Pith. sign in

REVIEW 3 major objections 5 minor 49 references

Dynamic Tsetlin Machine Accelerators for On-Chip Training at the Edge using FPGAs

T0 review · 3 major / 5 minor · reviewed 2026-08-16 · deepseek-v4-flash

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

Pith's one-line read This paper claims one FPGA accelerator, the Dynamic Tsetlin Machine, can train and retrain both Vanilla and Coalesced Tsetlin Machine models on chip with run-time reconfiguration, offering 2.54x GOP/s/W and 6x lower power than the closest…

desk verdict A genuinely useful TM training accelerator architecture, but the abstract's efficiency claims outrun what Table I actually supports. read the letter →

arxiv 2504.19797 v1 pith:MXTVHXAQ submitted 2025-04-28 cs.AR cs.LG

Gang Mao , Tousif Rahman , Sidharth Maheshwari , Bob Pattison , Zhuang Shao , Rishad Shafik , Alex Yakovlev This is my paper · ORCID
classification cs.ARcs.LG
keywords TsetlinMachineFPGAacceleratoredgetrainingruntimereconfigurationCoalescedon-chiplearningLUT-basedclausecomputationkeywordspotting
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 claims that edge devices can train and retrain machine-learning models on chip using Tsetlin Machines instead of deep neural networks, and that one FPGA accelerator design—the Dynamic Tsetlin Machine (DTM)—can serve both Vanilla and Coalesced Tsetlin Machine algorithms. The central promise is flexibility: the same synthesized bitstream can switch between datasets, model sizes, and algorithm variants at run time, without resynthesis, which matters for privacy-preserving on-field recalibration of IoT sensors. The reported payoff is efficiency: 2.54x more operations per watt and 6x less power than the next-best comparable FPGA design. The argument is that Tsetlin Machine training is logic-based—clauses built from AND, OR, and NOT of Boolean literals updated by finite-state automata—so it maps to LUTs and integer arithmetic instead of floating-point backpropagation.

What carries the argument

The load-bearing object is the partial clause matrix and its partner, the partial weight and class-sum matrix. Instead of computing all clauses in one combinational block, the DTM fixes an x-by-y slice (for example, 32 literals by 27 clauses) and iterates over the model in slices, masking remainder cells. This fixed slice makes the design runtime-programmable: model size, feature count, class count, and TM variant become configuration words rather than hardware parameters. The efficiency claim rests on this reuse plus the feedback-skipping optimization, which skips reading TA RAM slices when no clause in the slice received feedback, giving roughly 40% training-time saving for CoTM as the model converges.

What would settle it

Measure actual DC power of the DTM-L on a ZCU104 board during MNIST training while counting a concrete operation, such as one clause-literal AND evaluation, and apply the same operation definition to the Conv TM design. If the measured GOP/s/W ratio is not 2.54x and system power is not 6x lower, the headline claim fails; likewise, if switching from CoTM to Vanilla TM or from 500 to 2000 clauses requires resynthesis, the runtime-reconfiguration claim fails.

Watch

Extended reading notes

Core claim

The discovery is that Tsetlin Machine training and inference can be unified in a single FPGA architecture with a fixed compute array. The DTM decomposes full clause computation into partial slices of x literals by y clauses, applies masks to handle remainders, stores TA states and CoTM weights in BRAM, and computes clause outputs directly in LUT6 elements. Because the array dimensions, masks, and iteration counts are programmable at run time, the same hardware executes Vanilla TMs, where each class has polarity-divided clause teams, and Coalesced TMs, where a shared clause pool feeds class-specific signed integer weights. Training is done through class-level, clause-level, and TA-level feedback using integer comparisons with random numbers from a re-seeded LFSR cluster, and clause groups with no feedback are skipped in the TA update stage. The paper reports that on MNIST, FMNIST, KMNIST, and keyword spotting the accelerator trains to near the accuracy of prior Tsetlin Machine implementations, with the larger DTM-L configuration using 4.359 W system power and claiming 2.54x higher GOP/s/W than the closest comparable design.

Load-bearing premise

The headline efficiency comparison assumes that "Giga-operations per second" measures the same thing in every accelerator and that Vivado-reported power matches real board power; if operation counts or power sources are not normalized, the 2.54x and 6x numbers may not hold.

Editorial extensions

If this is right

  • An edge node can retrain a Tsetlin Machine model on device after deployment, responding to sensor drift or user personalization without sending data to the cloud.
  • A single FPGA bitstream covers multiple datasets and model configurations, so one product can serve different tasks without recompiling hardware.
  • CoTM users can cut training time by about 40% as clauses converge, because DTM skips feedback-free clause groups during TA updates.
  • Small LFSR lengths, such as 8 bits, are sufficient when seed refreshing is used, which reduces PRNG hardware cost while preserving training accuracy.
  • Training Tsetlin Machines requires very few DSPs, since multiply-accumulate work is replaced by LUT logic and small integer weight updates.

Reading between the lines

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

  • I would read the 2.54x GOP/s/W and 6x power numbers as design-target estimates rather than a settled ranking, since the comparison benchmarks different platforms and mixes system-level and IP-only power; the deeper claim is the runtime-reconfigurable training capability itself, which a board-level measurement could verify.
  • The same partial-clause masking and iteration scheme should extend to Convolutional and Regression Tsetlin Machines, making DTM a template for the whole TM family; adding those modules, which the paper lists as future work, would let the accelerator handle CIFAR-class image inputs.
  • The feedback-skipping optimization suggests a general hardware principle for converging learners: as sparse feedback emerges during training, skipping idle memory reads can cut training time with no accuracy loss; testing this on other automata-based or sparse trainers would show whether it transfers.
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. This paper presents a Dynamic Tsetlin Machine (DTM) FPGA accelerator that supports both inference and on-chip training for Vanilla and Coalesced Tsetlin Machines. The architecture uses partial clause computation, a master-slave LFSR PRNG cluster with seed refreshing, runtime reconfiguration through masks and iteration cycles, and an optimized TA-update skipping mechanism. The design is parameterized and instantiated on two platforms (DTM-S on XC7Z020, DTM-L on ZU-7EV). Evaluation reports accuracy on MNIST, FMNIST, KMNIST, and KWS-6, resource utilization, power from Vivado, and comparisons with earlier FPGA accelerators, claiming 2.54x higher GOP/s/W and 6x lower power than the next-best comparable design.

Significance. If the architectural claims are validated, the paper makes a useful contribution: it appears to be the first published CoTM hardware implementation, it demonstrates runtime reconfiguration without resynthesis, and it maps TM training to LUTs/BRAMs with modest DSP use. The architecture description is detailed and internally coherent, with algorithms for clause, class-sum, and feedback computation. However, the two headline quantitative advantages (2.54x GOP/s/W and 6x lower power) are not supported by the evidence in the manuscript as written, because the comparisons mix power boundaries and the GOP/s metric is undefined. The underlying design is valuable, but the paper should be revised to substantiate or properly qualify these claims.

major comments (3)
  1. [Section V, Table I and Abstract] The abstract's claim of '6x less power than the next-best comparable design' is not supported by the table. The only ratio in Table I that yields approximately 6x is FireFly's system-level power (2.550 W) divided by DTM-S's IP-only power (0.424 W). FireFly is an inference-only SNN accelerator, and DTM-S's IP-only power excludes the ARM/SoC power that is included in DTM-S's own system-power column (1.687 W). Apples-to-apples pairings in Table I give roughly 4.7x (Conv TM IP vs DTM-S IP), 2.7x (Conv TM SYS vs DTM-S SYS), or 3.1x (SNN SYS vs DTM-S SYS), none of which is 6x. Please state explicitly which designs and which power boundaries are being compared, and use consistent measurement boundaries for the headline claim.
  2. [Section V, Fig. 11] The 2.54x GOP/s/W claim is not checkable from the manuscript because 'operations' is never defined for the DTM or for the compared accelerators. No operation count per clause computation, per weight multiply, or per TA update is given, and Fig. 11 mixes training accelerators with inference-only designs. Without a normalized operation definition, the efficiency ratio cannot be established. The paper should either define the operation count and justify its comparability across papers, or replace the GOP/s/W ratio with a better-specified metric such as training datapoints/s/W, which would align with Table II.
  3. [Section V, Tables I-II] The accuracy results are presented without hyperparameter values (T, s, number of clauses for the specific rows, weight precision) or run-to-run variation, even though the architecture's learning efficacy depends on these parameters. The text states that accuracies are from 250 epochs, but without the exact configurations for each row the comparison is not reproducible. Please provide the hyperparameter configuration used for each reported accuracy, at least in supplementary material.
minor comments (5)
  1. [Section IV-A-a and IV-A-f] There are grammatical errors that should be corrected: 'Users does not need' should be 'Users do not need', and 'they will is always 0' should be 'they will always be 0'.
  2. [Section V] Power is reported from Vivado implementation reports, which is simulation-based power and may not reflect board-level measurements; the paper should note this limitation in the comparison.
  3. [Table I] Table I includes SATA, a 65nm ASIC simulation, in an FPGA comparison; while the text acknowledges that SATA is an ASIC, the table caption should also flag it as non-FPGA to avoid misleading readers.
  4. [Section II] The text claims 'the first CoTM hardware' without describing a prior-art search; a brief statement of how this was determined would strengthen the contribution claim.
  5. [Fig. 11 caption] The caption of Fig. 11 appears malformed in the manuscript ('YOLO v3 [23],' is repeated); please clean up the caption.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the paper's architecture and numerical claims rest on external algorithm definitions plus measured/simulated FPGA results; the GOP/s/W and power comparisons are a metric-consistency concern, not a derivation-from-inputs concern.

full rationale

The claimed contributions are an FPGA accelerator and its measured metrics, not a theorem derived from assumptions. The Vanilla TM and CoTM algorithms are taken from external references [8], [9], and the DTM is described as a hardware mapping of those algorithms; no equation in Section IV defines the TM behavior in terms of the accelerator's reported performance. Section V reports test accuracy from 250 training epochs, latency averages, Vivado implementation power, and resource utilization, so the central results are empirical measurements rather than quantities forced by construction. The 2.54x GOP/s/W and 6x power claims depend on comparing Table I entries and Fig. 11 designs, and the operation count per 'GOP' is never defined; that is a correctness/comparability limitation, not circularity, because the numbers are not algebraically equivalent to the inputs. Several references are from the same research group ([13], [15], [19], [36], [38], [39], [41], [42], [46]), and some architectural choices, such as partial clause decomposition, the PRNG cluster, and Booleanization, draw on those prior works. However, those citations supply implementation building blocks and hyperparameter guidance, not the headline efficiency or accuracy conclusions; the central flexibility and training-accuracy claims stand or fall on the measured results independently of these citations. No uniqueness theorem is imported from the authors' prior work, and no fitted parameter is renamed as a prediction. Accordingly, no specific circular step meets the quoted-reduction standard required for a circularity finding.

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

The paper has no invented physical entities. Its accuracy and efficiency results depend on several untested or unreported hyperparameters and on tool-based power estimates, so the ledger mostly captures algorithm hyperparameters and measurement assumptions.

free parameters (5)
  • Threshold hyperparameter T = not reported
    Controls clause update probability in Algorithms 3 and 4; accuracy results depend on it, but the evaluation does not list T values for MNIST, FMNIST, KMNIST, or KWS-6.
  • Sensitivity hyperparameter s = not reported
    Controls TA update probability in Algorithm 5; not reported for the evaluated datasets.
  • Weight precision = 12 bits (suggested)
    Fig. 14a shows accuracy saturates at 12-bit weights; likely the chosen width, but not stated in the evaluation setup.
  • LFSR length = 12-bit (DTM-S), 24-bit (DTM-L)
    Design parameters shown in Section V Point 1; chosen as a trade-off between randomness quality and resource/power.
  • Clause and feature matrix sizes = DTM-S: 32x16 and 2x4; DTM-L: 32x27 and 8x4
    Fixed compute matrix dimensions that define the reconfiguration envelope; selected for the target FPGA platforms.
assumptions (5)
  • domain assumption Tsetlin Machine feedback rules and convergence behavior are correct as described in [8].
    The entire training datapath implements these rules; the paper does not re-derive convergence.
  • domain assumption Booleanization into binary literals preserves enough information for the target tasks.
    The accelerator consumes Booleanized features; the KWS-6 Booleanization follows [46], and MNIST uses simple thresholding.
  • domain assumption The LFSR-based PRNG with seed refresh supplies random numbers of sufficient quality for TM training.
    Fig. 15 supports this empirically for the tested LFSR lengths, but it is assumed for all reported runs.
  • domain assumption Vivado power reports approximate real hardware power.
    The power numbers used in the efficiency comparison are from implementation reports, not board measurements.
  • domain assumption GOP/s values from different papers use equivalent operation definitions.
    Needed to compare DTM against FP-BNN, FlexBlock, Conv TM, and others; not justified in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dynamic Tsetlin Machine Accelerators for On-Chip Training at the Edge using FPGAs." pith.science (2026). https://pith.science/paper/MXTVHXAQ

@misc{pith2026250419797,
  author       = {Pith},
  title        = {Pith review of: Dynamic Tsetlin Machine Accelerators for On-Chip Training at the Edge using FPGAs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/MXTVHXAQ}},
  note         = {Machine review of arXiv:2504.19797}
}
read the original abstract

The increased demand for data privacy and security in machine learning (ML) applications has put impetus on effective edge training on Internet-of-Things (IoT) nodes. Edge training aims to leverage speed, energy efficiency and adaptability within the resource constraints of the nodes. Deploying and training Deep Neural Networks (DNNs)-based models at the edge, although accurate, posit significant challenges from the back-propagation algorithm's complexity, bit precision trade-offs, and heterogeneity of DNN layers. This paper presents a Dynamic Tsetlin Machine (DTM) training accelerator as an alternative to DNN implementations. DTM utilizes logic-based on-chip inference with finite-state automata-driven learning within the same Field Programmable Gate Array (FPGA) package. Underpinned on the Vanilla and Coalesced Tsetlin Machine algorithms, the dynamic aspect of the accelerator design allows for a run-time reconfiguration targeting different datasets, model architectures, and model sizes without resynthesis. This makes the DTM suitable for targeting multivariate sensor-based edge tasks. Compared to DNNs, DTM trains with fewer multiply-accumulates, devoid of derivative computation. It is a data-centric ML algorithm that learns by aligning Tsetlin automata with input data to form logical propositions enabling efficient Look-up-Table (LUT) mapping and frugal Block RAM usage in FPGA training implementations. The proposed accelerator offers 2.54x more Giga operations per second per Watt (GOP/s per W) and uses 6x less power than the next-best comparable design.

Figures

Figures reproduced from arXiv: 2504.19797 by the authors.

Figure 1
Figure 1. Block diagram of the fundamental components of the Tsetlin Machine algorithms for inference - they build the Vanilla TM (pink) and the Coalesced TM (CoTM) (yellow). Component breakdown: a) shows the pre-processing or Booleanization process (adapted from [13]) to generate the inputs to TMs using an MNIST datapoint example; b) shows the learning element - the Tsetlin Automata (TA); c) shows how each TA relates to its … view at source ↗
Figure 2
Figure 2. Block diagram of learning process used in the Vanilla TM and CoTM referred to as feedback. Feedback is used to transition the learning elements of the model, i.e., the TAs and weights (weights for CoTM). These processes are presented visually here, but will be explored as algorithm blocks in the subsequent section. Composite TMs involve a group composed predominantly of Convolution TMs [10] that each work on their o… view at source ↗
Figure 3
Figure 3. The log scale comparison of logic-based (clause compute) vs. integer-based (class sum compute) arithmetic operations in CoTM inference. The number of operations increase with clauses. outputs differently. Instead of instancing clauses for every class, it creates a shared pool of clauses (SC), shown in Fig. 1e. These clause outputs are shared across all classes. To generate a class sum, each class has its own set of … view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: The architecture of the proposed Dynamic Tsetlin Machine (DTM) training accelerator. The DTM accommodates both the Vanilla and CoTM algorithms. It consists of 5 core modules: 1-Clause Matrix, 2-Weight Matrix, 3-Weight Update Matrix, 4-TA State Update Matrix, 5-Argmax O…
Figure 5
Figure 5. Figure 5: Visualizing the iterations required to compute full clause and weight matrices. These figures show the rounds of compute required for the Clause Matrix and Weight Matrix. The colors of (a) and (b) correspond to the colors of the matrix modules in [PITH_FULL_IMAGE:figu…
Figure 6
Figure 6. Figure 6: Masks for remainder compute elements (marked in red). BRAM depending on the FPGA resources - this is the clause buffer. The clause buffer stores only 1 class of clause outputs. h) Class Sum Computation: A similar slicing strategy is also applied in the Weight Matrix fo…
Figure 7
Figure 7. Figure 7: Training time per epoch [PITH_FULL_IMAGE:figures/full_fig_p010_7.png]
Figure 9
Figure 9. Figure 9: Accelerator Timing: a) timing diagram for CoTM inference; b) timing diagram for Vanilla TM inference; c) timing diagram for DTM training (same procedure for both Vanilla and CoTM). in IV.For training, hyperparameters like the TA state update probability 1 s and thresho…
Figure 10
Figure 10. Figure 10: DTM accelerator data flows. V. EVALUATION The DTM accelerator is designed to train on edge applica￾tions, particularly cases where recalibration is required. This is typical with IoT sensor data inputs for edge tasks. Three Evaluation Points: This section presents thr…
Figure 11
Figure 11. Figure 11: Conv-TM accelerator [40], FP-BNN [20], Reconfig DNN [27], Low-batch CNN [29], F-CNN [30], YOLO v3 [23], [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 14
Figure 14. Figure 14: Resource utilization and power for different weight length [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 15
Figure 15. Figure 15: Resource utilization and power for different LFSR length. overhead in power and resources when varying the weight bit precision and LFSR length in the PRNG design. The results indicate that at least 12-bit weight precision is required to achieve good learning efficacy…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

49 extracted references · 40 canonical work pages

  1. [1]

    Real- time model calibration with deep reinforcement learning,

    Y . Tian, M. A. Chao, C. Kulkarni, K. Goebel, and O. Fink, “Real- time model calibration with deep reinforcement learning,” Mechanical Systems and Signal Processing , 2022. PREPRINT - ACCEPTED IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS 13 TABLE I: Comparison with other FPGA accelerators suitable for similar edge sensor-based applicatio...

    work page 2022

  2. [2]

    SemiPFL: Personalized Semi-Supervised Federated Learning Framework for Edge Intelligence,

    A. Tashakori, W. Zhang, Z. Jane Wang, and P. Servati, “SemiPFL: Personalized Semi-Supervised Federated Learning Framework for Edge Intelligence,” IEEE Internet of Things Journal , 2023

    work page 2023

  3. [3]

    Convergence of edge computing and deep learning: A comprehensive survey,

    X. Wang, Y . Han, V . C. M. Leung, D. Niyato, X. Yan, and X. Chen, “Convergence of edge computing and deep learning: A comprehensive survey,” IEEE Communications Surveys & Tutorials , 2020

    work page 2020

  4. [4]

    FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks,

    M. Blott, T. B. Preußer, N. J. Fraser, G. Gambardella, K. O’brien, Y . Umuroglu, M. Leeser, and K. Vissers, “FINN-R: An End-to-End Deep-Learning Framework for Fast Exploration of Quantized Neural Networks,” ACM Trans. Reconfigurable Technol. Syst. , 2018

    work page 2018

  5. [5]

    FINN: A Framework for Fast, Scalable Binarized Neural Network Inference,

    Y . Umuroglu, N. J. Fraser, G. Gambardella, M. Blott, P. Leong, M. Jahre, and K. Vissers, “FINN: A Framework for Fast, Scalable Binarized Neural Network Inference,” in ACM/SIGDA, 2017

    work page 2017

  6. [6]

    Machine learning on FPGAs to face the IoT revolution,

    X. Zhang, A. Ramachandran, C. Zhuge, D. He, W. Zuo, Z. Cheng, K. Rupnow, and D. Chen, “Machine learning on FPGAs to face the IoT revolution,” in IEEE/ACM ICCAD, 2017

    work page 2017

  7. [7]

    Edge FPGA-based Onsite Neural Network Training,

    R. Chen, H. Zhang, Y . Li, R. Zhang, G. Li, J. Yu, and K. Wang, “Edge FPGA-based Onsite Neural Network Training,” in IEEE ISCAS, 2023

    work page 2023

  8. [8]

    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,” 2021. [Online]. Available: https://arxiv.org/abs/1804.01508

    arXiv 2021

Show all 49 references
  1. [9]

    Coalesced Multi-Output Tsetlin Machines with Clause Sharing,

    S. Glimsdal and O.-C. Granmo, “Coalesced Multi-Output Tsetlin Machines with Clause Sharing,” 2021. [Online]. Available: https: //arxiv.org/abs/2108.07594

    2021 arXiv

  2. [10]

    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

    2019 arXiv

  3. [11]

    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,” 2019. [Online]. Available: https://arxiv.org/abs/1905.04206

    2019 arXiv

  4. [12]

    Efficient Acceleration of Deep Learning Inference on Resource-Constrained Edge Devices: A Review,

    M. M. H. Shuvo, S. K. Islam, J. Cheng, and B. I. Morshed, “Efficient Acceleration of Deep Learning Inference on Resource-Constrained Edge Devices: A Review,” Proceedings of the IEEE , 2023

    2023

  5. [13]

    Logic-Based Machine Learning with Reproducible Decision Model Using the Tsetlin Machine,

    O. Tarasyuk, A. Gorbenko, T. Rahman, R. Shafik, and A. Yakovlev, “Logic-Based Machine Learning with Reproducible Decision Model Using the Tsetlin Machine,” in IEEE IDAACS, 2023

    2023

  6. [14]

    The mnist database of handwritten digit images for machine learning research,

    L. Deng, “The mnist database of handwritten digit images for machine learning research,” IEEE Signal Processing Magazine , 2012

    2012

  7. [15]

    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,” in IEEE ISTM, 2022

    2022

  8. [16]

    An Optimized Toolbox for Advanced Image Processing with Tsetlin Machine Composites,

    Y . Grønningsæter, H. S. Smørvik, and O.-C. Granmo, “An Optimized Toolbox for Advanced Image Processing with Tsetlin Machine Composites,” 2024. [Online]. Available: https://arxiv.org/abs/2406.00704

    2024 arXiv

  9. [17]

    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

    2023 arXiv

  10. [18]

    On the conver- gence of tsetlin machines for the identity- and not operators,

    X. Zhang, L. Jiao, O.-C. Granmo, and M. Goodwin, “On the conver- gence of tsetlin machines for the identity- and not operators,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2022

    2022

  11. [19]

    Systematic Search for Optimal Hyper-parameters of the Tsetlin Machine on MNIST Dataset,

    O. Tarasyuk, T. Rahman, R. Shafik, A. Yakovlev, A. Gorbenko, O.-C. Granmo, and L. Jiao, “Systematic Search for Optimal Hyper-parameters of the Tsetlin Machine on MNIST Dataset,” in 2023 IEEE ISTM , 2023

    2023

  12. [20]

    Liang, S

    S. Liang, S. Yin, L. Liu, W. Luk, and S. Wei, “FP-BNN,” Neurocomput,

  13. [21]

    FracBNN: Accurate and FPGA-Efficient Binary Neural Networks with Fractional Activations,

    Y . Zhang, J. Pan, X. Liu, H. Chen, D. Chen, and Z. Zhang, “FracBNN: Accurate and FPGA-Efficient Binary Neural Networks with Fractional Activations,” 2020. [Online]. Available: https://arxiv.org/abs/2012.12206

    2020 arXiv

  14. [22]

    Implementation of Data-optimized FPGA-based Accelerator for Convolutional Neural Network,

    M. Cho and Y . Kim, “Implementation of Data-optimized FPGA-based Accelerator for Convolutional Neural Network,” in IEEE ICEIC, 2020

    2020

  15. [23]

    A Low- Latency FPGA Accelerator for YOLOv3-Tiny With Flexible Layerwise Mapping and Dataflow,

    M. Kim, K. Oh, Y . Cho, H. Seo, X. T. Nguyen, and H.-J. Lee, “A Low- Latency FPGA Accelerator for YOLOv3-Tiny With Flexible Layerwise Mapping and Dataflow,” IEEE Transactions on Circuits and Systems I: Regular Papers, 2024

    2024

  16. [24]

    High-Throughput and Energy-Efficient FPGA-Based Accelerator for All Adder Neural Networks,

    N. Zhang, S. Ni, L. Chen, T. Wang, and H. Chen, “High-Throughput and Energy-Efficient FPGA-Based Accelerator for All Adder Neural Networks,” IEEE Internet of Things Journal , 2025

    2025

  17. [25]

    WSQ- AdderNet: Efficient Weight Standardization based Quantized AdderNet FPGA Accelerator Design with High-Density INT8 DSP-LUT Co- Packing Optimization,

    Y . Zhang, B. Sun, W. Jiang, Y . Ha, M. Hu, and W. Zhao, “WSQ- AdderNet: Efficient Weight Standardization based Quantized AdderNet FPGA Accelerator Design with High-Density INT8 DSP-LUT Co- Packing Optimization,” in 2022 IEEE/ACM ICCAD , 2022

    2022

  18. [26]

    DNNBuilder: an Automated Tool for Building High-Performance DNN Hardware Accelerators for FPGAs,

    X. Zhang, J. Wang, C. Zhu, Y . Lin, J. Xiong, W.-m. Hwu, and D. Chen, “DNNBuilder: an Automated Tool for Building High-Performance DNN Hardware Accelerators for FPGAs,” in IEEE/ACM ICCAD, 2018

    2018

  19. [27]

    A Reconfigurable DNN Training Accelerator on FPGA,

    J. Lu, J. Lin, and Z. Wang, “A Reconfigurable DNN Training Accelerator on FPGA,” in IEEE SiPS, 2020

    2020

  20. [28]

    Training Deep Neural Networks in Low-Precision with High Accuracy Using FPGAs,

    S. Fox, J. Faraone, D. Boland, K. Vissers, and P. H. Leong, “Training Deep Neural Networks in Low-Precision with High Accuracy Using FPGAs,” in IEEE ICFPT, 2019

    2019

  21. [29]

    FPGA-based Low-Batch Training Accelerator PREPRINT - ACCEPTED IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS 14 for Modern CNNs Featuring High Bandwidth Memory,

    S. K. Venkataramanaiah, H.-S. Suh, S. Yin, E. Nurvitadhi, A. Dasu, Y . Cao, and J.-S. Seo, “FPGA-based Low-Batch Training Accelerator PREPRINT - ACCEPTED IN IEEE TRANSACTIONS ON CIRCUITS AND SYSTEMS—I: REGULAR PAPERS 14 for Modern CNNs Featuring High Bandwidth Memory,” in IEEE...

    2020

  22. [30]

    F-CNN: An FPGA-based framework for training Convolutional Neural Networks,

    W. Zhao, H. Fu, W. Luk, T. Yu, S. Wang, B. Feng, Y . Ma, and G. Yang, “F-CNN: An FPGA-based framework for training Convolutional Neural Networks,” in IEEE ASAP, 2016

    2016

  23. [31]

    An On-Chip Fully Connected Neural Network Training Hardware Accelerator Based on Brain Float Point and Sparsity Awareness,

    T.-H. Tsai and D.-B. Lin, “An On-Chip Fully Connected Neural Network Training Hardware Accelerator Based on Brain Float Point and Sparsity Awareness,” IEEE Open Journal of Circuits and Systems , 2023

    2023

  24. [32]

    FlexBlock: A Flexible DNN Training Accelerator With Multi-Mode Block Floating Point Support,

    S.-H. Noh, J. Koo, S. Lee, J. Park, and J. Kung, “FlexBlock: A Flexible DNN Training Accelerator With Multi-Mode Block Floating Point Support,” IEEE Transactions on Computers , 2023

    2023

  25. [33]

    FPGA-NHAP: A General FPGA- Based Neuromorphic Hardware Acceleration Platform With High Speed and Low Power,

    Y . Liu, Y . Chen, W. Ye, and Y . Gui, “FPGA-NHAP: A General FPGA- Based Neuromorphic Hardware Acceleration Platform With High Speed and Low Power,” IEEE Transactions on Circuits and Systems I: Regular Papers, 2022

    2022

  26. [34]

    FireFly: A High- Throughput Hardware Accelerator for Spiking Neural Networks With Efficient DSP and Memory Optimization,

    J. Li, G. Shen, D. Zhao, Q. Zhang, and Y . Zeng, “FireFly: A High- Throughput Hardware Accelerator for Spiking Neural Networks With Efficient DSP and Memory Optimization,” IEEE Transactions on Very Large Scale Integration (VLSI) Systems , 2023

    2023

  27. [35]

    SATA: Sparsity-Aware Training Accelerator for Spiking Neural Networks,

    R. Yin, A. Moitra, A. Bhattacharjee, Y . Kim, and P. Panda, “SATA: Sparsity-Aware Training Accelerator for Spiking Neural Networks,” IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems, 2023

    2023

  28. [36]

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

    A. Wheeldon, R. Shafik, T. Rahman, J. Lei, A. Yakovlev, and O.-C. Granmo, “Learning automata based energy-efficient AI hardware design for IoT applications,” Philos. Trans. R. Soc. A Math. Phys. Eng. Sci. , 2020

    2020

  29. [37]

    R. A. Fisher, “Iris,” UCI Machine Learning Repository, 1988, DOI: https://doi.org/10.24432/C56C76

    1988 doi

  30. [38]

    REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,

    S. Maheshwari, T. Rahman, R. Shafik, A. Yakovlev, A. Rafiev, L. Jiao, and O.-C. Granmo, “REDRESS: Generating Compressed Models for Edge Inference Using Tsetlin Machines,” IEEE Transactions on Pattern Analysis and Machine Intelligence , 2023

    2023

  31. [39]

    MATADOR: Automated System-on-Chip Tsetlin Machine Design Gen- eration for Edge Applications,

    T. Rahman, G. Mao, S. Maheshwari, R. Shafik, and A. Yakovlev, “MATADOR: Automated System-on-Chip Tsetlin Machine Design Gen- eration for Edge Applications,” in IEEE DATE, 2024

    2024

  32. [40]

    Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,

    S. A. Tunheim, L. Jiao, R. Shafik, A. Yakovlev, and O.-C. Granmo, “Tsetlin Machine-Based Image Classification FPGA Accelerator With On-Device Training,” IEEE Transactions on Circuits and Systems I: Regular Papers, 2025

    2025

  33. [41]

    IMBUE: In-Memory Boolean-to-CUrrent Inference ArchitecturE for Tsetlin Machines,

    O. Ghazal, S. Singh, T. Rahman, S. Yu, Y . Zheng, D. Balsamo, S. Patkar, F. Merchant, F. Xia, A. Yakovlev, and R. Shafik, “IMBUE: In-Memory Boolean-to-CUrrent Inference ArchitecturE for Tsetlin Machines,” in IEEE/ACM ISLPED, 2023

    2023

  34. [42]

    Parallel Symbiotic Random Number Generator for Training Tsetlin Machines on FPGA,

    T. Rahman, G. Mao, S. Maheshwari, K. Krishnamurthy, R. Shafik, and A. Yakovlev, “Parallel Symbiotic Random Number Generator for Training Tsetlin Machines on FPGA,” in IEEE ISTM, 2023

    2023

  35. [43]

    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,” 2017. [Online]. Available: https://arxiv.org/abs/1708.07747

    2017 arXiv

  36. [44]

    Deep learning for classical japanese literature,

    T. Clanuwat, M. Bober-Irizar, A. Kitamoto, A. Lamb, K. Yamamoto, and D. Ha, “Deep learning for classical japanese literature,” arXiv preprint arXiv:1812.01718, 2018

    2018 arXiv

  37. [45]

    Speech commands: A dataset for limited-vocabulary speech recognition,

    P. Warden, “Speech commands: A dataset for limited-vocabulary speech recognition,” arXiv preprint arXiv:1804.03209 , 2018

    2018 arXiv

  38. [46]

    Low-Power Audio 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 Audio Keyword Spotting Using Tsetlin Machines,” Journal of Low Power Electronics and Applications ,

  39. [47]

    Learning under Concept Drift: A Review,

    J. Lu, A. Liu, F. Dong, F. Gu, J. Gama, and G. Zhang, “Learning under Concept Drift: A Review,” IEEE Transactions on Knowledge and Data Engineering, 2019. Gang Mao received B.E degree from Northeastern University, China in 2016, and M.E degree from Newcastle University, Newcas...

    2019

  40. [2018]

    Available: https://doi.org/10.1016/j.neucom.2017.09.046

    [Online]. Available: https://doi.org/10.1016/j.neucom.2017.09.046

    2017 doi

  41. [2021]

    Available: https://www.mdpi.com/2079-9268/11/2/18

    [Online]. Available: https://www.mdpi.com/2079-9268/11/2/18

Pith tools

Reviewed August 16, 2026 · model on record in the stance chip above.