Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

DEBUG-HD: Debugging TinyML models on-device using Hyper-Dimensional computing

T0 review · 4 major / 5 minor · reviewed 2026-08-12 · deepseek-v4-flash

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

Pith's one-line read DEBUG-HD claims that a binary hyperdimensional classifier initialized with the binarized hidden-layer weights of a small trained MLP can diagnose input-corruption type on-device, outperforming prior binary HDC methods by 27 percent on…

desk verdict A clean internal ablation supports DEBUG-HD's core mechanism, but the headline 27% gain over prior binary HDC is not yet established because the baselines get no tuning at low hyper-dimensions. read the letter →

arxiv 2411.10692 v1 pith:HW35CL63 submitted 2024-11-16 cs.LG cs.AIcs.NE

Nikhil P Ghanathe , Steven J E Wilton This is my paper · ORCID
classification cs.LGcs.AIcs.NE
keywords TinyMLon-devicedebugginghyperdimensionalcomputingbinaryHDCencoderMLP-assistedencodinginputcorruptiondiagnosismodelreliabilityembeddedmachinelearning
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

DEBUG-HD tries to establish that a hyperdimensional (HDC) classifier small enough for kilobyte-scale TinyML hardware can still identify what kind of corruption broke a deployed model. Standard binary HDC methods need hyper-dimensions above 10,000 to work well, which is too much memory for such devices; below 1,000 dimensions their random projection encoder loses accuracy sharply. The paper's fix is to replace the random projection with the binarized hidden-layer weight matrix of a small two-layer MLP trained on the same corruption-classification task. With this encoder, DEBUG-HD reports 27 percent average improvement over prior binary HDC methods (12 percent higher accuracy than the best prior binary baseline) at hyper-dimensions of 300 to 400, plus top-2 and top-3 accuracy high enough to narrow down root causes. The value, if correct, is that a tiny device can diagnose its own failures without transmitting raw sensor data or halting operation.

What carries the argument

The load-bearing object is the MLP-initialized binary HDC encoder. In standard HDC, the encoder is a random projection matrix that maps a $d$-dimensional feature vector to a $D_H$-dimensional hypervector; with $D_H$ below 1000, random vectors lose the near-orthogonality that makes the representation useful. DEBUG-HD replaces that matrix with $P = \mathrm{sgn}(W)$, where $W$ is the hidden-layer weight matrix of a two-layer MLP trained to classify the corruption types, with hidden size equal to $D_H$ and no bias or activation. Because the hidden layer's matrix-vector product has the same form as HDC encoding, the trained weights project inputs into a hyperspace that is already separated by corruption class; binarization keeps the encoder memory-light, and the class hypervectors are formed by summing the encoded training hypervectors. The paper couples this with a feature-source choice: it taps an intermediate layer of the base network so the HDC classifier runs on compact features without preprocessing and without pausing the model. The layer choice is selected by exhaustive or binary search in the appendix.

What would settle it

Feed the deployed HDC classifier inputs corrupted by a type not in its training list (such as a sensor dropout or compression artifact absent from the benchmark corruption sets) and check whether its top-1 and top-2 outputs behave differently from chance; if unseen corruptions are simply force-labeled as the nearest known type, the closed-world assumption, and the reported field accuracy, fail.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the encoding stage of a binary HDC classifier can be learned rather than randomly generated, and that the learned encoder is what makes low-hyper-dimension HDC viable for on-device debugging. The authors train a two-layer MLP (hidden layer sized to the target hyper-dimension, no bias, no activation) on the task of classifying corrupted inputs, then take its hidden-layer weight matrix $W$, apply $\mathrm{sgn}(\cdot)$ to obtain a bipolar projection, and use that matrix in the HDC encoder in place of a random projection. The input to the encoder is not the raw sensor data but the output of a chosen intermediate layer of the base TinyML model, obtained without interrupting the model. In the evaluation, DEBUG-HD reaches top-1 corruption-type accuracy of 0.68 on CIFAR10-C, 0.78 on TinyImageNet-C, and 0.45 on SpeechCmd-C at hyper-dimensions 300, 400, and 300, outperforming all binary HDC baselines and approaching the accuracy of a conventional MLP. The authors also argue that the gap is largest exactly in the low-hyper-dimension regime where memory-constrained devices must operate.

Load-bearing premise

The method assumes the corruptions that appear after deployment are drawn from the same finite list the HDC classifier was trained on (19, 15, or 11 types), so a novel corruption would still be labeled as one of the known types.

Editorial extensions

If this is right

  • If the central claim holds, a TinyML device can run corruption diagnosis with only the added cost of a binary HDC classifier, because the encoder is a single matrix-vector product and the class hypervectors need only a few thousand bits.
  • The reported top-2 and top-3 accuracies (up to 0.95 on CIFAR10 and TinyImageNet) mean the tool can narrow candidate root causes even when exact top-1 naming is wrong.
  • Because the diagnostic is invoked only after an accuracy-monitoring mechanism flags a drop, it slots into an existing reliability pipeline without slowing normal inference.
  • The improvement is concentrated where memory is tight: at hyper-dimension 2000 the gap over the best binary baseline shrinks to about one point, so the learned encoder buys accuracy exactly in the low-hyper-dimension regime.
  • Developers can prune near-duplicate corruptions (for example, defocus blur and gaussian blur) to raise top-1 accuracy from 0.68 to 0.79 while keeping coverage of the remaining failure modes.

Reading between the lines

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

  • Inference: the learned-projection mechanism is not tied to a two-layer MLP; the same argument suggests any pretrained network's intermediate features, binarized, could initialize the HDC encoder, which would eliminate the separate MLP training step and make the method plug into existing deployments.
  • Inference: because the evaluation is closed-world, the reported 0.68 to 0.78 top-1 accuracies are an upper bound on field diagnosis; a Hamming-distance-based 'unknown' rejection rule would be a natural testable extension for novel corruptions.
  • Inference: the exhaustive layer-search result implies that layer choice, not just encoder initialization, dominates accuracy, so a cheap separability-based selection criterion could replace the search and is a testable prediction of the approach.
  • Inference: the SSIM tables imply a label-design rule: drop corruption types that are pairwise too similar, trading diagnosis granularity for accuracy, as the 19-to-12 corruption reduction on CIFAR10 shows with top-1 accuracy rising from 0.68 to 0.79.
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

4 major / 5 minor

Summary. The paper proposes DEBUG-HD, a binary hyperdimensional computing (HDC) classifier for on-device debugging of TinyML models, targeting the task of identifying the corruption type that caused an input distribution shift. The core idea is to replace the random projection matrix used by conventional HDC encoders with the binarized hidden-layer weights of a trained two-layer MLP (no bias, no activation), which the authors argue preserves the HDC training/inference pipeline while improving separability in low hyper-dimensions. The method is evaluated on CIFAR10-C, TinyImageNet-C, and SpeechCmd-C, using features tapped from an intermediate layer of a pretrained base network, at hyper-dimensions of 300–400. The paper reports an average 12 percentage point (approximately 27% relative) accuracy improvement over the best binary HDC baseline, and positions the approach as enabling corruption-type diagnosis within a few KB of added memory.

Significance. If the empirical claims hold, DEBUG-HD is a meaningful contribution to TinyML reliability: it offers a small-memory (kilobyte-scale) on-device mechanism for root-cause diagnosis of input corruptions, which is a practical gap in deployed tinyML systems. The clean ablation between Vanilla HDC and DEBUG-HD, which differ only in projection initialization while sharing the same single-pass training and inference pipeline, supports the direction of the claimed improvement. The proposed explanation for why multi-epoch binary HDC baselines degrade at low hyper-dimensions (dimension regeneration and binary-versus-real class hypervector maintenance) is plausible. However, the headline quantitative advantage is not yet firmly established: the binary baselines appear untuned for the low hyper-dimension regime, no repeated-seed statistics are reported, and the gain on TinyImageNet is only 3 accuracy points, so the magnitude of the 12-point average gain is uncertain.

major comments (4)
  1. [Section 4, Table 3a] The comparison against binary HDC baselines is not yet convincing because the baselines are run with torch-hd defaults and no reported hyperparameter tuning at the low hyper-dimensions (300–400) used in the evaluation. Since the paper itself argues that the baselines' multi-epoch retraining and dimension-regeneration behavior degrades sharply in this regime, a fair test requires reporting validation-based tuning of each baseline's training epochs, learning rate, and encoder update rules under the same memory budget, or a demonstration that the torch-hd default settings are already optimal at hyper-dimension 300–400.
  2. [Section 4, Table 3a] No variance or repeated-seed statistics are reported for any of the methods. On TinyImageNet, DEBUG-HD's advantage over AdaptHD is only 3 accuracy points (0.78 vs 0.75), which may be within run-to-run noise; the large average gain is driven almost entirely by the CIFAR10 difference (0.68 vs 0.45). Please report means and standard deviations over at least 3–5 independent runs for all methods and characterize the statistical significance of the claimed 12-point average gain.
  3. [Abstract and Section 4/Limitations] The abstract states that DEBUG-HD 'outperforms prior binary HDC methods by 27% on average' without the hyper-dimension qualifier, but the Limitations section concedes that the gain is 'notable only in smaller hyperspaces,' with AdaptHD and DEBUG-HD nearly tied at hyper-d=2000 on TinyImageNet (0.83 vs 0.84). The abstract and Section 1 should be revised to state that the improvement holds at hyper-dimensions of 300–400, which is the intended operating regime, so that readers do not infer a general superiority across all hyper-dimensional settings.
  4. [Appendix A.3.1] The procedure for selecting the tapped intermediate layer uses a Vanilla HDC's 'validation accuracy,' but the text does not make clear whether the validation set is a held-out split of DHDC-train or the same DHDC-test used for final accuracy reporting. If the final test distribution is used for layer selection, the reported accuracies are partially selected on the test set. Please clarify the validation protocol and, if necessary, perform layer selection on a separate validation split.
minor comments (5)
  1. [Section 4 and Appendix A.1.2] The number of SpeechCmd-C corruptions is inconsistent: Section 4 says 11 corruptions, but Appendix A.1.2 says '11 corruptions' and then lists 12 types (including both 'time mask' and 'time stretch'). Please reconcile the count and the list.
  2. [Throughout] There are several typos: 'TinyImagnet' (Section 4), 'benchmarck' (Section 4), 'libary' (Appendix A.1.2), and 'expect' instead of 'except' (Appendix A.2). A thorough proofread is needed.
  3. [Figure 3b] The caption 'HDC base model' is ambiguous; the figure should clearly label the base model size versus the total size including DEBUG-HD, or otherwise clarify what the bars represent.
  4. [Table 3a] The row labeled 'MLP' is a conventional neural network reference, not an HDC method; consider relabeling it or adding a footnote to distinguish it from the HDC baselines.
  5. [Section 2.1 and Section 3] The notation switches between 'hyper-d' and 'DH'; define the relationship explicitly, for example by stating that DH denotes the integer value of the hyper-dimension, and use the terms consistently throughout.

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation chain: the HDC accuracy comparison is external and measured against torch-hd baselines, with only a minor, non-load-bearing self-citation in the trigger simulation.

full rationale

DEBUG-HD's central claim is an empirical comparison against prior binary HDC methods implemented with the external torch-hd library on standard corrupted benchmarks. The MLP-assisted encoder is a supervised initialization: the MLP hidden-layer weights are learned on the same corruption-label task, binarized, and then used as the HDC projection matrix. This is a legitimate transfer of a learned representation rather than a fitted parameter renamed as a prediction, because the HDC class hypervectors are computed independently by summing encoded training samples and the reported HDC accuracies (0.68/0.78/0.45) are substantially below the MLP reference (0.85/0.90/0.54), so the result is not forced by construction. The paper transparently reports the MLP reference and limits the claimed advantage to low hyper-dimensions. The only self-citation is Ghanathe and Wilton [15], used in Appendix A.4 to simulate the accuracy-drop trigger; this is a procedural borrowing and does not carry the central accuracy claim. No equation reduces to its own input, and no prediction is equivalent to a fitted value. Accordingly, no circular steps are identified; the score reflects a single minor, non-load-bearing self-citation.

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

The method adds no new physical entities; the MLP-derived projection is a trained parameter, not a postulate. The fitted choices that the 27% claim actually rests on are hyper-dimension (300/400/300), the per-dataset tapped layer, severity level 5, the MLP training schedule, and the inherited monitor threshold. The axioms are: standard HDC orthogonality of random hypervectors; the load-bearing domain assumption that tapped intermediate-layer features encode corruption type (conceded as limiting in Section 4); the ad hoc no-bias/no-activation encoder design; the closed-world corruption set; and the SSIM-based retraining-transfer argument borrowed from Kindermans et al. [38]. The two premises most likely to break outside this paper's three datasets are the intermediate-feature assumption and the closed-world assumption.

free parameters (5)
  • hyper-dimension DH per dataset = 300 (CIFAR10), 400 (TinyImageNet), 300 (SpeechCmd)
    Chosen by hand to balance resource use and accuracy (Section 4, Appendix A.4). The headline 27% claim is tied to this operating point; the Limitations report the advantage shrinks to 0.01 at DH=2000 (AdaptHD 0.83 vs DEBUG-HD 0.84 on TinyImageNet).
  • tapped intermediate layer of base network = after 3rd residual stack (ResNet-8); 5th inverted residual block (MobileNetV2); after 1st depthwise layer (DSCNN)
    Selected per dataset by exhaustive or binary search over layers on validation accuracy (Appendix A.3.1). All downstream numbers depend on this fitted choice.
  • corruption severity level = 5 for CIFAR10-C and TinyImageNet-C
    Fixed at severity 5 (Appendix A.1.2); results may not transfer to other severities, and no severity sweep is reported.
  • MLP training schedule = 20 epochs, batch size 256
    Reported in Appendix A.2; learning rate, optimizer, and hidden-layer detail beyond size are unstated, so the encoder quality is not fully pinned down.
  • monitor trigger threshold = sliding window 100, trigger at mean minus 3 sigma
    Inherited from the authors' prior QUTE work [15] and used to simulate accuracy-drop invocation (Appendix A.4); the end-to-end system using it is never evaluated.
assumptions (5)
  • standard math Random hypervectors are nearly orthogonal, and this property weakens for hyper-dimensions below 1000 (Section 2.1, Figure 2).
    Standard HDC assumption from Kanerva [33]; the paper's motivation for learned projections rests on the low-dimensional failure of this property.
  • domain assumption Corruption type information survives in the tapped intermediate-layer output of the base network (Section 3, Appendix A.3).
    If false, no HDC classifier on these features could work; the paper validates it only by per-dataset layer search and concedes the tapped output 'may not be the most informative' (Section 4).
  • ad hoc to paper A 2-layer MLP hidden layer with no bias and no activation is a faithful surrogate for the HDC encoder (Section 3.1, Appendix A.4.1).
    The authors report that adding bias or activation hurt accuracy; the no-bias/no-activation configuration is an empirical choice specific to this method, not a derived necessity.
  • domain assumption Deployment corruptions come from the finite pre-enumerated set used in training (Section 4).
    Closed-world assumption: 'a finite set of commonly-occurring corruptions... used to train the HDC before deployment.' Unseen corruption types would be force-labeled into known classes.
  • domain assumption Retraining on one corruption (defocus blur) improves robustness to SSIM-similar corruptions (gaussian blur), via the signal/distractor argument of Kindermans et al. [38] (Appendix A.5.1).
    Used to justify excluding similar corruptions from the debug set; verified only in the one DB/GB example, so it is an assumption for general use.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DEBUG-HD: Debugging TinyML models on-device using Hyper-Dimensional computing." pith.science (2026). https://pith.science/paper/HW35CL63

@misc{pith2026241110692,
  author       = {Pith},
  title        = {Pith review of: DEBUG-HD: Debugging TinyML models on-device using Hyper-Dimensional computing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HW35CL63}},
  note         = {Machine review of arXiv:2411.10692}
}
read the original abstract

TinyML models often operate in remote, dynamic environments without cloud connectivity, making them prone to failures. Ensuring reliability in such scenarios requires not only detecting model failures but also identifying their root causes. However, transient failures, privacy concerns, and the safety-critical nature of many applications-where systems cannot be interrupted for debugging-complicate the use of raw sensor data for offline analysis. We propose DEBUG-HD, a novel, resource-efficient on-device debugging approach optimized for KB-sized tinyML devices that utilizes hyper-dimensional computing (HDC). Our method introduces a new HDC encoding technique that leverages conventional neural networks, allowing DEBUG-HD to outperform prior binary HDC methods by 27% on average in detecting input corruptions across various image and audio datasets.

Figures

Figures reproduced from arXiv: 2411.10692 by the authors.

Figure 1
Figure 1. DEBUG-HD overview with training (solid lines) and inference (dot￾ted lines) flows Traditional ML debug TensorFlow Debugger (tfdbg) [8] provides deep insights into TensorFlow graphs for software-like debugging, while TensorBoard [61] and Ten￾sorWatch [54] allow visual monitoring of model statistics. Typically, lossy compression (e.g., statistical summaries) is sufficient for machine learning debugging. Tools like ML-… view at source ↗
Figure 2
Figure 2. Top-1 accuracy of Vanilla HDC for hyper-d from 200 to 10k (x-axis logscale) where δ(.) denotes similarity function, and for real-valued HVs it is usually cosine similarity. For bipolar HVs, this reduces to Hamming distance, where class with lowest distance is predicted. Traditional HDC classifiers project input data into a hyper-d > 10, 000 [34], causing overhead that is impractical for extreme edge devices . Prior … view at source ↗
Figure 3
Figure 3. Experimental results employ a retraining step over multiple epochs, with some methods [14, 28] regenerating the least contributing hyper-dimensions by changing the encoding. However, we find that with lower hyper-d values (e.g., 300, 400), changes to encoder after a train-pass leads to an immediate drop in accuracy, with the next epoch barely recovering the lost performance, resulting in suboptimal outcomes. In cont… view at source ↗

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. Ariel-ML: Computing Parallelization with Embedded Rust for Neural Networks on Heterogeneous Multi-core Microcontrollers

    cs.LG 2025-12 conditional novelty 6.0 of 10

    Ariel-ML combines the IREE compiler with a Rust operating system to give microcontrollers automatic multi-core parallel inference for TinyML, with a measured 1.5x speedup on a dual-core board.

Reference graph

Works this paper leans on

72 extracted references · 50 canonical work pages · cited by 1 Pith paper

  1. [1]

    Abubakar Abid, Mert Yuksekgonul, and James Zou. 2022. Meaningfully debugging model mistakes using conceptual counterfactual explanations. In International Conference on Machine Learning. PMLR, 66–88

    work page 2022

  2. [2]

    Amina Adadi and Mohammed Berrada. 2018. Peeking inside the black-box: a survey on explainable artificial intelligence (XAI). IEEE access 6 (2018), 52138–52160

    2018

  3. [3]

    Khaled Ahmed and Mohamed Hassan. 2022. tinycare: A tinyml-based low-cost continuous blood pressure estimation on the extreme edge. In 2022 IEEE 10th International Conference on Healthcare Informatics (ICHI). IEEE, 264–275

    work page 2022

  4. [4]

    Sajid Ali, Tamer Abuhmed, Shaker El-Sappagh, Khan Muhammad, Jose M Alonso-Moral, Roberto Confalonieri, Riccardo Guidotti, Javier Del Ser, Natalia Díaz-Rodríguez, and Francisco Herrera. 2023. Explainable Artificial Intelligence (XAI): What we know and what is left to attain Trustworthy Artificial Intelligence. Information Fusion (2023), 101805

    work page 2023

  5. [5]

    David Baehrens, Timon Schroeter, Stefan Harmeling, Motoaki Kawanabe, Katja Hansen, and Klaus-Robert Müller. 2010. How to explain individual classification decisions. The Journal of Machine Learning Research 11 (2010), 1803–1831

    work page 2010

  6. [6]

    Colby R. Banbury, Vijay Janapa Reddi, Peter Torelli, Jeremy Holleman, Nat Jeffries, Csaba Király, Pietro Montino, David Kanter, Sebastian Ahmed, Danilo Pau, Urmish Thakker, Antonio Torrini, Pete Warden, Jay Cordaro, Giuseppe Di Guglielmo, Javier M. Duarte, Stephen Gibellini, Videet Parekh, Honson Tran, Nhan Tran, Wenxu Niu, and Xuesong Xu. 2021. MLPerf Ti...

    arXiv 2021

  7. [7]

    Kulkarni, Rashmi Rane, Sarika Bobde, and Ruhi Patankar

    Mamta Bhamare, Pradnya V . Kulkarni, Rashmi Rane, Sarika Bobde, and Ruhi Patankar

    work page

  8. [8]

    Shanqing Cai, Eric Breck, Eric Nielsen, Michael Salib, and D Sculley. 2016. Tensorflow debugger: Debugging dataflow graphs for machine learning. (2016)

    work page 2016

Show all 72 references
  1. [9]

    Cheng-Yang Chang, Yu-Chuan Chuang, Chi-Tse Huang, and An-Yeu Wu. 2023. Recent Progress and Development of Hyperdimensional Computing (HDC) for Edge Intelligence. IEEE Journal on Emerging and Selected Topics in Circuits and Systems13, 1 (2023), 119–136. https://doi.org/10.1109/...

    2023

  2. [10]

    Eyal Cidon, Evgenya Pergament, Zain Asgar, Asaf Cidon, and Sachin Katti. 2021. Characteriz- ing and taming model instability across edge devices. Proceedings of Machine Learning and Systems 3 (2021), 624–636

    2021

  3. [11]

    Cody Coleman, Christopher Yeh, Stephen Mussmann, Baharan Mirzasoleiman, Peter Bailis, Percy Liang, Jure Leskovec, and Matei Zaharia. 2020. Selection via Proxy: Efficient Data Selection for Deep Learning. arXiv:1906.11829 [cs.LG] https://arxiv.org/abs/1906. 11829

    2020 arXiv

  4. [12]

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. 2009. Imagenet: A large-scale hierarchical image database. In 2009 IEEE conference on computer vision and pattern recognition. Ieee, 248–255

    2009

  5. [13]

    Duarte Dias and João Paulo Silva Cunha. 2018. Wearable health devices—vital sign monitoring, systems and technologies. Sensors 18, 8 (2018), 2414

    2018

  6. [14]

    Shijin Duan, Yejia Liu, Shaolei Ren, and Xiaolin Xu. 2022. Lehdc: Learning-based hyper- dimensional computing classifier. In Proceedings of the 59th ACM/IEEE Design Automation Conference. 1111–1116. 5

    2022

  7. [15]

    Nikhil P Ghanathe and Steve Wilton. 2024. QUTE: Quantifying Uncertainty in TinyML models with Early-exit-assisted ensembles. arXiv:2404.12599 [cs.LG] https://arxiv.org/abs/ 2404.12599

    2024 arXiv

  8. [16]

    Wei Hao, Zixi Wang, Lauren Hong, Lingxiao Li, Nader Karayanni, Chengzhi Mao, Junfeng Yang, and Asaf Cidon. 2023. Monitoring and Adapting ML Models on Mobile Devices. arXiv preprint arXiv:2305.07772 (2023)

    2023 arXiv

  9. [17]

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. 2016. Deep Residual Learning for Image Recognition. In 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). 770–778. https://doi.org/10.1109/CVPR.2016.90

    2016 doi

  10. [18]

    Mike Heddes, Igor Nunes, Pere Vergés, Denis Kleyko, Danny Abraham, Tony Givargis, Alexan- dru Nicolau, and Alexander Veidenbaum. 2023. Torchhd: An open source python library to support research on hyperdimensional computing and vector symbolic architectures. Journal of Machine...

    2023

  11. [19]

    Dan Hendrycks and Thomas Dietterich. 2019. Benchmarking neural network robustness to common corruptions and perturbations. arXiv preprint arXiv:1903.12261 (2019)

    2019 arXiv

  12. [20]

    Alejandro Hernández-Cano, Namiko Matsumoto, Eric Ping, and Mohsen Imani. 2021. Onlinehd: Robust, efficient, and single-pass online learning using hyperdimensional system. In 2021 Design, Automation & Test in Europe Conference & Exhibition (DATE). IEEE, 56–61

    2021

  13. [21]

    Daniel Holanda Noronha, Ruizhe Zhao, Jeff Goeders, Wayne Luk, and Steven JE Wilton. 2019. On-chip FPGA debug instrumentation for machine learning applications. In Proceedings of the 2019 ACM/SIGDA International Symposium on Field-Programmable Gate Arrays. 110–115

    2019

  14. [22]

    Daniel Holanda Noronha, Ruizhe Zhao, Zhiqiang Que, Jeffrey Goeders, Wayne Luk, and Steve Wilton. 2019. An Overlay for Rapid FPGA Debug of Machine Learning Applications. In 2019 International Conference on Field-Programmable Technology (ICFPT) . 135–143. https://doi.org/10.1109...

    2019

  15. [23]

    Andrew G Howard, Menglong Zhu, Bo Chen, Dmitry Kalenichenko, Weijun Wang, Tobias Weyand, Marco Andreetto, and Hartwig Adam. 2017. Mobilenets: Efficient convolutional neural networks for mobile vision applications. arXiv preprint arXiv:1704.04861 (2017)

    2017 arXiv

  16. [24]

    Yen-Chang Hsu, Yilin Shen, Hongxia Jin, and Zsolt Kira. 2020. Generalized odin: Detecting out-of-distribution image without learning from out-of-distribution data. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition. 10951–10960

    2020

  17. [25]

    Zhaolan Huang, Koen Zandberg, Kaspar Schleiser, and Emmanuel Baccelli. 2023. U-TOE: Universal TinyML On-Board Evaluation Toolkit for Low-Power IoT. In2023 12th IFIP/IEEE International Conference on Performance Evaluation and Modeling in Wired and Wireless Networks (PEMWN). IEEE, 1–6

    2023

  18. [26]

    Zhaolan Huang, Koen Zandberg, Kaspar Schleiser, and Emmanuel Baccelli. 2024. RIOT-ML: toolkit for over-the-air secure updates and performance evaluation of TinyML models. Annals of Telecommunications (2024), 1–15

    2024

  19. [27]

    Jonathan J. Hull. 1994. A database for handwritten text recognition research.IEEE Transactions on pattern analysis and machine intelligence 16, 5 (1994), 550–554

    1994

  20. [28]

    Mohsen Imani, Samuel Bosch, Sohum Datta, Sharadhi Ramakrishna, Sahand Salamat, Jan M Rabaey, and Tajana Rosing. 2019. Quanthd: A quantization framework for hyperdimensional computing. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems 39, 10 (2019), ...

    2019

  21. [29]

    Mohsen Imani, Justin Morris, Samuel Bosch, Helen Shu, Giovanni De Micheli, and Tajana Ros- ing. 2019. Adapthd: Adaptive efficient training for brain-inspired hyperdimensional computing. In 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS). IEEE, 1–4. 6

    2019

  22. [30]

    Mohsen Imani, Sahand Salamat, Behnam Khaleghi, Mohammad Samragh, Farinaz Koushanfar, and Tajana Rosing. 2019. SparseHD: Algorithm-Hardware Co-optimization for Efficient High-Dimensional Computing. In 2019 IEEE 27th Annual International Symposium on Field- Programmable Custom C...

    2019

  23. [31]

    Md Shafayat Jamil, Sirdarta Prashad Banik, GM Atiqur Rahaman, and Sajib Saha. 2023. Ad- vanced GradCAM++: Improved Visual Explanations of CNN Decisions in Diabetic Retinopathy. In Computer Vision and Image Analysis for Industry 4.0. CRC Press 6000 Broken Sound Park- way NW, Su...

    2023

  24. [32]

    Iver Jordal. 2024. Audiomentations. https://github.com/iver56/audiomentations Version 0.32.0

    2024

  25. [33]

    Pentti Kanerva. 1998. Encoding Structure in Boolean Space. In ICANN 98, Lars Niklasson, Mikael Bodén, and Tom Ziemke (Eds.). Springer London, London, 387–392

    1998

  26. [34]

    Pentti Kanerva. 2009. Hyperdimensional computing: An introduction to computing in dis- tributed representation with high-dimensional random vectors. Cognitive computation 1 (2009), 139–159

    2009

  27. [35]

    Daniel Kang, Deepti Raghavan, Peter Bailis, and Matei Zaharia. 2018. Model assertions for debugging machine learning. In NeurIPS MLSys Workshop, V ol. 3. 10

    2018

  28. [36]

    Daniel Kang, Deepti Raghavan, Peter Bailis, and Matei Zaharia. 2020. Model assertions for monitoring and improving ml models. Proceedings of Machine Learning and Systems 2 (2020), 481–496

    2020

  29. [37]

    Been Kim, Martin Wattenberg, Justin Gilmer, Carrie Cai, James Wexler, Fernanda Viegas, et al

  30. [38]

    Pieter-Jan Kindermans, Kristof T Schütt, Maximilian Alber, Klaus-Robert Müller, Dumitru Erhan, Been Kim, and Sven Dähne. 2017. Learning how to explain neural networks: Patternnet and patternattribution. arXiv preprint arXiv:1705.05598 (2017)

    2017 arXiv

  31. [39]

    Alex Krizhevsky. 2009. Learning multiple layers of features from tiny images. Technical Report. Citeseer

    2009

  32. [40]

    Ya Le and Xuan Yang. 2015. Tiny imagenet visual recognition challenge. CS 231N 7, 7 (2015), 3

    2015

  33. [41]

    Yann LeCun, Léon Bottou, Yoshua Bengio, and Patrick Haffner. 1998. Gradient-based learning applied to document recognition. Proc. IEEE 86, 11 (1998), 2278–2324

    1998

  34. [42]

    Ji Lin, Ligeng Zhu, Wei-Ming Chen, Wei-Chen Wang, Chuang Gan, and Song Han. 2022. On-device training under 256kb memory. arXiv preprint arXiv:2206.15472 (2022)

    2022 arXiv

  35. [43]

    Jialong Liu, Mingyuan Ma, Zhenhua Zhu, Yu Wang, and Huazhong Yang. 2019. Hdc-im: Hyperdimensional computing in-memory architecture based on rram. In 2019 26th IEEE International Conference on Electronics, Circuits and Systems (ICECS). IEEE, 450–453

    2019

  36. [44]

    Maqueda, Carlos R

    Antonio Loquercio, Ana I. Maqueda, Carlos R. del Blanco, and Davide Scaramuzza. 2018. DroNet: Learning to Fly by Driving. IEEE Robotics and Automation Letters 3, 2 (2018), 1088–1095. https://doi.org/10.1109/LRA.2018.2795643

    2018

  37. [45]

    Dongning Ma, Cong Hao, and Xun Jiao. 2024. Hyperdimensional computing vs. neural networks: Comparing architecture and learning process. In 2024 25th International Symposium on Quality Electronic Design (ISQED). IEEE, 1–5

    2024

  38. [46]

    Justin Morris, Mohsen Imani, Samuel Bosch, Anthony Thomas, Helen Shu, and Tajana Rosing

  39. [47]

    Daniel Holanda Noronha, Zhiqiang Que, Wayne Luk, and Steven J. E. Wilton. 2021. Flex- ible Instrumentation for Live On-Chip Debug of Machine Learning Training on FPGAs. In 2021 IEEE 29th Annual International Symposium on Field-Programmable Custom Computing Machines (FCCM). 20–...

    2021

  40. [48]

    Keiichi Ochiai, Kohei Senkawa, Naoki Yamamoto, Yuya Tanaka, and Yusuke Fukazawa. 2019. Real-time On-Device Troubleshooting Recommendation for Smartphones. In Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (Anchorage, AK, USA) (...

    2019

  41. [49]

    Patil, Don Kurian Dennis, Chirag Pabbaraju, Nadeem Shaheer, Harsha Vardhan Simhadri, Vivek Seshadri, Manik Varma, and Prateek Jain

    Shishir G. Patil, Don Kurian Dennis, Chirag Pabbaraju, Nadeem Shaheer, Harsha Vardhan Simhadri, Vivek Seshadri, Manik Varma, and Prateek Jain. 2019. GesturePod: Enabling On-Device Gesture-Based Interaction for White Cane Users. In Proceedings of the 32nd Annual ACM Symposium o...

    2019

  42. [50]

    Hang Qiu, Ioanna Vavelidou, Jian Li, Evgenya Pergament, Pete Warden, Sandeep Chinchali, Zain Asgar, and Sachin Katti. 2022. ML-EXray: Visibility into ML deployment on the edge. Proceedings of Machine Learning and Systems 4 (2022), 337–351

    2022

  43. [51]

    Alexander Redding, Xiaofan Yu, Shengfan Hu, Pat Pannuto, and Tajana Rosing. 2023. EmbHD: A Library for Hyperdimensional Computing Research on MCU-Class Devices. In Proceedings of the 2nd Workshop on Networked Sensing Systems for a Sustainable Society. 187–192

    2023

  44. [52]

    Haoyu Ren, Darko Anicic, and Thomas A Runkler. 2021. Tinyol: Tinyml with online-learning on microcontrollers. In 2021 international joint conference on neural networks (IJCNN). IEEE, 1–8

    2021

  45. [53]

    Burr Settles. 2009. Active learning literature survey. (2009)

    2009

  46. [54]

    Shital Shah, Roland Fernandez, and Steven Drucker. 2019. A system for real-time interac- tive analysis of deep learning training. In Proceedings of the ACM SIGCHI Symposium on Engineering Interactive Computing Systems. 1–6

    2019

  47. [55]

    Jost Tobias Springenberg, Alexey Dosovitskiy, Thomas Brox, and Martin Riedmiller. 2014. Striving for simplicity: The all convolutional net. arXiv preprint arXiv:1412.6806 (2014)

    2014 arXiv

  48. [56]

    Mukund Sundararajan, Ankur Taly, and Qiqi Yan. 2017. Axiomatic attribution for deep networks. In International conference on machine learning. PMLR, 3319–3328

    2017

  49. [57]

    Jorge Vargas, Suleiman Alsweiss, Onur Toker, Rahul Razdan, and Joshua Santos. 2021. An overview of autonomous vehicles sensors and their vulnerability to weather conditions. Sensors 21, 16 (2021), 5397

    2021

  50. [58]

    Junyao Wang, Sitao Huang, and Mohsen Imani. 2023. Disthd: A learner-aware dynamic encod- ing method for hyperdimensional classification. In 2023 60th ACM/IEEE Design Automation Conference (DAC). IEEE, 1–6

    2023

  51. [59]

    Pete Warden. 2018. Speech Commands: A Dataset for Limited-V ocabulary Speech Recognition. CoRR abs/1804.03209 (2018). arXiv:1804.03209 http://arxiv.org/abs/1804.03209

    2018 arXiv

  52. [60]

    Pete Warden and Daniel Situnayake. 2019. TinyML. O’Reilly Media, Incorporated

    2019

  53. [61]

    Kanit Wongsuphasawat, Daniel Smilkov, James Wexler, Jimbo Wilson, Dandelion Mane, Doug Fritz, Dilip Krishnan, Fernanda B Viégas, and Martin Wattenberg. 2017. Visualizing dataflow graphs of deep learning models in tensorflow. IEEE transactions on visualization and computer grap...

    2017

  54. [62]

    Jingjing Yang, Yuanning Li, Yonghong Tian, Lingyu Duan, and Wen Gao. 2009. Group- sensitive multiple kernel learning for object categorization. In Computer Vision, 2009 IEEE 12th International Conference on. IEEE, 436–443. 8

    2009

  55. [63]

    Matthew D Zeiler and Rob Fergus. 2014. Visualizing and understanding convolutional networks. In Computer Vision–ECCV 2014: 13th European Conference, Zurich, Switzerland, September 6-12, 2014, Proceedings, Part I 13. Springer, 818–833

    2014

  56. [64]

    Yundong Zhang, Naveen Suda, Liangzhen Lai, and Vikas Chandra. 2017. Hello Edge: Keyword Spotting on Microcontrollers. https://doi.org/10.48550/ARXIV.1711.07128

    2017 doi arXiv

  57. [65]

    Zhuowen Zou, Yeseong Kim, Farhad Imani, Haleh Alimohamadi, Rosario Cammarota, and Mohsen Imani. 2021. Scalable edge-based hyperdimensional learning system with brain- like neural adaptation. In Proceedings of the International Conference for High Performance Computing, Network...

    2021

  58. [69]

    SpeechCommands Speech-Cmd is a collection of short audio clips, each spanning 1 second

    SpeechCmd [59], 2) CIFAR10 [39] and 3) TinyImagenet [40]. SpeechCommands Speech-Cmd is a collection of short audio clips, each spanning 1 second. The dataset consists of utterances for 35 words and is commonly used for benchmarking keyword spotting systems. We train our model ...

  59. [70]

    The base network has been trained only on ID

    First, we iterate through all the corrupted-ID images for each dataset by passing each corrupted image through the pretrained base network. The base network has been trained only on ID

  60. [71]

    We use this to create our train and test datasets for the HDC classifier denoted by DHDC −train and DHDC −test respectively

    Second, we select an appropriate intermediate layer of the base network, and record the output of this intermediate layer as we iterate through the corrupted images. We use this to create our train and test datasets for the HDC classifier denoted by DHDC −train and DHDC −test ...

  61. [72]

    A.3.1 Selecting an appropriate intermediate layer Since tinyML models often have only a few layers, the number of intermediate layers available for tapping is limited

    Next, we use DHDC −train to train the HDC classifier as described in Section 3, and use DHDC −test for validation/testing. A.3.1 Selecting an appropriate intermediate layer Since tinyML models often have only a few layers, the number of intermediate layers available for tappin...

  62. [2018]

    In International conference on machine learning

    Interpretability beyond feature attribution: Quantitative testing with concept activation vectors (tcav). In International conference on machine learning. PMLR, 2668–2677

  63. [2019]

    In 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED)

    CompHD: Efficient hyperdimensional computing using model compression. In 2019 IEEE/ACM International Symposium on Low Power Electronics and Design (ISLPED). IEEE, 1–6. 7

    2019

  64. [2024]

    In TinyML for Edge Intelligence in IoT and LPWAN Networks , Bharat S

    Chapter 14 - TinyML applications and use cases for healthcare. In TinyML for Edge Intelligence in IoT and LPWAN Networks , Bharat S. Chaudhari, Sheetal N. Ghor- pade, Marco Zennaro, and Rytis Paškauskas (Eds.). Academic Press, 331–353. https: //doi.org/10.1016/B978-0-44-322202...

    doi

Pith tools

Reviewed August 12, 2026 · model on record in the stance chip above.