Pith. sign in

REVIEW 4 major objections 5 minor 55 references

A network of privacy-preserving Tsetlin agents can—by exchanging only one-bit clause outputs with their neighbors—match the accuracy of a centralized model on vertically partitioned data.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 10:41 UTC pith:B5AC7ONZ

load-bearing objection Decentralized TM ensemble with a two-layer hierarchy is a genuinely new combination, but the headline accuracy claim rests on a single unperturbed sensor run and an apples-to-oranges baseline comparison. the 4 major comments →

arxiv 2607.20124 v1 pith:B5AC7ONZ submitted 2026-07-22 cs.LG cs.MA

Autonomous Collaborative Learning Among an Ensemble of Tsetlin Machines with Consensus-Based Inference

Yehuda Rudin , Osnat Keren , Michal Yemini , Alexander Fish This is my paper
classification cs.LG cs.MA
keywords Tsetlin Machinedecentralized learningvertical federated learningconsensus-based inferencegossip protocoledge computingprivacy-preserving machine learninglogical clauses
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper tries to establish that decentralized collaborative learning with Tsetlin Machines works without a central server and without sharing raw data or model parameters. Each agent trains a small local Tsetlin Machine on its private feature slice, then a second-layer Tsetlin Machine learns to read the one-bit clause outputs of its neighbors, and a gossip protocol diffuses vote sums until every agent reaches the same global prediction. The support is experimental: on a synthetic 25-agent sensor network the distributed ensemble reaches 93% accuracy, matching a centralized ANN, while on MNIST and Fashion-MNIST the two-layer ensemble reaches 93–94% and 81–84%, approaching but not matching a classic Tsetlin Machine. The authors are careful to state that these results assume ideal communication channels and that no formal privacy-leakage analysis is provided.

Core claim

The paper's central claim is that a two-layer hierarchical Tsetlin Machine—an Input Layer that learns from local Boolean features and a Neighborhood Aggregation Layer that learns from binary clause outputs arriving from neighboring agents—together with a gossip-based vote-sum dissemination protocol, lets a decentralized ensemble of heterogeneous agents classify data as accurately as a centralized model. The claim is supported by accuracy measurements showing parity with a centralized ANN on a sensor network and near-parity with a classic TM on image benchmarks, and by a comparison showing the learned aggregation layer outperforms simple vote-sum aggregation.

What carries the argument

The load-bearing object is the two-layer TM agent architecture. A Tsetlin Machine is a rule-based classifier in which teams of two-action Tsetlin Automata build conjunctive logical clauses over Boolean literals; here, each agent's first-layer TM produces clause outputs that can be encoded as single bits, and the second-layer TM receives those bits from the agent and its eight immediate neighbors as its feature input. The second layer learns correlations between local clause activity and the cluster's input patterns, so a global decision emerges without any central aggregator. The gossip scheme in Algorithm 2 completes the mechanism: vote sums are broadcast, re-broadcast when new information

Load-bearing premise

The results assume ideal, error-free, latency-free communication and that every agent holds the ground-truth label for every training sample; if real channels drop packets, Algorithm 1 skips the aggregation-layer update and the consensus property can degrade.

What would settle it

Run the 25-agent sensor-network experiment with a controlled packet-loss rate—say 10%, 30%, and 50%—and plot accuracy against the rate. The paper's Algorithm 1 explicitly omits aggregation-layer training when a packet is lost, so a sharp accuracy collapse with even modest loss would show that the 'comparable to centralized' claim depends on the ideal-channel assumption. A second check: intercept one agent's clause-output stream and attempt to reconstruct the local tile's class structure; success would falsify the confidentiality claim.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If the claim holds, vertically partitioned data can be learned across a sensor network with no central server and no raw-data transfer, which supports privacy-conscious edge deployments.
  • Per-clause outputs being single bits means neighbor communication is extremely cheap, so the method is compatible with energy-limited devices.
  • Heterogeneous agents—different sensor counts, noise levels, or computational budgets—can participate in one ensemble without architectural alignment.
  • The learned aggregation layer systematically beats non-learned vote-sum aggregation, implying that the second-layer TM captures genuine cross-agent structure.
  • Accuracy improves as more agents join the network and approaches centralized performance, so larger deployments should help rather than hurt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's privacy claim is conditional: it shows that raw data and model parameters are not exchanged, but it does not analyze whether binary clause-output trajectories leak information about local data; a reconstruction attack on clause outputs would be a natural stress test.
  • Because the gossip protocol is proven to terminate with full information on connected graphs under ideal channels, the accuracy results should transfer to any connected topology, but real packet loss breaks the aggregation-layer training step; simulating random aggregation-skip rates is a cheap way to predict real-world degradation.
  • The MNIST/Fashion-MNIST gap to a classic TM suggests the bottleneck may be the small per-agent clause budgets or shallow depth; scaling clauses per layer, or adding a third aggregation tier, is a testable path toward closing that gap.
  • The consensus layer could be reused not just for classification but for other TM-based tasks such as regression or anomaly detection in distributed sensing, since the mechanism only depends on exchanging clause outputs.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes a decentralized collaborative learning framework for an ensemble of Tsetlin Machines (TMs) under vertical feature partitioning. Each agent trains a first-layer TM on its local Boolean features and a second-layer 'neighborhood aggregation' TM on the binary clause outputs from itself and its neighbors, with no raw data or model parameters exchanged. Inference combines class vote sums from all agents through a gossip-based consensus protocol. The authors report experiments on MNIST, Fashion-MNIST, and a synthetic heterogeneous sensor network, claiming that the resulting accuracies are 'comparable to those of centralized models.' The main positive result is that the two-layer architecture consistently outperforms simple vote-sum aggregation in the MNIST and Fashion-MNIST tables; however, the comparison to centralized/classic models is weakened by single-run results, missing error bars, and baselines trained under different conditions.

Significance. If fully supported, the proposed architecture would be a useful contribution to privacy-preserving and communication-efficient decentralized learning with Tsetlin Machines, particularly in edge and sensor-network settings. The idea of learning a second-layer TM over neighboring agents' clause outputs is novel, and the paper's demonstration that learned aggregation beats vote-sum aggregation is valuable. The claimed parity with centralized models, however, is the central selling point, and the current evidence does not establish it: the benchmark gaps are 2–4 percentage points against baselines trained for twice as many epochs, and the single synthetic parity result has no variance estimates. The paper also honestly acknowledges its ideal-communication assumption, but that limitation has not been stress-tested. Overall, the architecture is interesting and worth further investigation, but the headline claim requires substantially stronger empirical support.

major comments (4)
  1. [Abstract; §IV.A, Tables IV–V; §IV.B, Tables VI–VII] The abstract and §V claim accuracies 'comparable to those of centralized models,' but the benchmark evidence does not support this. Best ensemble results are 94.40–94.65% on MNIST and 83.85–83.99% on Fashion-MNIST, versus 96.73% and 87.75% for the classic TM baselines cited from [50], which were trained for 100 epochs while the ensemble was trained for 50. No repeated runs or error bars are reported, so the 2–4 point gaps cannot be assessed statistically. Please report like-for-like comparisons (same epoch count and baselines under the same protocol) and multiple random seeds with mean and standard deviation.
  2. [§IV.C, Table VIII, Fig. 7, Table X] The only direct parity result with a centralized model is the 25-agent sensor network (93% vs 93%), but it is based on a single synthetic dataset realization and a single run. No confidence intervals, standard deviations, or sensitivity to the random data-generation process are given. Fig. 7's monotone convergence is a single trajectory. Table X likewise reports 76% vs 80% for the 9-agent connected graph and calls this 'comparable' without error bars. The parity claim needs multiple dataset draws and seeds, with the reported variation.
  3. [§III.A, Algorithm 1, lines 11–15] The 'alt class label' selection is underspecified. It is stated that 'the different class label is known to all agents,' but the mechanism is not defined: is it sampled uniformly among all classes except the true label, and do all agents use the same alt label for the same training sample? Since the aggregation-layer TM receives neighbor clause outputs for both the true-label and alt-label TM pairs, inconsistent alt-label choices across agents could inject noise into the aggregation-layer input and affect results. Please specify the selection rule and its rationale.
  4. [§III.B and §V] The learning and consensus protocols assume ideal, error-free, latency-free communication. The paper acknowledges this and defers realistic communication to future work, which is a legitimate scope choice. However, the algorithm's robustness to packet loss, delays, or partial updates is not examined even at a high level, and the complexity statements ("all agents become fully informed in Θ(n) time," "total number of broadcast events is Θ(n^2)") are stated without proof or a precise scheduler model. Since the paper's contribution is explicitly a distributed/decentralized algorithm, please provide a short derivation or a precise citation supporting these bounds, and clarify what happens to the consensus guarantee when messages are lost.
minor comments (5)
  1. [Tables IV–VII] The caption "Accuracy of a classic TM ... s=7,5" contains a typo (comma instead of decimal point: s=7.5).
  2. [Table VIII] The F1 entry for the single-agent ANN is printed as '60' instead of '0.60'.
  3. [Fig. 3 and Fig. 4] The tile vote figures are hard to parse because the numbers are embedded in a grid without a color scale or a clear separation between the class-sum rows. Consider formatting these as heatmaps with numeric labels or using separate panels for each class.
  4. [§IV.C] The binarization and sensor-generation process is described algorithmically but not all hyperparameters are specified (e.g., exact bin counts, number of training samples per agent). Please provide the full experimental setup or a reproducibility appendix.
  5. [§VI] Only data are promised 'upon reasonable request.' No code release is mentioned. Given that the claims rely on empirical results from an unusual architecture, I would strongly encourage the authors to release code to enable independent verification.

Circularity Check

0 steps flagged

No significant circularity: the paper is an empirical benchmark study; no prediction reduces by construction to a fitted parameter, and no load-bearing self-citation is present.

full rationale

The paper's central claim is empirical: a decentralized two-layer Tsetlin Machine ensemble with consensus-based inference achieves accuracies 'comparable to those of centralized models' on MNIST, Fashion-MNIST, and a synthetic sensor network. The derivation chain consists of standard TM training (Algorithm 1), a gossip-based vote dissemination protocol (Algorithm 2), and hand-specified hyperparameter configurations. No equation is fitted to a target quantity and then reported as a prediction; the accuracy tables are direct experimental measurements against external baselines (classic TM from [50]/[42], centralized ANN). Hyperparameters are enumerated as configuration variants rather than fitted to test accuracy, so the 'fitted input called prediction' pattern does not apply. The consensus procedure in Algorithm 2 is definitional to the protocol—a node is 'fully informed' when it has all vote sums—but the paper does not present this as a theoretical prediction; it is a protocol specification. There are no author-overlapping self-citations in the reference list, and no uniqueness theorem or prior-work ansatz is imported to force the architecture. The paper explicitly defers realistic communication impairments and formal leakage analysis to future work; these are scope limitations, not circular inputs. Thus the central empirical claim, while possibly under-supported by the small number of runs and unequal training epochs, is not circular. Honest non-finding: score 0.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central empirical claims rest on several hand-chosen hyperparameters and on domain assumptions about label availability, ideal communication, and TM convergence. No new physical entities are introduced; the only new component is the aggregation-layer TM, which is evaluated only within the paper.

free parameters (5)
  • Input layer TM hyperparameters (clauses, T, s) = e.g., 32 clauses, T=3, s=3 (sensor); 16–128 clauses, T=3–8, s=3–10 (MNIST)
    Chosen by hand per experiment; directly control the learned clauses and accuracy.
  • Aggregation layer TM hyperparameters (clauses, T, s) = 16–64 clauses, T=3–6, s=4–10
    Chosen by hand per experiment; control the aggregation quality.
  • Number of training epochs = 20 or 50
    Results are reported after fixed epochs; accuracy depends on this choice.
  • Binarization bit-width per sensor = 8 bits per sensor
    Number of Boolean features per agent; affects input granularity and model size.
  • Neighborhood radius / grid connectivity = 8 immediate neighbors (3x3) on toroidal grid; graph with 1–4 neighbors
    Defines the receptive field of the aggregation layer and the gossip network.
axioms (4)
  • domain assumption Tsetlin Machine learning dynamics converge to accurate conjunctive clauses under Type I/II feedback
    Relies on cited TM literature ([1], [42]); not re-derived or formally verified here.
  • domain assumption Training samples are vertically aligned and class labels are available to all agents
    Stated in Section III; required for every agent to run Algorithm 1.
  • ad hoc to paper The communication graph is fixed, connected, and communication is error-free and latency-free
    Assumed in Section IV; packet loss can skip aggregation-layer updates (Algorithm 1) and break consensus (Algorithm 2); deferred to future work.
  • domain assumption Quantile binning thresholds chosen as approximately evenly spaced order statistics of observed feature values
    Data-dependent preprocessing used in the sensor network experiment.
invented entities (1)
  • Neighborhood Aggregation Layer TM no independent evidence
    purpose: Learns to combine clause outputs from an agent and its immediate neighbors to produce a more global prediction
    This is the paper's new architectural component; its effectiveness is shown only through in-paper experiments (Tables IV–VII), not through an independent external handle.

pith-pipeline@v1.3.0-alltime-deepseek · 21919 in / 13420 out tokens · 126252 ms · 2026-08-01T10:41:21.693820+00:00 · methodology

0 comments
read the original abstract

Tsetlin Machine (TM) is a rule-based machine-learning algorithm comprising collectives of two-action Tsetlin Automata (TAs) that cooperatively form conjunctive logical clauses from Boolean inputs through stochastic feedback. Although few recent studies have examined TM Federated Learning, the broader area of distributed and decentralized TM learning has not received much attention in the existing literature and warrants further exploration. In this work, we propose a paradigm for decentralized collaborative learning under a vertical feature-partitioning setting among an ensemble of Tsetlin Machines using consensus-based inference. Within this decentralized paradigm, each agent maintains its own private TM model, and there is no exchange of raw data among agents. Inference combines individual agents model predictions into a global consensus. The paradigm accommodates heterogeneous TM-based agents with differing data acquisition means, local data distributions, or computational resources, thereby facilitating the integration and fusion of information in settings such as multi-modal sensing environments. Experiments conducted using two-dimensional grid and connected graph network topologies demonstrate that the classification accuracies achieved are comparable to those of centralized models.

Figures

Figures reproduced from arXiv: 2607.20124 by Alexander Fish, Michal Yemini, Osnat Keren, Yehuda Rudin.

Figure 1
Figure 1. Figure 1: Architecture of an Agent. scale sensor networks and multi-modal sensing en￾vironments. • We propose an inter-agent asynchronous gossip￾based communication protocol for coordination and data exchange. • We analyze the model’s performance, using soft￾ware implementation, on the MNIST and Fashion￾MNIST datasets, partitioned among several agents. • We analyze the model’s performance on a synthetic heterogeneou… view at source ↗
Figure 2
Figure 2. Figure 2: The Tsetlin Machine. (a) Block Diagram of a Tsetlin Machine. (b) State diagram of Tsetlin learning [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: MNIST benchmark: Class vote sums. The number [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fashion-MNIST benchmark: agents’ voting be [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sensor data distribution for 9 agents. representing the global feature set (all 25 agents). When data from all agents are pooled, a global class structure emerges in the 2D feature space. Class 0 Class 1 Class 2 18.2 cm 3.1cm 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1… view at source ↗
Figure 7
Figure 7. Figure 7: Classification accuracy as a function of the [PITH_FULL_IMAGE:figures/full_fig_p011_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: In this setting, agents possess one, three, or [PITH_FULL_IMAGE:figures/full_fig_p012_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

55 extracted references · 2 canonical work pages

  1. [1]

    The tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic,

    O.-C. Granmo, “The tsetlin machine - a game theoretic bandit driven approach to optimal pattern recognition with propositional logic,”arXiv preprint arXiv:1804.01508, 2018

    Pith/arXiv arXiv 2018

  2. [2]

    Convolutional tsetlin machine: Learning interpretable representations through convolutional clauses,

    O.-C. Granmo and L. Jiao, “Convolutional tsetlin machine: Learning interpretable representations through convolutional clauses,” inInternational Joint Conference on Neural Networks (IJCNN), 2019

    2019

  3. [3]

    The regression tsetlin machine: A tsetlin machine for continuous output problems,

    K. D. Abeyrathna, O.-C. Granmo, L. Jiao, and M. Goodwin, “The regression tsetlin machine: A tsetlin machine for continuous output problems,”arXiv preprint arXiv:1905.04206, 2019

    Pith/arXiv arXiv 1905

  4. [4]

    Multi-class tsetlin machine,

    L. Jiao, K. D. Abeyrathna, and O.-C. Granmo, “Multi-class tsetlin machine,”arXiv preprint arXiv:2001.08227, 2020

    Pith/arXiv arXiv 2001

  5. [5]

    Weighted tsetlin machines: Towards interpretable accurate low- complexity classifiers,

    K. D. Abeyrathna, O.-C. Granmo, L. Jiao, and M. Goodwin, “Weighted tsetlin machines: Towards interpretable accurate low- complexity classifiers,” inIEEE Symposium Series on Computa- tional Intelligence (SSCI), 2020

    2020

  6. [6]

    Integer weighted tsetlin machines,

    K. D. Abeyrathna, O.-C. Granmo, and M. Goodwin, “Integer weighted tsetlin machines,”arXiv preprint arXiv:2101.10347, 2021

    Pith/arXiv arXiv 2021

  7. [7]

    Coalesced multi- output tsetlin machine,

    K. D. Abeyrathna, L. Jiao, and O.-C. Granmo, “Coalesced multi- output tsetlin machine,” inInternational Joint Conference on Neural Networks (IJCNN), 2021

    2021

  8. [8]

    Continuous input tsetlin machine,

    V . Yadav, O.-C. Granmo, and L. Jiao, “Continuous input tsetlin machine,”arXiv preprint arXiv:2101.10346, 2021

    Pith/arXiv arXiv 2021

  9. [9]

    The recur- rent tsetlin machine,

    K. D. Abeyrathna, O.-C. Granmo, and M. Goodwin, “The recur- rent tsetlin machine,”arXiv preprint arXiv:2104.01409, 2021

    Pith/arXiv arXiv 2021

  10. [10]

    The fuzzy tsetlin ma- chine,

    V . Yadav, O.-C. Granmo, and L. Jiao, “The fuzzy tsetlin ma- chine,”arXiv preprint arXiv:2201.04977, 2022

    Pith/arXiv arXiv 2022

  11. [11]

    Relational tsetlin ma- chines,

    A. Yazidi, O.-C. Granmo, and L. Jiao, “Relational tsetlin ma- chines,” inInternational Workshop on Mining and Learning with Graphs (MLG), 2022

    2022

  12. [12]

    Multi- layer tsetlin machine: Architecture and performance evaluation,

    O. Tarasyuk, A. Gorbenko, R. Shafik, and A. Yakovlev, “Multi- layer tsetlin machine: Architecture and performance evaluation,” inInternational Symposium on Tsetlin Machines (ISTM), 2024

    2024

  13. [13]

    The probabilistic tsetlin machine: A novel approach to uncertainty quantification,

    D. A. Kuruge, S. El Mekkaoui, A. Hafver, and C. Agrell, “The probabilistic tsetlin machine: A novel approach to uncertainty quantification,” inProceedings of the 2024 8th International Conference on Advances in Artificial Intelligence. ACM, Oct. 2024, p. 39–47. [Online]. Available: http://dx.doi.org/10.1145/3704137.3704143

    arXiv 2024

  14. [14]

    Stochastic and deterministic processes in asymmetric tsetlin machine,

    N. Elmisadr, M.-B. Belaid, and A. Yazidi, “Stochastic and deterministic processes in asymmetric tsetlin machine,”Frontiers in Artificial Intelligence, vol. V olume 8 - 2025, 2025. [On- line]. Available: https://www.frontiersin.org/journals/artificial- intelligence/articles/10.3389/frai.2025.1377944

    arXiv 2025

  15. [15]

    The tsetlin machine goes deep: Logical learning and reasoning with graphs,

    O.-C. Granmo, Y . Abdelwahab, P.-A. Andersen, K. A. K. Borgersen, P. F. A. Clarke, K. Dumbre, Y . Grønningsæter, V . Halenka, R. Helin, L. Jiao, A. Khalid, R. Omslandseter, R. Saha, M. Shende, and X. Zhang, “The tsetlin machine goes deep: Logical learning and reasoning with graphs,” 2026. [Online]. Available: https://arxiv.org/abs/2507.14874

    Pith/arXiv arXiv 2026

  16. [16]

    M. L. Tsetlin,Automation Theory and Modelling of Biological Systems, ser. Mathematics in Science and Engineering. Academic Press, 1974. [Online]. Available: https://books.google.co.il/books?id=3wLEDm bnsC

    1974

  17. [17]

    D. A. P. V . I. Varshavsky,Puppets Without Strings. MIR, 1988

    1988

  18. [18]

    Finite automata and models of simple forms of behaviour,

    M. L. Tsetlin, “Finite automata and models of simple forms of behaviour,”Russian Mathematical Surveys, vol. 18, no. 4, p. 1, aug 1963. [Online]. Available: https://dx.doi.org/10.1070/RM1963v018n04ABEH001139

    work page doi:10.1070/rm1963v018n04abeh001139 1963

  19. [19]

    The biological princi- ples of swarm intelligence,

    S. Garnier, J. Gautrais, and G. Theraulaz, “The biological princi- ples of swarm intelligence,”Swarm Intelligence, vol. 1, pp. 3–31, 10 2007

    2007

  20. [20]

    From arithmetic to logic based ai: A comparative analysis of neural networks and tsetlin machine,

    J. Lei, A. Wheeldon, R. A. Shafik, A. Yakovlev, and O.-C. Granmo, “From arithmetic to logic based ai: A comparative analysis of neural networks and tsetlin machine,”2020 27th IEEE International Conference on Electronics, Circuits and Systems (ICECS), pp. 1–4, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:230513121

    2020

  21. [21]

    Learning automata based energy-efficient AI hardware design for IoT applications,

    A. Wheeldon, R. A. Shafik, Tousif, Rahman, J. Lei, A. Yakovlev, and O.-C. Granmo, “Learning automata based energy-efficient AI hardware design for IoT applications,” Philosophical transactions. Series A, Mathematical, physical, and engineering sciences, vol. 378, 2020. [Online]. Available: https://api.semanticscholar.org/CorpusID:221244332

    2020

  22. [22]

    Adaptive intelligence for batteryless sensors using software- accelerated Tsetlin machines,

    A. Bakar, T. Rahman, R. Shafik, F. Kawsar, and A. Montanari, “Adaptive intelligence for batteryless sensors using software- accelerated Tsetlin machines,” inProceedings of the 20th ACM Conference on Embedded Networked Sensor Systems, ser. SenSys ’22. New York, NY , USA: Association for Computing Machinery, 2023, p. 236–249. [Online]. Available: https://do...

    arXiv 2023

  23. [23]

    Federated learning: Collaborative machine learning without centralized training data,

    A. AbhishekV , S. Binny, R. JohanT, N. Raj, and V . Thomas, “Federated learning: Collaborative machine learning without centralized training data,”international journal of engineering technology and management sciences, 2022. [Online]. Available: https://api.semanticscholar.org/CorpusID:251659795

    2022

  24. [24]

    Communication-efficient learning of deep networks from decentralized data,

    H. B. McMahan, E. Moore, D. Ramage, S. Hampson, and B. A. y Arcas, “Communication-efficient learning of deep networks from decentralized data,” 2023. [Online]. Available: https://arxiv.org/abs/1602.05629

    Pith/arXiv arXiv 2023

  25. [25]

    Federated learning: A cutting-edge survey of the latest advancements and applications,

    A. Akhtarshenas, M. A. Vahedifar, N. Ayoobi, B. Maham, T. Al- izadeh, S. Ebrahimi, and D. L ´opez-P´erez, “Federated learning: A cutting-edge survey of the latest advancements and applications,”

  26. [26]

    Federated machine learning: Concept and applications,

    Q. Yang, Y . Liu, T. Chen, and Y . Tong, “Federated machine learning: Concept and applications,” 2019. [Online]. Available: https://arxiv.org/abs/1902.04885

    Pith/arXiv arXiv 2019

  27. [27]

    Multi-modal anchor gated transformer with knowledge distillation for emotion recognition in conversation,

    J. Li, S. Ding, L. Guo, and X. Li, “Multi-modal anchor gated transformer with knowledge distillation for emotion recognition in conversation,” inProceedings of the Thirty- ThirdInternational Joint Conference on Artificial Intelligence. International Joint Conferences on Artificial Intelligence Organization, Aug. 2024, p. 8141–8149. [Online]. Available: ht...

    work page doi:10.24963/ijcai.2024/905 2024

  28. [28]

    Vertical federated knowledge transfer via representation distillation for healthcare collaboration networks,

    C.-j. Huang, L. Wang, and X. Han, “Vertical federated knowledge transfer via representation distillation for healthcare collaboration networks,” inProceedings of the ACM Web Conference 2023. ACM, Apr. 2023, p. 4188–4199. [Online]. Available: http://dx.doi.org/10.1145/3543507.3583874

    arXiv 2023

  29. [29]

    Vertical federated learning in practice: The good, the bad, and the ugly,

    Z. Wu, Z. Qin, J. Hou, H. Zhao, Q. Li, B. He, and L. Fan, “Vertical federated learning in practice: The good, the bad, and the ugly,” 2025. [Online]. Available: https://arxiv.org/abs/2502.08160

    Pith/arXiv arXiv 2025

  30. [30]

    Fate: An industrial grade platform for collaborative learning with data protection,

    Y . Liu, T. Fan, T. Chen, Q. Xu, and Q. Yang, “Fate: An industrial grade platform for collaborative learning with data protection,” Journal of Machine Learning Research, vol. 22, no. 226, pp. 1–6,

  31. [31]

    Advances and open problems in federated learning,

    P. Kairouz, H. B. McMahan, B. Avent, A. Bellet, M. Bennis, A. N. Bhagoji, K. Bonawitz, Z. Charles, G. Cormode, R. Cummings, R. G. L. D’Oliveira, H. Eichner, S. E. Rouayheb, D. Evans, J. Gardner, Z. Garrett, A. Gasc ´on, B. Ghazi, P. B. Gibbons, M. Gruteser, Z. Harchaoui, C. He, L. He, Z. Huo, B. Hutchinson, J. Hsu, M. Jaggi, T. Javidi, G. Joshi, M. Khodak...

  32. [32]

    Fedmd: Heterogenous federated learning via model distillation,

    D. Li and J. Wang, “Fedmd: Heterogenous federated learning via model distillation,” 2019. [Online]. Available: https://arxiv.org/abs/1910.03581

    Pith/arXiv arXiv 2019

  33. [33]

    Fedtm: Memory and communication efficient federated learning with tsetlin machine,

    S. H. S. Qi, J. Chauhan, G. V . Merrett, and J. Hare, “Fedtm: Memory and communication efficient federated learning with tsetlin machine,”2023 International Symposium on the Tsetlin Machine (ISTM), pp. 1–8, 2023. [Online]. Available: https://api.semanticscholar.org/CorpusID:268254400

    2023

  34. [34]

    Available: https://arxiv.org/abs/1912.04977

    [Online]. Available: https://arxiv.org/abs/1912.04977

    Pith/arXiv arXiv 1912

  35. [35]

    Tpfl: Tsetlin- personalized federated learning with confidence-based cluster- ing,

    R. J. Gohari, L. Aliahmadipour, and E. Valipour, “Tpfl: Tsetlin- personalized federated learning with confidence-based cluster- ing,” 2024. [Online]. Available: https://arxiv.org/abs/2409.10392

    Pith/arXiv arXiv 2024

  36. [36]

    Towards iot anomaly detection with tsetlin machines,

    O. Gunvaldsen, H. B. Thorsen, P.-A. Andersen, O.-C. Granmo, and M. Goodwin, “Towards iot anomaly detection with tsetlin machines,” in2023 International Symposium on the Tsetlin Machine (ISTM), 2023, pp. 1–8

    2023

  37. [37]

    FedTMOS: Efficient one-shot federated learning with tsetlin machine,

    ——, “FedTMOS: Efficient one-shot federated learning with tsetlin machine,” inThe Thirteenth International Conference on Learning Representations, 2025. [Online]. Available: https://openreview.net/forum?id=44hcrfzydU

    2025

  38. [38]

    Using Tsetlin machine to discover interpretable rules in natural language processing applications,

    R. Saha, O.-C. Granmo, and M. Goodwin, “Using Tsetlin machine to discover interpretable rules in natural language processing applications,”Expert Systems, vol. 40, no. 4, p. e12873, 2023. [Online]. Available: https://onlinelibrary.wiley.com/doi/abs/10.1111/exsy.12873

    doi:10.1111/exsy.12873 2023

  39. [39]

    Tsetlin machine embedding: Representing words using logical expressions,

    B. Bhattarai, O.-C. Granmo, L. Jiao, R. Yadav, and J. Sharma, “Tsetlin machine embedding: Representing words using logical expressions,”arXiv preprint arXiv:2301.00709, 2023

    Pith/arXiv arXiv 2023

  40. [40]

    Low-power au- dio keyword spotting using Tsetlin machines,

    J. Lei, T. Rahman, R. Shafik, A. Wheeldon, A. Yakovlev, O.-C. Granmo, F. Kawsar, and A. Mathur, “Low-power au- dio keyword spotting using Tsetlin machines,”arXiv preprint arXiv:2101.11336, 2021

    Pith/arXiv arXiv 2021

  41. [41]

    Human- level interpretable learning for aspect-based sentiment analysis,

    R. K. Yadav, L. Jiao, O.-C. Granmo, and M. Goodwin, “Human- level interpretable learning for aspect-based sentiment analysis,” Proceedings of the AAAI Conference on Artificial Intelligence, vol. 35, no. 16, pp. 14 203–14 212, May 2021. [Online]. Avail- able: https://ojs.aaai.org/index.php/AAAI/article/view/17671

    2021

  42. [42]

    The convolutional tsetlin machine,

    O.-C. Granmo, S. Glimsdal, L. Jiao, M. Goodwin, C. W. Omlin, and G. T. Berge, “The convolutional tsetlin machine,” 2019. [Online]. Available: https://arxiv.org/abs/1905.09688

    Pith/arXiv arXiv 2019

  43. [43]

    Measuring the novelty of natural language text using the conjunctive clauses of a Tsetlin machine text classifier,

    B. Bhattarai, O.-C. Granmo, and L. Jiao, “Measuring the novelty of natural language text using the conjunctive clauses of a Tsetlin machine text classifier,”arXiv preprint arXiv:2011.08755, 2020

    Pith/arXiv arXiv 2011

  44. [44]

    Asynchronous broadcast-based decentralized learning in sensor networks,

    L. Zhao, W.-Z. Song, X. Ye, and Y . Gu, “Asynchronous broadcast-based decentralized learning in sensor networks,” International Journal of Parallel, Emergent and Distributed Systems, vol. 33, no. 6, pp. 589–607, 2018. [Online]. Available: https://doi.org/10.1080/17445760.2017.1294690

    arXiv 2018

  45. [45]

    Broadcast gossip algorithms for consensus,

    T. C. Aysal, M. E. Yildiz, A. D. Sarwate, and A. Scaglione, “Broadcast gossip algorithms for consensus,”IEEE Transactions on Signal Processing, vol. 57, no. 7, pp. 2748–2761, 2009

    2009

  46. [46]

    Tmcomposites: Plug-and-play collaboration between specialized tsetlin machines,

    O.-C. Granmo, “Tmcomposites: Plug-and-play collaboration between specialized tsetlin machines,” 2023. [Online]. Available: https://arxiv.org/abs/2309.04801

    Pith/arXiv arXiv 2023

  47. [47]

    Peleg,Distributed computing: a locality-sensitive approach

    D. Peleg,Distributed computing: a locality-sensitive approach. USA: Society for Industrial and Applied Mathematics, 2000

    2000

  48. [48]

    Gossip-based computation of aggregate information,

    D. Kempe, A. Dobra, and J. Gehrke, “Gossip-based computation of aggregate information,” in44th Annual IEEE Symposium on Foundations of Computer Science, 2003. Proceedings., 2003, pp. 482–491

    2003

  49. [49]

    N. A. Lynch,Distributed Algorithms. San Francisco, CA, USA: Morgan Kaufmann Publishers Inc., 1996

    1996

  50. [50]

    pytsetlinmachine,

    O.-C. Granmo, “pytsetlinmachine,” 2019. [Online]. Available: https://github.com/cair/pyTsetlinMachine

    2019

  51. [51]

    Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,

    H. Xiao, K. Rasul, and R. V ollgraf, “Fashion-mnist: a novel image dataset for benchmarking machine learning algorithms,”ArXiv, vol. abs/1708.07747, 2017. [Online]. Available: https://api.semanticscholar.org/CorpusID:702279

    Pith/arXiv arXiv 2017

  52. [52]

    Mnist handwritten digit database,

    Y . LeCun, C. Cortes, and C. Burges, “Mnist handwritten digit database,”ATT Labs [Online]. Available: http://yann.lecun.com/exdb/mnist, vol. 2, 2010

    2010

  53. [55]

    Data booleanization for energy efficient on-chip learning using logic driven ai,

    T. Rahman, A. Wheeldon, R. Shafik, A. Yakovlev, J. Lei, O.-C. Granmo, and S. Das, “Data booleanization for energy efficient on-chip learning using logic driven ai,” in2022 International Symposium on the Tsetlin Machine (ISTM), 2022, pp. 29–36

    2022

  54. [2021]

    Available: http://jmlr.org/papers/v22/20-815.html

    [Online]. Available: http://jmlr.org/papers/v22/20-815.html

  55. [2024]

    Available: https://arxiv.org/abs/2310.05269

    [Online]. Available: https://arxiv.org/abs/2310.05269

    Pith/arXiv arXiv