Pith. sign in

REVIEW 3 major objections 5 minor 42 references

Optimising TinyML with Quantization and Distillation of Transformer and Mamba Models for Indoor Localisation on Edge Devices

T0 review · 3 major / 5 minor · reviewed 2026-08-11 · deepseek-v4-flash

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

Pith's one-line read Indoor localisation models can be compressed down to 32–64 KB of microcontroller memory.

desk verdict A systematically useful TinyML comparison whose RAM-feasibility headline is a proxy, not a measurement. read the letter →

arxiv 2412.09289 v1 pith:S4GQLCNS submitted 2024-12-12 cs.LG cs.SE

Thanaphon Suwannaphong , Ferdian Jovan , Ian Craddock , Ryan McConville This is my paper · ORCID
classification cs.LGcs.SE
keywords TinyMLIndoorlocalisationQuantizationKnowledgedistillationMambaTransformerEdgedevicesMicrocontrollers
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

The paper sets out to show that advanced sequence models for indoor localisation can be compressed enough to run on low-power microcontrollers with only 32–64 KB of RAM. It takes a state-of-the-art transformer-based model (MDCSA), a compact Mamba state-space model, and applies post-training quantization plus knowledge distillation to shrink them. The central claim is that the quantized transformer performs well within a 64 KB budget for in-home room-level localisation, while a small Mamba model does well under a 32 KB budget without needing compression. On a large multi-building Wi-Fi dataset, the Mamba model becomes the better choice under 64 KB. The authors argue this makes on-device indoor localisation viable for wearable health monitoring, with lower latency and better privacy than sending data to a server.

What carries the argument

The machinery is the size–performance trade-off curve obtained by varying hidden size $H$ and number of layers $L$ for two architectures: the Multihead Dual Convolutional Self-Attention (MDCSA) transformer, and a Mamba structured state-space model with linear embedding and a conditional random field (CRF) classification head. Post-training quantization — both static LLM.int8() and dynamic weight-only quantization — converts FP32 linear layers to int8, and knowledge distillation trains small students against the best-performing teacher using a weighted cross-entropy loss with $\alpha=0.1$. The model size in kilobytes is the quantity that ties the whole argument together: it is what must fall below the 64 KB or 32 KB memory budget while the F1 score stays competitive.

What would settle it

Measure peak RAM usage on an actual low-power MCU with 64 KB of RAM while running the quantized MDCSA H16L1 model, and on a 32 KB MCU while running the Mamba H8L1 and H16L1 models on the same in-home and UJIIndoorLoc test sets; if peak usage exceeds the respective budgets during inference, the feasibility claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is a sizing map: sweeping hidden size and layer count produces transformer and Mamba models whose static sizes land under 64 KB and 32 KB, and post-training quantization brings the best transformer variant (MDCSA H16L1) down to about 44 KB with F1 scores of 73.84–84.36 across four homes, while a one-layer Mamba with hidden size 8 or 16 fits in 12–21 KB and outperforms the transformer in the stricter 32 KB band. For the larger UJIIndoorLoc dataset, the one-layer Mamba with hidden size 8 and knowledge distillation is the strongest model under 64 KB (44 KB, F1 64%), while the transformer does not fit even after quantization. The paper also reports that quantization can inflate the size of already-tiny models because of the 32-bit scale and zero-point overhead, and that simple output-level knowledge distillation yields only marginal gains. The concluding recommendation is architecture-dependent: transformer with quantization for 64 KB in-home use, unquantized compact Mamba for 32 KB in-home use, and Mamba with distillation for large-building localisation under 64 KB.

Load-bearing premise

The claim that the models fit in 32–64 KB of RAM rests on comparing static model size in kilobytes against the MCU's RAM budget, without measuring what the device actually consumes at runtime (activations, buffers, framework overhead), and the paper's own future-work section says validation on physical devices remains to be done.

Editorial extensions

If this is right

  • If the claims hold, a wearable with 64 KB of RAM can run room-level localisation in homes using a quantized transformer, keeping location data on the device.
  • Under a 32 KB budget, a compact Mamba model (hidden size 8 or 16, one layer) is the recommended choice for in-home use, without requiring quantization.
  • For multi-building, high-dimensional Wi-Fi settings, the Mamba architecture with knowledge distillation is the viable option under 64 KB; the transformer baseline is not.
  • Quantization should be applied selectively: on models already under roughly 12 KB it can increase size rather than reduce it.
  • Simple final-layer knowledge distillation does not by itself deliver large gains, so its value in this setting is limited to cases like the large dataset where it nudges performance.

Reading between the lines

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

  • Editorial extension: the 32–64 KB numbers are static model sizes, not measured on-device RAM; validating the same models on a physical MCU (with activations and runtime buffers) would be the direct test of the feasibility claim the paper leaves for future work.
  • Editorial extension: on datasets with hundreds of access points, the input dimension itself may become the dominant memory cost, so input-side reduction (AP selection or feature compression) could push Mamba below 32 KB for large buildings — a path the paper does not explore.
  • Editorial extension: the reported quantization overhead suggests a general rule for TinyML: below some model-size threshold, architecture shrinkage beats precision reduction, and the threshold can be identified by comparing the cost of scale/zero-point parameters against the weight savings.
  • Editorial extension: the finding that final-layer distillation helps less than expected points to layer-wise or feature-based distillation as a testable follow-up, and the paper itself names this direction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper develops TinyML versions of a transformer-based model (MDCSA) and a Mamba-based state-space model for RSSI indoor localisation, targeting low-power MCUs with 32-64 KB memory budgets. It applies post-training quantization (a claimed static LLM.int8() scheme and PyTorch dynamic quantization), knowledge distillation, and a hybrid of both, sweeping hidden size and layer count on four in-home BLE houses and on UJIIndoorLoc. The main reported results are F1, accuracy, and model size in KB for each configuration; the central claim is that the quantized MDCSA model fits in 64 KB RAM and the compact Mamba model fits in 32 KB RAM without compression.

Significance. If the deployment claim were substantiated, the paper would make a useful empirical contribution: it is, to my knowledge, the first systematic comparison of Mamba versus transformer architectures for indoor localisation under TinyML constraints, it covers two public datasets, and it reports a consistent sweep of model sizes with a clear comparison of quantization, distillation, and their combination. The repository link is a welcome element. However, the headline feasibility claim is currently supported only by static model size, not by measured memory on hardware, so the significance remains conditional; the comparative F1 results are useful, but the deployment claim needs either direct measurement or substantial qualification.

major comments (3)
  1. [Abstract; Tables 4-8; 'Complexity and Practical Challenges of Model Deployment'] The paper's central claim that the quantized transformer 'performs well within a 64 KB RAM constraint' and that Mamba 'has strong performance under ... 32 KB of RAM' is not established by the reported evidence. The only support is the 'Size (KB)' columns in Tables 4-8, but no definition of this size is given and no on-device measurement is reported. Static checkpoint size omits activation tensors, CRF/Viterbi buffers, the Mamba selective-SSM scan state, the input window, and the inference interpreter or runtime arena. The paper's own 'Complexity and Practical Challenges of Model Deployment' section states that 'profiling and testing are essential to evaluate memory usage, inference time, and power consumption on the target hardware' and concludes that 'Future work will include validating these models on physical edge devices.' The RAM feasibility claims should be replaced by 'model storage size' claims, or supported by actual peak-memory measurements on a representative MCU with a deployment framework.
  2. [Methodology, 'Quantization'; 'Limitations of Quantization'] The method used for the 'static quant' columns is LLM.int8(), cited as Dettmers et al. (2022); this is a mixed FP16/int8 outlier-decomposition inference method developed for GPU transformer workloads, not a static integer-only quantization scheme for MCU kernels. The paper does not explain how FP16 outlier values, scale, and zero-point are stored, nor how the Mamba selective scan and CRF layers would be mapped to integer or otherwise MCU-compatible operations. Consequently, the 'static quant' size values do not by themselves establish compatibility with low-power MCUs. Please clarify the intended deployment path or adjust the TinyML feasibility claim accordingly.
  3. [Evaluations, 'Model Size'] The model-size metric is not reproducible as reported. The Evaluations section states only that model size is measured in KB and that the number of parameters is reported, without stating whether the value is the FP32 parameter count, the state-dict byte size, the serialized artifact size, or the size after framework conversion. Since the central claims in Tables 4-8 hinge on this number, the exact computation or measurement tool must be specified, and the relationship between this number and actual RAM usage should be stated explicitly.
minor comments (5)
  1. [Experimental Setting] No random seed or number of repeated runs is reported; F1 differences of 1-2% between configurations should be interpreted with caution, and adding variance estimates or seeds would strengthen the comparisons.
  2. [Table 8] The MDCSA H16L1 static-quant row shows accuracy collapsing from 63.37% to 19.05% while F1 remains around 46%; the text notes the drop, but the magnitude deserves an explanation, such as calibration sensitivity or the interaction of outlier handling with majority classes.
  3. [Methodology, 'Knowledge Distillation'] The distillation loss in Equations (4)-(5) uses the teacher's hard class predictions rather than soft probability targets; this is a legitimate choice given the CRF output, but the method should be described as hard-label distillation to avoid confusion with conventional soft-label KD.
  4. [Tables 4-7] The asterisk notation is confusing: in Table 4, MDCSA H16L1 has a baseline size of 64 KB, which does not strictly exceed a 64 KB constraint, yet it is marked with an asterisk indicating that the size exceeds the memory constraint; the caption wording should be clarified.
  5. [Related Work] MDCSA, used as the 'state-of-the-art' transformer baseline, is the authors' own prior model; including at least one external transformer baseline would make the state-of-the-art claim more convincing.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the paper's claims are empirical benchmarks, the self-cited MDCSA baseline is used only as a comparison model, and the RAM-feasibility caveat is a verification gap, not a circular step.

full rationale

This paper is an empirical benchmarking study rather than a derivation. The central claims are measured F1/accuracy scores and reported model sizes across architectures and compression techniques. No result is obtained by fitting a parameter and then renaming it as a prediction; the distillation weight alpha is tuned on validation and then applied, which is standard practice. The transformer baseline MDCSA is cited from the authors' prior work, but it is used as an architecture to compress and compare, not as evidence that the current results hold. The paper's own limitation sections explicitly state that profiling on target hardware and validation on physical edge devices remain future work, which is a verification gap concerning the static-size proxy for RAM usage, not a circularity in the argument. The only potential concern, equating checkpoint size with RAM feasibility, is an unverified proxy but not a definitional or self-referential reduction. Therefore, no circular steps exist and the paper is self-contained as an empirical comparison.

Assumptions & free parameters 3 free parameters · 3 assumptions · 0 invented entities

The central claim is empirical, so the main ledger items are the proxy assumption connecting model size to RAM, the tuned distillation weight, and the architecture hyperparameters. No new physical or conceptual entities are introduced.

free parameters (3)
  • alpha (distillation weight) = 0.1
    Selected as the proportion between student loss and distillation loss because it gave highest validation performance (Methodology, Knowledge Distillation).
  • in-home window size = 4 seconds, 50% overlap
    Choice of temporal window for in-home data preprocessing; affects input shape and model size (Data Pre-Processing).
  • hidden size H and number of layers L = e.g., H16L1 for best transformer under 64 KB; H8L1 or H16L1 for best Mamba under 32 KB
    Architecture size hyperparameters swept to find models under the 64 KB and 32 KB constraints; selection is based on validation F1 (Experimental Setting).
assumptions (3)
  • domain assumption Model size in KB approximates MCU RAM usage
    The paper equates fitting within a 64 KB RAM constraint with model weight size in KB, but actual RAM usage includes activations, runtime buffers, and framework overhead; no on-device profiling is performed (Evaluations, Model Size; Conclusion).
  • domain assumption RSSI fingerprints are stable enough for room-level classification
    Fingerprinting assumes that measured signal strengths are sufficiently repeatable across time and devices for the classification labels to be learnable; the paper relies on this standard assumption without testing it.
  • domain assumption MDCSA from the authors' prior work is a valid state-of-the-art baseline
    The MDCSA model is adopted from the authors' own paper [11] without independent replication, and is used as the transformer baseline for all comparisons (Related Work, Methodology).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Optimising TinyML with Quantization and Distillation of Transformer and Mamba Models for Indoor Localisation on Edge Devices." pith.science (2026). https://pith.science/paper/S4GQLCNS

@misc{pith2026241209289,
  author       = {Pith},
  title        = {Pith review of: Optimising TinyML with Quantization and Distillation of Transformer and Mamba Models for Indoor Localisation on Edge Devices},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S4GQLCNS}},
  note         = {Machine review of arXiv:2412.09289}
}
read the original abstract

This paper proposes small and efficient machine learning models (TinyML) for resource-constrained edge devices, specifically for on-device indoor localisation. Typical approaches for indoor localisation rely on centralised remote processing of data transmitted from lower powered devices such as wearables. However, there are several benefits for moving this to the edge device itself, including increased battery life, enhanced privacy, reduced latency and lowered operational costs, all of which are key for common applications such as health monitoring. The work focuses on model compression techniques, including quantization and knowledge distillation, to significantly reduce the model size while maintaining high predictive performance. We base our work on a large state-of-the-art transformer-based model and seek to deploy it within low-power MCUs. We also propose a state-space-based architecture using Mamba as a more compact alternative to the transformer. Our results show that the quantized transformer model performs well within a 64 KB RAM constraint, achieving an effective balance between model size and localisation precision. Additionally, the compact Mamba model has strong performance under even tighter constraints, such as a 32 KB of RAM, without the need for model compression, making it a viable option for more resource-limited environments. We demonstrate that, through our framework, it is feasible to deploy advanced indoor localisation models onto low-power MCUs with restricted memory limitations. The application of these TinyML models in healthcare has the potential to revolutionize patient monitoring by providing accurate, real-time location data while minimizing power consumption, increasing data privacy, improving latency and reducing infrastructure costs.

Figures

Figures reproduced from arXiv: 2412.09289 by the authors.

Figure 1
Figure 1. Mamba architecture for RSSI classification [PITH_FULL_IMAGE:figures/full_fig_p009_1.png] view at source ↗
Figure 2
Figure 2. Illustration shows an example case when a quantized model is bigger than the full model due [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 34 canonical work pages

  1. [1]

    Tinyml applications and use cases for healthcare,

    M. Bhamare, P. V. Kulkarni, R. Rane, S. Bobde, and R. Patankar, “Tinyml applications and use cases for healthcare,” inTinyML for Edge Intelligence in IoT and LPWAN Networks, pp. 331–353, Elsevier, 2024

    work page 2024

  2. [2]

    Activity monitoring and location sensory system for people with mild cognitive impairments,

    A. García-Requejo, M. C. Pérez-Rubio, J. M. Villadangos, and Á. Hernández, “Activity monitoring and location sensory system for people with mild cognitive impairments,”IEEE Sensors Journal, vol. 23, no. 5, pp. 5448–5458, 2023

    work page 2023

  3. [3]

    Survey of indoor location technologies and wayfinding systems for users with cognitive disabilities in emergencies,

    M. García-Catalá, M. C. Rodriguez-Sánchez, and E. Martín-Barroso, “Survey of indoor location technologies and wayfinding systems for users with cognitive disabilities in emergencies,”Behaviour & Information Technology, vol. 41, no. 4, pp. 879–903, 2022

    work page 2022

  4. [4]

    A tinyml deep learning approach for indoor tracking of assets,

    D. Avellaneda, D. Mendez, and G. Fortino, “A tinyml deep learning approach for indoor tracking of assets,” Sensors, vol. 23, no. 3, p. 1542, 2023

    work page 2023

  5. [5]

    Detecting Signatures of Early-stage Dementia with Behavioural Models Derived from Sensor Data

    R. Poyiadzi, W. Yang, Y. Ben-Shlomo, I. Craddock, L. Coulthard, R. Santos-Rodriguez, J. Selwood, and N. Twomey, “Detecting signatures of early-stage dementia with behavioural models derived from sensor data,” arXiv preprint arXiv:2007.03615, 2020

    work page Pith review arXiv 2007

  6. [6]

    Vesta: A digital health analytics platform for a smart home in a box,

    R. McConville, G. Archer, I. Craddock, M. Kozłowski, R. Piechocki, J. Pope, and R. Santos- Rodriguez, “Vesta: A digital health analytics platform for a smart home in a box,”Future Generation Computer Systems, vol. 114, pp. 106 – 119, 2021

    work page 2021

  7. [7]

    Tinyml using neural networks for resource-constrained devices,

    V. Kulkarni and V. Jujare, “Tinyml using neural networks for resource-constrained devices,” in TinyML for Edge Intelligence in IoT and LPWAN Networks, pp. 87–101, Elsevier, 2024

    work page 2024

  8. [8]

    A survey of quantiza- tion methods for efficient neural network inference,

    A. Gholami, S. Kim, Z. Dong, Z. Yao, M. W. Mahoney, and K. Keutzer, “A survey of quantiza- tion methods for efficient neural network inference,” inLow-Power Computer Vision, pp. 291–326, Chapman and Hall/CRC, 2022

    work page 2022

Show all 42 references
  1. [9]

    Distilling the knowledge in a neural network,

    G. Hinton, O. Vinyals, and J. Dean, “Distilling the knowledge in a neural network,”arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  2. [10]

    Machine learning techniques for indoor localization on edge devices: Integrating ai with embedded devices for indoor localization purposes,

    D. Méndez, D. Crovo, and D. Avellaneda, “Machine learning techniques for indoor localization on edge devices: Integrating ai with embedded devices for indoor localization purposes,” inTinyML for Edge Intelligence in IoT and LPWAN Networks, pp. 355–376, Elsevier, 2024

    2024

  3. [11]

    Multimodal indoor localisation in parkinson’s disease for detecting medication use: Observational pilot study in a free- living setting,

    F. Jovan, C. Morgan, R. McConville, E. L. Tonkin, I. Craddock, and A. Whone, “Multimodal indoor localisation in parkinson’s disease for detecting medication use: Observational pilot study in a free- living setting,” inProceedings of the 29th ACM SIGKDD Conference on Knowledge ...

    2023

  4. [12]

    A review on tinyml: State-of-the-art and prospects,

    P. P. Ray, “A review on tinyml: State-of-the-art and prospects,”Journal of King Saud University- Computer and Information Sciences, vol. 34, no. 4, pp. 1595–1623, 2022

    2022

  5. [13]

    To prune, or not to prune: exploring the efficacy of pruning for model compression,

    M. Zhu and S. Gupta, “To prune, or not to prune: exploring the efficacy of pruning for model compression,” arXiv preprint arXiv:1710.01878, 2017

    2017 arXiv

  6. [14]

    Low-rank matrix factorization for deep neural network training with high-dimensional output targets,

    T. N. Sainath, B. Kingsbury, V. Sindhwani, E. Arisoy, and B. Ramabhadran, “Low-rank matrix factorization for deep neural network training with high-dimensional output targets,” in2013 IEEE international conference on acoustics, speech and signal processing, pp. 6655–6659, IEEE, 2013

    2013

  7. [15]

    A compre- hensive survey on tinyml,

    Y. Abadade, A. Temouden, H. Bamoumen, N. Benamar, Y. Chtouki, and A. S. Hafid, “A compre- hensive survey on tinyml,”IEEE Access, 2023

    2023

  8. [16]

    Model compression via distillation and quantization,

    A. Polino, R. Pascanu, and D. Alistarh, “Model compression via distillation and quantization,”arXiv preprint arXiv:1802.05668, 2018

    2018 arXiv

  9. [17]

    Channel state information based device free wireless sensing for iot devices employing tinyml,

    A. M. Hayajneh, S. Aldalahmeh, S. A. R. Zaidi, D. McLernon, H. Obeidollah, and R. Alsakarnah, “Channel state information based device free wireless sensing for iot devices employing tinyml,” in 2022 4th IEEE Middle East and North Africa COMMunications Conference (MENACOMM), pp...

    2022

  10. [18]

    Design space exploration of a multi-model ai-based indoor localization system,

    K. Kotrotsios, A. Fanariotis, H.-C. Leligou, and T. Orphanoudakis, “Design space exploration of a multi-model ai-based indoor localization system,”Sensors, vol. 22, no. 2, p. 570, 2022

    2022

  11. [19]

    A tinyml-approach to detect the proximity of people based on bluetooth low energy beacons,

    M. Girolami, F. Fattori, and S. Chessa, “A tinyml-approach to detect the proximity of people based on bluetooth low energy beacons,” in2023 19th International Conference on Intelligent Environ- ments (IE), pp. 1–4, IEEE, 2023

    2023

  12. [20]

    Tiny but mighty: Embedded machine learning for indoor wireless localization,

    B. Jones, U. Raza, and A. Khan, “Tiny but mighty: Embedded machine learning for indoor wireless localization,” in 2023 IEEE 20th Consumer Communications & Networking Conference (CCNC), pp. 176–181, IEEE, 2023

    2023

  13. [21]

    A fast indoor positioning using a knowledge-distilled convolutional neural network (kd-cnn),

    A. B. Mazlan, Y. H. Ng, and C. K. Tan, “A fast indoor positioning using a knowledge-distilled convolutional neural network (kd-cnn),”IEEE access, vol. 10, pp. 65326–65338, 2022

    2022

  14. [22]

    Teacher-assistant knowledge distillation based indoor positioning system,

    A. B. Mazlan, Y. H. Ng, and C. K. Tan, “Teacher-assistant knowledge distillation based indoor positioning system,” Sustainability, vol. 14, no. 21, p. 14652, 2022

    2022

  15. [23]

    Knowledge distillation for a lightweight deep learning-based indoor positioning system on edge environments,

    A. G. Putrada, N. Alamsyah, S. F. Pane, M. N. Fauzan, and D. Perdana, “Knowledge distillation for a lightweight deep learning-based indoor positioning system on edge environments,” in2023 International Seminar on Intelligent Technology and Its Applications (ISITIA), pp. 370–37...

    2023

  16. [24]

    Knowledge distillation based deep learning model for user equipment positioning in massive mimo systems using flying reconfigurable intelligent surfaces,

    A. Al-Ahmadi, “Knowledge distillation based deep learning model for user equipment positioning in massive mimo systems using flying reconfigurable intelligent surfaces,”IEEE Access, 2024

    2024

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polo- sukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  18. [26]

    A survey of transformers,

    T. Lin, Y. Wang, X. Liu, and X. Qiu, “A survey of transformers,”AI open, vol. 3, pp. 111–132, 2022

    2022

  19. [27]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    2023 arXiv

  20. [28]

    Timemachine: A time series is worth 4 mambas for long-term forecasting,

    M. A. Ahamed and Q. Cheng, “Timemachine: A time series is worth 4 mambas for long-term forecasting,” arXiv preprint arXiv:2403.09898, 2024

    2024 arXiv

  21. [29]

    Residential wearable rssi and accelerometer measurements with detailed location annotations,

    D. Byrne, M. Kozlowski, R. Santos-Rodriguez, R. Piechocki, and I. Craddock, “Residential wearable rssi and accelerometer measurements with detailed location annotations,”Scientific data, vol. 5, no. 1, pp. 1–14, 2018

    2018

  22. [30]

    Ujiindoorloc: A new multi-building and multi-floor database for wlan fingerprint-based indoor localization problems,

    J. Torres-Sospedra, R. Montoliu, A. Martínez-Usó, J. P. Avariento, T. J. Arnau, M. Benedito- Bordonau, and J. Huerta, “Ujiindoorloc: A new multi-building and multi-floor database for wlan fingerprint-based indoor localization problems,” in2014 international conference on indoo...

    2014

  23. [31]

    A survey on indoor positioning systems for iot-based applications,

    P. S. Farahsari, A. Farahzadi, J. Rezazadeh, and A. Bagheri, “A survey on indoor positioning systems for iot-based applications,”IEEE Internet of Things Journal, vol. 9, no. 10, pp. 7680–7699, 2022

    2022

  24. [32]

    Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,

    T. Dettmers, M. Lewis, Y. Belkada, and L. Zettlemoyer, “Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,”Advances in Neural Information Processing Systems, vol. 35, pp. 30318– 30332, 2022

    2022

  25. [33]

    A survey of techniques for optimizing transformer inference,

    K. T. Chitty-Venkata, S. Mittal, M. Emani, V. Vishwanath, and A. K. Somani, “A survey of techniques for optimizing transformer inference,”Journal of Systems Architecture, p. 102990, 2023

    2023

  26. [34]

    Mamba in speech: Towards an alternative toself-attention,

    X. Zhang, Q. Zhang, H. Liu, T. Xiao, X. Qian, B. Ahmed, E. Ambikairajah, H. Li, and J. Epps, “Mamba in speech: Towards an alternative toself-attention,”arXiv preprint arXiv:2405.12609, 2024

    2024 arXiv

  27. [35]

    Saturn: Sample-efficient generative molecular design using memory ma- nipulation,

    J. Guo and P. Schwaller, “Saturn: Sample-efficient generative molecular design using memory ma- nipulation,” arXiv preprint arXiv:2405.17066, 2024

    2024 arXiv

  28. [36]

    Accurate post training quantiza- tion with small calibration sets,

    I. Hubara, Y. Nahshan, Y. Hanani, R. Banner, and D. Soudry, “Accurate post training quantiza- tion with small calibration sets,” inInternational Conference on Machine Learning, pp. 4466–4475, PMLR, 2021. 18

    2021

  29. [37]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference,

    B. Jacob, S. Kligys, B. Chen, M. Zhu, M. Tang, A. Howard, H. Adam, and D. Kalenichenko, “Quantization and training of neural networks for efficient integer-arithmetic-only inference,” in Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 2704–2713, 2018

    2018

  30. [38]

    Up or down? adap- tive rounding for post-training quantization,

    M. Nagel, R. A. Amjad, M. Van Baalen, C. Louizos, and T. Blankevoort, “Up or down? adap- tive rounding for post-training quantization,” in International Conference on Machine Learning, pp. 7197–7206, PMLR, 2020

    2020

  31. [39]

    Less is more: Task-aware layer-wise distillation for language model compression,

    C. Liang, S. Zuo, Q. Zhang, P. He, W. Chen, and T. Zhao, “Less is more: Task-aware layer-wise distillation for language model compression,” in International Conference on Machine Learning, pp. 20852–20867, PMLR, 2023

    2023

  32. [40]

    Fitnets: Hints for thin deep nets,

    A. Romero, N. Ballas, S. E. Kahou, A. Chassang, C. Gatta, and Y. Bengio, “Fitnets: Hints for thin deep nets,”arXiv preprint arXiv:1412.6550, 2014

    2014 arXiv

  33. [41]

    Darkrank: Accelerating deep metric learning via cross sample similarities transfer,

    Y. Chen, N. Wang, and Z. Zhang, “Darkrank: Accelerating deep metric learning via cross sample similarities transfer,” inProceedings of the AAAI conference on artificial intelligence, vol. 32, 2018

    2018

  34. [42]

    Categories of response-based, feature-based, and relation- based knowledge distillation,

    C. Yang, X. Yu, Z. An, and Y. Xu, “Categories of response-based, feature-based, and relation- based knowledge distillation,” inAdvancements in Knowledge Distillation: Towards New Horizons of Intelligent Systems, pp. 1–32, Springer, 2023. 19

    2023

Pith tools

Reviewed August 11, 2026 · model on record in the stance chip above.