Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

Embedded FPGA Acceleration of Brain-Like Neural Networks: Online Learning to Scalable Inference

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a Bayesian Confidence Propagation Neural Network (BCPNN), a brain-like model with sparse Hebbian learning, can be accelerated on an embedded FPGA so that online learning and inference run on-device, with up to 17.5x…

desk verdict Solid first embedded BCPNN accelerator with plausible absolute measurements, but the headline speedups rest on an underspecified ARM baseline and the 'without sacrificing accuracy' claim overreaches. read the letter →

arxiv 2506.18530 v1 pith:STYIYD3K submitted 2025-06-23 cs.AR cs.AI

Muhammad Ihsan Al Hafiz , Naresh Ravichandran , Anders Lansner , Pawel Herman , Artur Podobas This is my paper · ORCID
classification cs.ARcs.AI
keywords BCPNNbrain-likeneuralnetworksembeddedFPGAZynqUltraScale+high-levelsynthesisneuromorphiccomputingonlinelearninghalfprecision
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 aims to show that a brain-like neural network called BCPNN, which learns with biologically plausible local rules rather than error backpropagation, can be moved from datacenter hardware onto a low-power embedded FPGA. The authors build the first embedded accelerator for BCPNN on a Zynq UltraScale+ system-on-chip, supporting both full online learning and inference-only operation, and evaluate it on MNIST, a chest X-ray pneumonia set, and a breast cancer set. They report up to 17.5x lower latency and 94% lower energy than a host ARM A53 core at essentially unchanged test accuracy, with half-precision arithmetic giving the best balance. The value is practical: devices that cannot rely on cloud connectivity could learn and adapt on-device with a neuromorphic model.

What carries the argument

The load-bearing machinery is the BCPNN learning rule and its dependence on synaptic traces. Each connection stores exponential moving averages of pre-synaptic, post-synaptic, and joint activation probabilities, from which biases and weights are computed as log-probabilities and log pointwise mutual information. On the FPGA, the accelerator is a stream-based high-level-synthesis design: AXI bursts fetch matrix slices from DDR into FIFO streams, sub-kernels process them in parallel with loop unrolling, and the unsafe-math flag plus precision tuning (FP32, FP16, and mixed FP16/FXP16) trade resources for throughput. The sparse connectivity and soft-winner-take-all hypercolumn organization (groups of minicolumn units that compete to represent features) are what let the design fit in a 1728-DSP, 11 Mb-BRAM fabric.

What would settle it

Compile the same BCPNN model for the ARM Cortex-A53 with aggressive optimization (-O3, NEON/SIMD, cache-friendly layout) and re-measure latency and energy. If the speedups and energy savings over this optimized baseline fall below the reported 17.5x and 94%, the paper's central efficiency claim is weakened.

Watch

Extended reading notes

Core claim

The paper's central claim is that the BCPNN model — with its sparse columnar architecture, local Bayesian-Hebbian learning, and three-population feedforward structure — fits naturally onto the stream-based dataflow pattern of FPGA acceleration, and that this mapping can fit within the tight BRAM, DSP, and power budgets of an embedded SoC-FPGA. The authors present two kernels: a full online-learning kernel that updates synaptic traces (p_i, p_j, and p_ij) and computes weights as log(p_ij/(p_i p_j)) and biases as log(p_j), and a leaner inference-only kernel. The inference-only kernel with 16-bit half precision runs 11-17.5x faster than the ARM baseline, cuts energy by roughly 90%, and stays within about a point of floating-point test accuracy, while the online-learning kernel manages a more modest 2-2.6x speedup because of BRAM pressure and a lower parallelism factor.

Load-bearing premise

The headline speedups and energy savings rest on the assumption that the ARM A53 baseline is a fair, representative implementation; the paper gives no compiler flags, optimization level, or memory layout for the baseline, so a weaker-than-necessary baseline would inflate the reported ratios.

Editorial extensions

If this is right

  • BCPNN online learning can run entirely on an embedded device, removing the need for cloud round-trips during training.
  • Half-precision (FP16) is the recommended edge configuration: it roughly halves latency and energy relative to FP32 while preserving accuracy on the tested datasets.
  • The stream-based architecture holds resource usage roughly constant as model size varies, so larger BCPNN models can be deployed on the same FPGA without redesign.
  • The recommended model-scaling order for latency and energy is to reduce HCUs first, then MCUs, and only last adjust connectivity sparsity, since aggressive sparsification rapidly degrades accuracy.
  • Mixed fixed/float precision is risky for complex datasets (accuracy dropped from 86.2% to 84.1% on Pneumonia and from 84.0% to 78.8% on Breast Cancer), so it is not a general-purpose choice.

Reading between the lines

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

  • A fairer ARM baseline (compiled with -O3, NEON/SIMD, and tuned memory layout) might shrink the reported 17.5x and 94% margins; the paper does not document baseline build settings, so the absolute efficiency advantage of the FPGA is less certain than the relative one.
  • The same design recipe — stream-based high-level synthesis with precision scaling — could be applied to other Bayesian-Hebbian models or BCPNN variants with reward traces, which the paper does not demonstrate.
  • Because resource usage stays flat while latency scales with model size, the practical limit of this accelerator is likely memory bandwidth for weight fetches rather than fabric capacity; a bandwidth-bound analysis would show where the design saturates.
  • A direct comparison against an optimized edge inference engine (for example, a Cortex-M with CMSIS-NN or a neural processing unit) would position BCPNN against conventional deep networks; the paper only compares against a scalar ARM core.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents an HLS-based FPGA accelerator for the Bayesian Confidence Propagation Neural Network (BCPNN) on a Zynq UltraScale+ (ZCU104) SoC. It implements a full online-learning kernel and an inference-only kernel with FP32, FP16, and mixed-precision (FP16/FXP16) variants. The accelerator is evaluated on MNIST, Pneumonia, and Breast Cancer datasets, reporting up to 17.56x latency speedup and 94.1% energy savings over an ARM A53 baseline, with accuracy figures claimed to be preserved. Additional experiments examine model-size scaling and precision trade-offs.

Significance. If the claims hold, the work is a useful engineering contribution for neuromorphic edge computing: it demonstrates a complete, on-device BCPNN accelerator on a low-power embedded FPGA, including online learning, and provides a systematic evaluation of precision and model-scaling trade-offs. The paper's strengths include real measurements on a ZCU104 board, the use of three public datasets, detailed resource-utilization reports, and a clear separation of full-learning and inference-only kernels. The main limitation is that the headline speedup and energy figures are relative to an ARM baseline that is not adequately documented, and no variance/error statistics are provided, so the quantitative claims are not yet reproducible. The abstract's 'without sacrificing accuracy' is also too strong given the mixed-precision results on complex datasets.

major comments (4)
  1. [V-A, Table III] The ARM A53 baseline used to compute the speedup and energy-saving ratios is not sufficiently described. The manuscript reports only 'similar model and execution parameters' and omits compiler flags, optimization level, scalar vs. NEON/SIMD code, single- vs. multi-core execution, memory layout, and whether the baseline is the same C++ source as the one synthesized by HLS. On a Cortex-A53, differences between -O0 scalar loops and -O3 with NEON vectorization can easily exceed 2-4x, so the headline 2.06x-17.56x speedups and 87.8%-94.1% energy savings are not reproducible as written. Please document the baseline build and execution environment fully, and consider reporting absolute FPGA latencies and energies alongside relative ratios.
  2. [IV-C, V-A, Tables III and IV] No repeated-run statistics or error bars are reported for latency, energy, or accuracy. Latency is measured with gettimeofday but the number of trials and mean/median/variance are not given; accuracy is a single test-set point per configuration. As a result, the full-kernel accuracy difference (94.6% on ARM vs. 94.3% on FPGA) cannot be assessed as noise or a real regression, and the abstract's 'without sacrificing accuracy' is not supported by the data. Please report multiple repetitions (at least 5-10) with standard deviations for latency and energy, and confidence intervals or repeated-training/test noise for accuracy.
  3. [Abstract, V-B, Fig. 5] The abstract's claim that the accelerator achieves its speedups 'without sacrificing accuracy' is contradicted by the paper's own mixed-precision results: for Breast Cancer, mixed precision (FXP16/FP16) drops test accuracy from 84.0% (FP32) to 78.8%, and for Pneumonia from 86.2% to 84.1%. Since mixed precision is presented as a contribution (Section III-C), the abstract and Section V-B need to be scoped explicitly, e.g., by stating that FP16 preserves accuracy while mixed precision trades accuracy for resources on complex datasets.
  4. [IV-C2, V-A, Table IV] The abstract's headline '94% energy savings' is computed from board power, which includes all ZCU104 board components, whereas the execution-power basis yields 87.8% for the same configuration. Both numbers are reported in Table IV, but the abstract does not state which basis is used. Please clarify which energy metric is used in the headline and discuss whether board power is the intended deployment metric or includes fixed overheads that may not reflect a production edge device.
minor comments (5)
  1. [Section VI] The text attributes the Artix-7 BCPNN accelerator to 'Liu et al. (2020) ... [12]' but reference [12] is a different Liu paper on an upper-limb rehabilitation device; the correct citation appears to be [18] (L. Liu et al., NorCAS 2020).
  2. [Figure 5] The caption of Figure 5 says 'Latency vs Power' but the axes are labeled 'Latency (ms)' and 'Energy (mJ)'; the caption should say 'Latency vs Energy'.
  3. [IV-C2] The power measurement section states that more than 1000 samples were recorded for idle and active states, but no standard deviation or confidence interval is reported in Table IV or the text; please include these statistics.
  4. [Section I] The claim of 'the world's first Edge-capable BLNN accelerator' should be qualified relative to prior Artix-7 implementations (e.g., [18]) by specifying that the novelty is an embedded SoC-FPGA implementation with complete online-learning and inference-only kernels for rate-based BCPNN.
  5. [Section III-B] The phrase 'Vitis-HLS unsafe-math is enabled' should be reworded to 'Vitis HLS unsafe-math optimizations are enabled' for readability and consistency with the tool name.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the accelerator's speedups, energy, and accuracy are measured against an ARM A53 baseline and public datasets, not derived from or fitted to the model inputs.

full rationale

This is a measurement and engineering paper rather than a derivation. The BCPNN equations (1) and (2) are adopted from the authors' prior work [1] as background definitions, and the stream-based pipeline is adapted from [3]; these self-citations establish provenance, but no load-bearing numerical claim is justified solely by them. The central claims—latency speedups up to 17.56x, energy savings up to 94.1%, and matching accuracy—are obtained by direct measurement against an ARM A53 host and public MNIST, Pneumonia, and Breast Cancer datasets, i.e., externally benchmarked rather than fitted or predicted from the model's own outputs. No equation's output is defined as its own input, and no fitted parameter is renamed as a prediction. The underspecified ARM baseline (Section V-A) is a legitimate methodological reproducibility concern about the fairness of the speedup ratios, but it is not a circularity: the accelerator's absolute operation is still measured, not entailed by its construction. Similarly, the 'world's first' novelty claim rests on a literature survey, not on a self-citation chain that forces the result. Therefore the paper exhibits no significant circularity.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim is an engineering measurement, so the ledger is dominated by model hyperparameters chosen by hand (hidden sizes, sparsity, learning rates) and by unstated benchmarking assumptions (ARM baseline fairness, power isolation, test-set non-leakage). No new physical entities are introduced.

free parameters (4)
  • Hidden-layer size (HCU/MCU) = MNIST 32/128; Pneumonia 10-30/200-400; Breast 10/1000
    Set by empirical iteration (Table II) to balance accuracy and hardware cost; accuracy and latency both depend on these values.
  • Connectivity sparsity (n_act/n_sil) = MNIST 64/64; Pneumonia 80-320/24-80; Breast 676/156
    Structural-plasticity parameters chosen per dataset; Section V-C shows these strongly affect accuracy, so they are fitted rather than derived.
  • Learning epochs and tau_p = MNIST 5/3; Pneumonia 5/0.3; Breast 15/0.2
    Training hyperparameters from Table II; they set the learned weights and therefore the reported accuracies.
  • Mixed-precision fixed-point format Q3.12 = 4 integer bits, 12 fractional bits
    Chosen by hand to represent trained parameters; its adequacy is only checked empirically via the accuracy drop on complex datasets in Section V-B.
assumptions (5)
  • domain assumption The BCPNN equations (1)-(2) map probability traces to a usable classifier on the three datasets.
    The accelerator inherits the model from [1]; the paper validates it only through final test accuracy, not by re-deriving or benchmarking against other learning rules.
  • domain assumption Vitis HLS 2023.2 produces functionally correct hardware, and unsafe-math optimizations preserve accuracy to within the reported small differences.
    Section IV-A enables config_compile -unsafe_math_optimizations; no formal equivalence or bit-accurate verification is provided, only RTL co-simulation for FIFO depth.
  • domain assumption The INA226 idle-subtraction protocol isolates the accelerator's dynamic energy.
    Section IV-C.2 defines execution power as active board power minus idle board power, assuming static power is constant and the delta is entirely due to the kernel.
  • domain assumption The ARM A53 baseline is a fair and reasonably optimized comparison.
    Section V-A gives no compiler flags, optimization level, or ARM code details, so the validity of the speedup and energy ratios rests on this unstated assumption.
  • domain assumption Test-set accuracy is not biased by the empirical iteration used to select model configurations.
    Section IV-B says parameters were selected through empirical iteration without describing a separate validation split, so the accuracy claims assume no test-set leakage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Embedded FPGA Acceleration of Brain-Like Neural Networks: Online Learning to Scalable Inference." pith.science (2026). https://pith.science/paper/STYIYD3K

@misc{pith2026250618530,
  author       = {Pith},
  title        = {Pith review of: Embedded FPGA Acceleration of Brain-Like Neural Networks: Online Learning to Scalable Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/STYIYD3K}},
  note         = {Machine review of arXiv:2506.18530}
}
read the original abstract

Edge AI applications increasingly require models that can learn and adapt on-device with minimal energy budget. Traditional deep learning models, while powerful, are often overparameterized, energy-hungry, and dependent on cloud connectivity. Brain-Like Neural Networks (BLNNs), such as the Bayesian Confidence Propagation Neural Network (BCPNN), propose a neuromorphic alternative by mimicking cortical architecture and biologically-constrained learning. They offer sparse architectures with local learning rules and unsupervised/semi-supervised learning, making them well-suited for low-power edge intelligence. However, existing BCPNN implementations rely on GPUs or datacenter FPGAs, limiting their applicability to embedded systems. This work presents the first embedded FPGA accelerator for BCPNN on a Zynq UltraScale+ SoC using High-Level Synthesis. We implement both online learning and inference-only kernels with support for variable and mixed precision. Evaluated on MNIST, Pneumonia, and Breast Cancer datasets, our accelerator achieves up to 17.5x latency and 94% energy savings over ARM baselines, without sacrificing accuracy. This work enables practical neuromorphic computing on edge devices, bridging the gap between brain-like learning and real-world deployment.

Figures

Figures reproduced from arXiv: 2506.18530 by the authors.

Figure 1
Figure 1. Architecture of a BLNN based on the BCPNN model, featuring HCU [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. BCPNN system architecture on embedded FPGA. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 5
Figure 5. Latency vs Power across Datasets and Precision Types. Accuracy is [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: FPGA Resource Usage for Baseline BLNN Kernels float precision [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 6
Figure 6. Figure 6: FPGA Resource Usage for precision variants [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Latency vs Energy for model size variations on the Pneumonia dataset. [PITH_FULL_IMAGE:figures/full_fig_p007_7.png]
Figure 8
Figure 8. Figure 8: FPGA Resource Usage for the pneumonia BLNN as model size scales [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. A Hardware-oriented Approach for Efficient Bayesian Inference Computation and Deployment

    cs.AI 2026-07 conditional novelty 6.0 of 10

    Reorganizing the memory layout of tensor contractions speeds up discrete Bayesian message passing on an embedded GPU by 2-2.5x typically, up to about 5x, with mathematically identical outputs.

Reference graph

Works this paper leans on

19 extracted references · 19 canonical work pages · cited by 1 Pith paper

  1. [1]

    Unsupervised representa- tion learningwith Hebbian synaptic and structural plasticity inbrain-like feedforward neural networks,

    N. Ravichandran, A. Lansner, and P. Herman, “Unsupervised representa- tion learningwith Hebbian synaptic and structural plasticity inbrain-like feedforward neural networks,”Neurocomputing, p. 129440, Jan. 2025

    work page 2025

  2. [2]

    Elusive scents: neurocompu- tational mechanisms of verbal omissions in free odor naming,

    N. Chrysanthidis, R. Raj, T. H ¨orberg, R. Lindroos, A. Lansner, E. Laukka, J. Olofsson, and P. Herman, “Elusive scents: neurocompu- tational mechanisms of verbal omissions in free odor naming,”bioRxiv, pp. 2025–02, 2025

    work page 2025

  3. [3]

    A Reconfigurable Stream-Based FPGA Accelerator for Bayesian Confidence Propagation Neural Networks,

    M. I. Al Hafiz, N. Ravichandran, A. Lansner, P. Herman, and A. Podobas, “A Reconfigurable Stream-Based FPGA Accelerator for Bayesian Confidence Propagation Neural Networks,” inApplied Reconfigurable Computing. Architectures, Tools, and Applications (R. Giorgi, M. Stojilovic, D. Stroobandt, P. Brox Jimenez, and A. Bar- riga Barros, eds.), (Cham), pp. 196–2...

    work page 2025

  4. [4]

    The columnar organization of the neocortex.,

    V . B. Mountcastle, “The columnar organization of the neocortex.,”Brain, vol. 120, pp. 701–722, Apr. 1997

    work page 1997

  5. [5]

    A Theory of How Columns in the Neocortex Enable Learning the Structure of the World,

    J. Hawkins, S. Ahmad, and Y . Cui, “A Theory of How Columns in the Neocortex Enable Learning the Structure of the World,”Frontiers in Neural Circuits, vol. 11, Oct. 2017. Publisher: Frontiers

    work page 2017

  6. [6]

    Brain-Like Ap- proaches to Unsupervised Learning of Hidden Representations - A Com- parative Study,

    N. B. Ravichandran, A. Lansner, and P. Herman, “Brain-Like Ap- proaches to Unsupervised Learning of Hidden Representations - A Com- parative Study,” inArtificial Neural Networks and Machine Learning – ICANN 2021(I. Farka ˇs, P. Masulli, S. Otte, and S. Wermter, eds.), (Cham), pp. 162–173, Springer International Publishing, 2021

    work page 2021

  7. [7]

    FPGA-Based HPC for Associative Memory System,

    D. Wang, Y . Wang, Y . Yang, D. Stathis, A. Hemani, A. Lansner, J. Xu, L.-R. Zheng, and Z. Zou, “FPGA-Based HPC for Associative Memory System,” in2024 29th Asia and South Pacific Design Automation Conference (ASP-DAC), pp. 52–57, Jan. 2024. ISSN: 2153-697X

    work page 2024

  8. [8]

    Scalable Multi-FPGA HPC Architecture for Associative Memory System,

    D. Wang, X. Yan, Y . Yang, D. Stathis, A. Hemani, A. Lansner, J. Xu, L.-R. Zheng, and Z. Zou, “Scalable Multi-FPGA HPC Architecture for Associative Memory System,”IEEE Transactions on Biomedical Circuits and Systems, vol. 19, pp. 454–468, Apr. 2025

    work page 2025

Show all 19 references
  1. [9]

    Reconfigurable Logic for Hardware IP Protection: Opportunities and Challenges (Invited Paper),

    L. Collini, B. Tan, C. Pilato, and R. Karri, “Reconfigurable Logic for Hardware IP Protection: Opportunities and Challenges (Invited Paper),” in2022 IEEE/ACM International Conference On Computer Aided Design (ICCAD), pp. 1–7, Oct. 2022. ISSN: 1558-2434

    2022

  2. [10]

    Implementation of Robotic Navigation Algorithms Using Partial Reconfiguration on Zynq SoC,

    M. Basha, M. S. Kumar, and M. C. Chinnaiah, “Implementation of Robotic Navigation Algorithms Using Partial Reconfiguration on Zynq SoC,”ECS Transactions, vol. 107, p. 13887, Apr. 2022. Publisher: IOP Publishing

    2022

  3. [11]

    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, vol. 4, pp. 85– 98, 2023

    2023

  4. [12]

    An FPGA-Based Upper-Limb Rehabilitation Device for Gesture Recognition and Motion Evaluation Using Multi-Task Recurrent Neural Networks,

    H. Liu, A. Panahi, D. Andrews, and A. Nelson, “An FPGA-Based Upper-Limb Rehabilitation Device for Gesture Recognition and Motion Evaluation Using Multi-Task Recurrent Neural Networks,” in2020 International Conference on Field-Programmable Technology (ICFPT), pp. 296–297, Dec. 2020

    2020

  5. [13]

    Optimized Memory Allocation and Power Minimization for FPGA- Based Image Processing,

    P. Garcia, D. Bhowmik, R. Stewart, G. Michaelson, and A. Wallace, “Optimized Memory Allocation and Power Minimization for FPGA- Based Image Processing,”Journal of Imaging, vol. 5, p. 7, Jan. 2019. Number: 1 Publisher: Multidisciplinary Digital Publishing Institute

    2019

  6. [14]

    FPGA(ZCU104) Based Energy Efficient Accelerator for MobileNet- V1,

    Y . R. M. Reddy, P. Muralidhar, G. Satya Narayana, and D. Jagan, “FPGA(ZCU104) Based Energy Efficient Accelerator for MobileNet- V1,” in2024 20th IEEE International Colloquium on Signal Processing & Its Applications (CSPA), pp. 57–62, Mar. 2024. ISSN: 2836-4090

    2024

  7. [15]

    An FPGA-Based Reconfig- urable Convolutional Neural Network Accelerator for Tiny YOLO-V3,

    T.-H. Tsai, N.-C. Tung, and C.-Y . Chen, “An FPGA-Based Reconfig- urable Convolutional Neural Network Accelerator for Tiny YOLO-V3,” Circuits, Systems, and Signal Processing, vol. 44, pp. 3388–3409, May 2025

    2025

  8. [16]

    Gradient-based learning applied to document recognition,

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, pp. 2278–2324, Nov. 1998

    1998

  9. [17]

    MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification,

    J. Yang, R. Shi, D. Wei, Z. Liu, L. Zhao, B. Ke, H. Pfister, and B. Ni, “MedMNIST v2 - A large-scale lightweight benchmark for 2D and 3D biomedical image classification,”Scientific Data, vol. 10, p. 41, Jan. 2023

    2023

  10. [18]

    A FPGA-based Hardware Accelerator for Bayesian Confidence Propagation Neural Network,

    L. Liu, D. Wang, Y . Wang, A. Lansner, A. Hemani, Y . Yang, X. Hu, Z. Zou, and L. Zheng, “A FPGA-based Hardware Accelerator for Bayesian Confidence Propagation Neural Network,” in2020 IEEE Nordic Circuits and Systems Conference (NorCAS), pp. 1–6, Oct. 2020

    2020

  11. [19]

    StreamBrain: An HPC Frame- work for Brain-like Neural Networks on CPUs, GPUs and FPGAs,

    A. Podobas, M. Svedin, S. W. D. Chien, I. B. Peng, N. B. Ravichandran, P. Herman, A. Lansner, and S. Markidis, “StreamBrain: An HPC Frame- work for Brain-like Neural Networks on CPUs, GPUs and FPGAs,” in Proceedings of the 11th International Symposium on Highly Efficient Accel...

    2021

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.