview study and experimental results. In 2020 11th international conference on information and communication systems (ICICS) , pages 243–248. IEEE, 2020

    2020

  12. [20]

    4-clique network minor embedding for quantum annealers

    Elijah Pelofske. 4-clique network minor embedding for quantum annealers. arXiv preprint arXiv:2301.08807 , 2023

    2023 arXiv

  13. [21]

    To- ward generating a new intrusion detection dataset and intrusion traffic characterization

    Iman Sharafaldin, Arash Habibi Lashkari, and Ali A Ghorbani. To- ward generating a new intrusion detection dataset and intrusion traffic characterization. ICISSp, 1:108–116, 2018

    2018

  14. [22]

    Quan- tum optimization of fully connected spin glasses

    Davide Venturelli, Dominique J Marchand, and Gian-Luca Rojo. Quan- tum optimization of fully connected spin glasses. Physical Review X , 5(3):031040, 2015. APPENDIX A. Algorithms Algorithm 1 Embedding a QRBM on Pegasus 1: Input: n visible, n hidden, periodicity v, periodicity ...

    2015

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.