Pith. sign in

REVIEW 4 major objections 5 minor 51 references

Post-Training Quantization of Generative and Discriminative LSTM Text Classifiers: A Study of Calibration, Class Balance, and Robustness

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

Pith's one-line read Generative LSTM text classifiers lose far more accuracy than discriminative ones under post-training quantization, and only remain viable below 6 bits when the calibration data is class-balanced.

desk verdict The empirical trend is plausible, but the paper's central mechanism—class imbalance in calibration—is not actually tested, because a 25% random sample of balanced AG News/DBPedia is nearly class-balanced. read the letter →

arxiv 2507.09687 v1 pith:D7ZKWSRK submitted 2025-07-13 cs.LG cs.AI

classification cs.LGcs.AI MSC 68T0768T50
keywords post-trainingquantizationgenerativeclassifiersdiscriminativeLSTMtextclassificationcalibrationdataclassimbalancenoiserobustness
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 tries to establish that generative LSTM text classifiers degrade far more than discriminative ones when post-training quantization (PTQ) is applied, and that the composition of the small calibration set used to set activation scales is a main driver of that degradation. On standard datasets, the generatively trained classifier falls sharply below 6-bit precision when calibration data is sampled without regard to class balance, while the discriminative classifier stays accurate down to 3 bits. The paper argues the mechanism is that class-imbalanced calibration yields skewed activation statistics, and the greedy refinement step then makes insufficient weight corrections in the final linear layer. If this is right, practitioners deploying generative classifiers on edge devices must use class-balanced calibration data, and the apparent robustness of generative classifiers to noise does not survive aggressive quantization. The finding matters because generative classifiers are otherwise attractive for noisy, safety-sensitive edge settings.

What carries the argument

The load-bearing mechanism is the calibration step of PTQ followed by Greedy Path-Following Quantization (GPFQ), a data-dependent post-training refinement that adjusts rows of a weight matrix to minimize the forward-pass error induced by quantization on a batch of calibration inputs. In this study GPFQ is applied only to the final linear layer, since the quantization library does not support it for the recurrent LSTM layers. The paper's diagnostic tool is the Kolmogorov–Smirnov (KS) statistic, a nonparametric measure of the largest distance between two distributions; it is used to compare weight distributions before and after calibration and to compare activation distributions produced by class-conditional versus class-unconditional calibration data. The argument proceeds by showing that unbalanced calibration shifts activation statistics relative to balanced calibration, and that those skewed statistics lead GPFQ to make smaller corrective adjustments at low bitwidths, leaving quantized weights misaligned with full-precision behavior.

What would settle it

Measure the per-class composition of the 25 percent random calibration sets on AG News and DBPedia; if those sets are nearly class-balanced while the accuracy gap between class-conditional and class-unconditional calibration persists, the proposed imbalance mechanism is not the explanation, and the decisive control would be to draw multiple random calibration sets with matched class counts and show that accuracy tracks the class distribution rather than the seed.

Watch

Extended reading notes

Core claim

The paper's central claim is that discriminative LSTM text classifiers remain robust under post-training quantization while generative LSTM classifiers are markedly more sensitive, with sensitivity appearing as a function of three things: bitwidth, the class composition of the calibration data, and input noise during quantized inference. Using a controlled comparison of class-conditional and class-unconditional calibration, the authors show that the generative classifier's accuracy collapses below 6 bits precisely when calibration is unbalanced, and that balancing the calibration set restores most of the lost accuracy, even reversing the ordering so that the generative model beats the discriminative one at several bitwidths. The authors attribute the failure to insufficient weight adaptation in the final linear layer during greedy path-following quantization, a consequence of activation statistics that do not represent all classes. They further show that the generative classifier's full-precision robustness to character-level noise erodes after quantization, especially at 3 and 4 bits.

Load-bearing premise

The claim that class imbalance in calibration causes the generative classifier's collapse assumes that the random 'class-unconditional' calibration set is actually class-imbalanced, but the paper never reports its empirical class distribution, so the effect could be driven by which particular samples were drawn rather than by class balance.

Editorial extensions

If this is right

  • Class-balanced calibration data should be treated as a requirement, not an option, for generative LSTM classifiers deployed below 6-bit precision.
  • Discriminative LSTM classifiers are comparatively indifferent to calibration class composition, so existing PTQ pipelines that ignore class balance are safer for them.
  • Full-precision robustness to noisy input is not preserved under quantization for generative classifiers; 3-bit and 4-bit generative models lose their noise advantage over discriminative ones.
  • At several bitwidths with class-conditional calibration, the generative classifier outperforms the discriminative one, so the two model families' rankings can flip depending on calibration quality.
  • Training-time noise injection can partially restore low-bitwidth robustness, suggesting robustness and quantization decisions interact.

Reading between the lines

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

  • The paper never reports the empirical class distribution of the 'class-unconditional' 25 percent sample; because AG News and DBPedia training sets are class-balanced, a random sample should be nearly balanced, so the observed gap may partly reflect seed or sample-selection effects rather than class imbalance per se.
  • Because GPFQ is applied only to the final linear layer, the demonstrated 'insufficient weight adaptation' concerns that layer; whether LSTM gate and cell weights would exhibit the same imbalance sensitivity is untested and is the most direct next experiment.
  • The results suggest a cheap deployment rule: when labels are unavailable, approximating class balance in calibration data with unsupervised clustering or class-prior estimates may recover much of the generative classifier's low-bitwidth accuracy.
  • The model comparison is limited to a single 100-hidden-unit LSTM configuration; the same calibration-sensitivity pattern may or may not transfer to transformer text classifiers, which is where edge NLP is moving.
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. This manuscript reports an empirical study of post-training quantization (PTQ) for LSTM-based text classifiers, comparing a discriminative classifier (embedding, LSTM, linear layer, softmax) with a generative classifier that scores each class by the autoregressive likelihood of the input under a class-conditioned LSTM. The authors train both models on AG News and DBPedia, then quantize weights and activations to 3-8 bits using Brevitas with class-unconditional (random) or class-conditional (stratified) calibration and GPFQ refinement on the final linear layer. They measure test accuracy across bitwidths, character-level input noise at several rates, and internal weight/activation shifts via KS statistics. The main claims are that discriminative classifiers remain robust under PTQ while generative classifiers are more sensitive to bitwidth, calibration composition, and input noise, and that class-imbalanced calibration data causes 'insufficient weight adaptation' at low bitwidths in generative models.

Significance. If the central claim were established, the paper would offer actionable guidance for deploying generative LSTM classifiers on edge devices: use class-balanced calibration and avoid aggressive low-bit quantization. The study covers two standard datasets, multiple bitwidths, a controlled class-coverage experiment, and a systematic KS-statistic analysis, which are useful contributions. However, the main causal claim currently rests on a comparison that may be confounded by calibration-set size and by the fact that random sampling from balanced datasets produces nearly balanced calibration sets; the paper does not report the empirical class distribution, seeds, or repeated runs. The controlled class-coverage experiment (Section 6.1) supports a weaker claim that extreme class coverage matters, but not the specific mechanism invoked for the Section 5.3-5.4 gap.

major comments (4)
  1. [Sections 5.3 and 5.4] The comparison that motivates the paper's central claim is confounded. AG News and DBPedia training sets are class-balanced, so a 25% random calibration sample is expected to be nearly class-balanced (for AG News, per-class proportions around 25% with a standard deviation below 1 percentage point on the calibration sample). The large accuracy gap between Figures 6 and 7 therefore cannot be attributed to class imbalance unless the authors report the actual class counts in the sampled calibration sets, the random seed, and repeated draws. Without this information, the observed gap may be caused by which particular samples were drawn or by other differences between the two calibration protocols. I ask the authors to report these details and to supplement the comparison with a matched experiment in which calibration sets have identical total size and controlled class proportions.
  2. [Section 5.4] The definition of class-conditional calibration is ambiguous. '25% from each class in a four-class classification task' could mean 25% of each class (preserving the same total calibration size as the class-unconditional setting) or 25% of the calibration set per class (making the calibration set four times larger). These two readings change the calibration-set size, which is known to affect PTQ accuracy, and would confound the comparison in Figure 7. Please specify the sample count per class and total sample size for both calibration protocols.
  3. [Section 4.3.3 and Figures 9-10] The noise injection procedure is not reproducible as stated. Section 4.3.3 says 'for each character in the input sentence, a substitution is performed with probability epsilon' but then says 'the replacement character is sampled uniformly from the model's vocabulary.' The vocabulary is word-level, so replacing a character with a vocabulary token is not defined; if word-level substitution was intended, the text should say so. In addition, the x-axes of Figures 9 and 10 label epsilon as 'Noise Level (Variance),' but epsilon is defined as a substitution probability. Please clarify the noise model and relabel the axes.
  4. [Section 6.1.1] The KS-statistic argument for 'insufficient weight adaptation' does not establish the claimed mechanism. A smaller KS distance between quantized and full-precision weights at low bitwidths (Figure 12) is interpreted as insufficient adaptation, but it could equally indicate that GPFQ made smaller changes because the calibration data were less representative, or because the optimization objective on a skewed sample did not demand larger changes. The authors should connect the KS magnitudes to accuracy directly, e.g., by showing that the class-unconditional calibration data are actually class-imbalanced and that repeated sampling with matched calibration size reproduces the accuracy gap.
minor comments (5)
  1. [Section 5.1] There are typos in the manuscript, including 'trianing' in Section 5.1, 'calibraiton' in Section 6.1, and 'calibtrated' in Section 6.1.1; these should be corrected.
  2. [Figure 15] The caption says 'loss difference' but the figure appears to show KDEs of token-level loss; please clarify whether the plotted quantity is the loss itself or a difference relative to full precision.
  3. [Sections 4.2.3 and 4.3.2] The text states GPFQ is applied only to the final linear layer, but the general formulation in Section 4.2.3 describes arbitrary layers; it would help to state explicitly in Section 4.2.3 that the implementation restricts GPFQ to the final linear layer.
  4. [Figures 12-14] The KS statistics are reported without p-values or confidence intervals; the abstract's phrase 'test statistics derived from nonparametric hypothesis testing' suggests formal tests, but no test outcomes are reported.
  5. [General] The paper does not mention whether code or model configurations are available; given the number of implementation details, a reproducibility appendix would strengthen the paper.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the paper's PTQ comparisons are measured empirical outcomes, not fitted or self-referential reductions.

full rationale

The paper reports measured comparisons of full-precision and post-training-quantized LSTM classifiers across bitwidths, calibration schemes, and noise levels. The central claims—that generative classifiers degrade more under class-unconditional calibration, that KS statistics reveal different weight and activation shifts, and that class-balanced calibration helps—are empirical results from experiments implemented in Brevitas; they are not derived by construction from a fitted parameter or from an assumed identity. The quantization and GPFQ equations (Eqs. 15-22) are standard PTQ definitions, not premises that entail the accuracy results. The paper contains a few self-citations (e.g., [35], [48]) used to motivate underexplored areas or cite prior quantization methods, but no load-bearing reduction relies on them: the robustness comparisons and KS analyses are independently measured. The main methodological weakness is experimental rather than circular: on the balanced AG News and DBPedia datasets, a random 25% calibration sample is nearly class-balanced, so the Section 5.3 versus 5.4 accuracy gap may be confounded by specific sample draws, seeds, or calibration set size. That is a validity threat, not a circularity. Therefore no step in the paper's derivation chain reduces to its own input.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new model entities are introduced. The paper's claims rest on three hand-chosen experimental constants (calibration fraction, clipping percentile, noise level), on the Bayes-rule definition of generative classification, on the chosen noise model, and on the interpretive leap that KS distribution shifts measure adaptation sufficiency.

free parameters (3)
  • calibration set fraction = 0.25 (25% of training data)
    Chosen after observing that increasing beyond 25% did not improve results (Section 5.3); the central comparison depends on this fraction.
  • activation clipping percentile = 99.99th percentile
    Fixed choice for percentile-based activation range estimation (Section 4.3.2); sensitivity to this value is not studied.
  • training-time noise level = 0.1 (Figure 16)
    The noise-injection-in-training experiment in the conclusions uses a single noise level with no methodology description, and results are only shown in one figure.
assumptions (4)
  • domain assumption Generative classification via Bayes rule: argmax_y p(x|y), with p(x|y) modeled autoregressively (Equations 6 and 7).
    The whole comparison rests on this definition of a generative LSTM classifier.
  • domain assumption Random character substitution with replacement uniformly sampled from the vocabulary models real-world input noise.
    Section 4.3.3 defines noise this way; the robustness conclusions depend on this noise model being representative.
  • domain assumption KS statistic between weight or activation distributions is a valid proxy for the sufficiency of GPFQ weight adaptation.
    Sections 6.1.1 and 6.1.2 interpret KS values as evidence of adaptation quality, but KS measures distribution shift, not adaptation quality.
  • standard math Standard PTQ mathematics from Brevitas and GPFQ as described in Section 4.2.
    Quantization equations (15) through (21) and the GPFQ objective (22) are standard background.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Post-Training Quantization of Generative and Discriminative LSTM Text Classifiers: A Study of Calibration, Class Balance, and Robustness." pith.science (2026). https://pith.science/paper/D7ZKWSRK

@misc{pith2026250709687,
  author       = {Pith},
  title        = {Pith review of: Post-Training Quantization of Generative and Discriminative LSTM Text Classifiers: A Study of Calibration, Class Balance, and Robustness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D7ZKWSRK}},
  note         = {Machine review of arXiv:2507.09687}
}
read the original abstract

Text classification plays a pivotal role in edge computing applications like industrial monitoring, health diagnostics, and smart assistants, where low latency and high accuracy are both key requirements. Generative classifiers, in particular, have been shown to exhibit robustness to out-of-distribution and noisy data, which is an extremely critical consideration for deployment in such real-time edge environments. However, deploying such models on edge devices faces computational and memory constraints. Post Training Quantization (PTQ) reduces model size and compute costs without retraining, making it ideal for edge deployment. In this work, we present a comprehensive comparative study of generative and discriminative Long Short Term Memory (LSTM)-based text classification models with PTQ using the Brevitas quantization library. We evaluate both types of classifier models across multiple bitwidths and assess their robustness under regular and noisy input conditions. We find that while discriminative classifiers remain robust, generative ones are more sensitive to bitwidth, calibration data used during PTQ, and input noise during quantized inference. We study the influence of class imbalance in calibration data for both types of classifiers, comparing scenarios with evenly and unevenly distributed class samples including their effect on weight adjustments and activation profiles during PTQ. Using test statistics derived from nonparametric hypothesis testing, we identify that using class imbalanced data during calibration introduces insufficient weight adaptation at lower bitwidths for generative LSTM classifiers, thereby leading to degraded performance. This study underscores the role of calibration data in PTQ and when generative classifiers succeed or fail under noise, aiding deployment in edge environments.

Figures

Figures reproduced from arXiv: 2507.09687 by the authors.

Figure 1
Figure 1. Post-training quantization workflow. Full Precision Model is trained, followed by quantization, calibration, [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Text preprocessing pipeline for both generative and discriminative classifiers. [PITH_FULL_IMAGE:figures/full_fig_p009_2.png] view at source ↗
Figure 3
Figure 3. Dataflow in the discriminative LSTM classifier during inference. Trainable layers are shaded in blue. [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figures from the paper (13 more)
Figure 4
Figure 4. Figure 4: Dataflow in the generative LSTM classifier during inference. Trainable layers are shaded in blue. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Deployment workflow for post-training quantization. The FP model is trained offline/in cloud, after which it [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Quantization Accuracy of Generative and discriminative classifiers using Class-Unconditional Calibration on [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Quantization accuracy of generative and discriminative classifiers using class-conditional calibration on AG [PITH_FULL_IMAGE:figures/full_fig_p013_7.png]
Figure 8
Figure 8. Figure 8: Accuracy drop with respect to full-precision (32-bit) model for both generative and discriminative classifiers [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Accuracy vs. Noise Level for Discriminative and generative classifiers on AG News and DBPedia datasets. [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Noise robustness of quantized discriminative and generative classifiers. Accuracy is measured across varying [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Accuracy of generative and discriminative classifiers under different calibration sampling schemes across [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: KS statistic comparison of weight shifts in generative and discriminative classifiers after class-conditional vs. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: KS statistic between class-conditional (CC) and class-unconditional (UC) data activations in the embedding [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: KS statistic between pre-calibrated model activations and class-conditional/class-unconditional calibrated [PITH_FULL_IMAGE:figures/full_fig_p019_14.png]
Figure 15
Figure 15. Figure 15: Kernel density plot showing loss difference (maximum values are shown in the legend) of token-level [PITH_FULL_IMAGE:figures/full_fig_p020_15.png]
Figure 16
Figure 16. Figure 16: Impact of training-time noise injection on accuracy across bit-widths. Models are evaluated on noisy input [PITH_FULL_IMAGE:figures/full_fig_p021_16.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

51 extracted references · 33 canonical work pages

  1. [1]

    Pattern recognition and machine learning

    Bishop, C.M., 2006. Pattern recognition and machine learning. volume 4. Springer

  2. [2]

    Understanding and overcoming the challenges of efficient transformer quantization

    Bondarenko, Y ., Nagel, M., Blankevoort, T., 2021. Understanding and overcoming the challenges of efficient transformer quantization. arXiv preprint arXiv:2109.12948

  3. [3]

    Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, march 2023

    Chiang, W.L., Li, Z., Lin, Z., Sheng, Y ., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y ., Gonzalez, J.E., et al., 2023. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, march 2023. URL https://lmsys. org/blog/2023-03-30-vicuna 3

  4. [4]

    A comparative survey of instance selection methods applied to non-neural and transformer-based text classification

    Cunha, W., Viegas, F., França, C., Rosa, T., Rocha, L., Gonçalves, M.A., 2023. A comparative survey of instance selection methods applied to non-neural and transformer-based text classification. ACM Computing Surveys 55, 1–52

  5. [5]

    Latent-Variable Generative Models for Data-Efficient Text Classification

    Ding, X., Gimpel, K., 2019. Latent-variable generative models for data-efficient text classification. arXiv preprint arXiv:1910.00382 URL: https://arxiv.org/abs/1910.00382, arXiv:1910.00382

  6. [6]

    Matrix completion via memoryless scalar quantization

    Eamaz, A., Yeganegi, F., Soltanalian, M., 2023. Matrix completion via memoryless scalar quantization. arXiv preprint arXiv:2311.05052

  7. [7]

    Learned step size quantization

    Esser, S.K., McKinstry, J.L., Bablani, D., Appuswamy, R., Modha, D.S., 2019. Learned step size quantization. arXiv preprint arXiv:1902.08153

  8. [8]

    Xilinx/brevitas

    Franco, G., Pappalardo, A., Fraser, N.J., 2025. Xilinx/brevitas. URL: https://doi.org/10.5281/zenodo. 3333552, doi:10.5281/zenodo.3333552

Show all 51 references
  1. [9]

    The elements of statistical learning: data mining, inference, and prediction

    Hastie, T., Tibshirani, R., Friedman, J.H., Friedman, J.H., 2009. The elements of statistical learning: data mining, inference, and prediction. volume 2. Springer

  2. [10]

    Long short-term memory

    Hochreiter, S., Schmidhuber, J., 1997. Long short-term memory. Neural computation 9, 1735–1780

  3. [11]

    spacy: Industrial-strength natural language processing in python

    Honnibal, M., Montani, I., Van Landeghem, S., Boyd, A., et al., 2020. spacy: Industrial-strength natural language processing in python

  4. [12]

    Fastai: A layered api for deep learning

    Howard, J., Gugger, S., 2020. Fastai: A layered api for deep learning. Information 11, 108

  5. [13]

    Empirical evaluation of post-training quantization methods for language tasks

    Hu, T., Meinel, C., Yang, H., 2022. Empirical evaluation of post-training quantization methods for language tasks. arXiv preprint arXiv:2210.16621 . 21 A PREPRINT - SEPTEMBER 4, 2025

  6. [14]

    Accurate post training quantization with small calibration sets, in: International Conference on Machine Learning, PMLR

    Hubara, I., Nahshan, Y ., Hanani, Y ., Banner, R., Soudry, D., 2021. Accurate post training quantization with small calibration sets, in: International Conference on Machine Learning, PMLR. pp. 4466–4475

  7. [15]

    Speech and Language Processing

    Jurafsky, D., Martin, J.H., 2023. Speech and Language Processing. 3rd ed., Pearson. Draft version available at https://web.stanford.edu/~jurafsky/slp3/

  8. [16]

    Sulla determinazione empirica di una legge di distribuzione

    Kolmogorov, A.N., 1933. Sulla determinazione empirica di una legge di distribuzione. Giornale dell’Istituto Italiano degli Attuari 4, 83–91

  9. [17]

    Generative models improve fairness of medical classifiers under distribution shifts

    Ktena, I., Wiles, O., Albuquerque, I., Rebuffi, S.A., Tanno, R., Roy, A.G., Azizi, S., Belgrave, D., Kohli, P., Cemgil, T., et al., 2024. Generative models improve fairness of medical classifiers under distribution shifts. Nature Medicine 30, 1166–1173

  10. [18]

    Precision and recall metrics for generative models, in: Advances in Neural Information Processing Systems

    Kynkäänniemi, T., Karras, T., Laine, S., Lehtinen, J., Aila, T., 2019. Precision and recall metrics for generative models, in: Advances in Neural Information Processing Systems. URL: https://proceedings.neurips.cc/ paper_files/paper/2018/file/f7696a9b362ac5a51c3dc8f098b73923-Paper.pdf

  11. [19]

    Robust inference via generative classifiers for handling noisy labels, in: Proceedings of the 36th International Conference on Machine Learning (ICML)

    Lee, K., Yun, S., Lee, K., Lee, H., Li, B., Shin, J., 2019. Robust inference via generative classifiers for handling noisy labels, in: Proceedings of the 36th International Conference on Machine Learning (ICML)

  12. [20]

    Fp8-bert: Post-training quantization for transformer

    Li, J., Zhang, T., Yen, I.E.H., Xu, D., 2023a. Fp8-bert: Post-training quantization for transformer. arXiv preprint arXiv:2312.05725

  13. [21]

    On the impact of calibration data in post-training quantization and pruning

    Li, S., Ganesh, V ., Yao, Z., Jain, P., Gholami, A., Gonzalez, J.E., 2023b. On the impact of calibration data in post-training quantization and pruning. arXiv preprint arXiv:2311.09755

  14. [22]

    Are generative classifiers more robust to adversarial attacks? arXiv preprint arXiv:1802.06552

    Li, Y ., Bradshaw, J., Sharma, Y ., 2018. Are generative classifiers more robust to adversarial attacks? arXiv preprint arXiv:1802.06552

  15. [23]

    Spinquant: Llm quantization with learned rotations

    Liu, Z., Zhao, C., Fedorov, I., Soran, B., Choudhary, D., Krishnamoorthi, R., Chandra, V ., Tian, Y ., Blankevoort, T., 2024. Spinquant: Llm quantization with learned rotations. arXiv preprint arXiv:2405.16406

  16. [24]

    A greedy algorithm for quantizing neural networks

    Lybrand, E., Saab, R., 2021. A greedy algorithm for quantizing neural networks. Journal of Machine Learning Research 22, 1–38. URL: http://jmlr.org/papers/v22/20-1233.html. arXiv:2010.15979

  17. [25]

    Efficient estimation of word representations in vector space

    Mikolov, T., Chen, K., Corrado, G., Dean, J., 2013. Efficient estimation of word representations in vector space. arXiv preprint arXiv:1301.3781

  18. [26]

    Up or down? adaptive round- ing for post-training quantization

    Nagel, M., Amjad, R.A., van Baalen, M., Louizos, C., Blankevoort, T., 2020. Up or down? adaptive round- ing for post-training quantization. CoRR abs/2004.10568. URL: https://arxiv.org/abs/2004.10568, arXiv:2004.10568

  19. [27]

    Data-free quantization through weight equalization and bias correction, in: Proceedings of the IEEE/CVF international conference on computer vision, pp

    Nagel, M., Baalen, M.v., Blankevoort, T., Welling, M., 2019. Data-free quantization through weight equalization and bias correction, in: Proceedings of the IEEE/CVF international conference on computer vision, pp. 1325–1334

  20. [28]

    A white paper on neural network quantization

    Nagel, M., et al., 2021. A white paper on neural network quantization. arXiv preprint arXiv:2106.08295

  21. [29]

    On discriminative vs

    Ng, A., Jordan, M., 2001. On discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes. Advances in neural information processing systems 14

  22. [30]

    Deep neural networks are easily fooled: High confidence predictions for unrecognizable images, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp

    Nguyen, A., Yosinski, J., Clune, J., 2015. Deep neural networks are easily fooled: High confidence predictions for unrecognizable images, in: Proceedings of the IEEE conference on computer vision and pattern recognition, pp. 427–436

  23. [31]

    On estimation of a probability density function and mode

    Parzen, E., 1962. On estimation of a probability density function and mode. The Annals of Mathematical Statistics 33, 1065–1076

  24. [32]

    Pytorch: An imperative style, high-performance deep learning library

    Paszke, A., 2019. Pytorch: An imperative style, high-performance deep learning library. arXiv preprint arXiv:1912.01703

  25. [33]

    GloVe: Global vectors for word representation, in: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp

    Pennington, J., Socher, R., Manning, C.D., 2014. GloVe: Global vectors for word representation, in: Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pp. 1532–1543

  26. [34]

    Remarks on some nonparametric estimates of a density function

    Rosenblatt, M., 1956. Remarks on some nonparametric estimates of a density function. The Annals of Mathemati- cal Statistics 27, 832–837

  27. [35]

    Generative and discriminative deep belief network classifiers: Comparisons under an approximate computing framework

    Ruan, S., Colbert, I., Kreutz-Delgado, K., Das, S., 2021. Generative and discriminative deep belief network classifiers: Comparisons under an approximate computing framework. arXiv preprint arXiv:2102.00534

  28. [36]

    Table for estimating the goodness of fit of empirical distributions

    Smirnov, N., 1948. Table for estimating the goodness of fit of empirical distributions. Annals of Mathematical Statistics 19, 279–281

  29. [37]

    How to fine-tune bert for text classification?, in: China national conference on Chinese computational linguistics, Springer

    Sun, C., Qiu, X., Xu, Y ., Huang, X., 2019. How to fine-tune bert for text classification?, in: China national conference on Chinese computational linguistics, Springer. pp. 194–206. 22 A PREPRINT - SEPTEMBER 4, 2025

  30. [38]

    Llama: Open and efficient foundation language models

    Touvron, H., Lavril, T., Izacard, G., Martinet, X., Lachaux, M.A., Lacroix, T., Rozière, B., Goyal, N., Hambro, E., Azhar, F., et al., 2023. Llama: Open and efficient foundation language models. arXiv preprint arXiv:2302.13971

  31. [39]

    Attention is all you need

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, Ł., Polosukhin, I., 2017. Attention is all you need. Advances in neural information processing systems 30

  32. [40]

    Haq: Hardware-aware automated quantization with mixed precision

    Wang, K., Liu, Z., Wu, Y ., Yang, J., Han, S., 2019. Haq: Hardware-aware automated quantization with mixed precision. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , 8612–8620arXiv:1811.08886

  33. [41]

    The effect of class imbalance on precision-recall curves

    Ward, R., Williams, C.K., 2021. The effect of class imbalance on precision-recall curves. Neu- ral Computation 33, 853–889. URL: https://direct.mit.edu/neco/article/33/4/853/97475/ The-Effect-of-Class-Imbalance-on-Precision-Recall , doi:10.1162/neco_a_01370

  34. [42]

    Easyquant: Post-training quantization via scale optimization, in: CVPR

    Xu, C., Yuan, L., Wu, Y ., Yu, W., Wang, L., 2022. Easyquant: Post-training quantization via scale optimization, in: CVPR

  35. [43]

    Zeroquant: Efficient and affordable post-training quantization for large-scale transformers

    Yao, Z., Yazdani Aminabadi, R., Zhang, M., Wu, X., Li, C., He, Y ., 2022. Zeroquant: Efficient and affordable post-training quantization for large-scale transformers. Advances in Neural Information Processing Systems 35, 27168–27183

  36. [44]

    Generative and discriminative text classification with recurrent neural networks

    Yogatama, D., Dyer, C., Ling, W., Blunsom, P., 2017. Generative and discriminative text classification with recurrent neural networks. arXiv preprint arXiv:1703.01898 URL: https://arxiv.org/abs/1703.01898, arXiv:1703.01898

  37. [45]

    Q8bert: Quantized 8bit bert

    Zafrir, O., Boudoukh, G., Izsak, P., Wasserblat, M., 2019. Q8bert: Quantized 8bit bert. arXiv preprint arXiv:1910.06188

  38. [46]

    Opt: Open pre-trained transformer language models

    Zhang, S., Roller, S., Goyal, N., Artetxe, M., et al., 2022a. Opt: Open pre-trained transformer language models. arXiv preprint arXiv:2205.01068

  39. [47]

    Qronos: Correcting the past by shaping the future

    Zhang, S., Zhang, H., Colbert, I., Saab, R., 2025a. Qronos: Correcting the past by shaping the future... in post-training quantization. arXiv preprint arXiv:2505.11695

  40. [48]

    Learning low-precision structured subnetworks using joint layerwise channel pruning and uniform quantization

    Zhang, X., Colbert, I., Das, S., 2022b. Learning low-precision structured subnetworks using joint layerwise channel pruning and uniform quantization. Applied Sciences 12, 7829

  41. [49]

    Character-level convolutional networks for text classification, in: Advances in Neural Information Processing Systems

    Zhang, X., Zhao, J., LeCun, Y ., 2015. Character-level convolutional networks for text classification, in: Advances in Neural Information Processing Systems. URL: https://arxiv.org/abs/1509.01626

  42. [50]

    A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification

    Zhang, Y ., Wallace, B.C., 2015. A sensitivity analysis of (and practitioners’ guide to) convolutional neural networks for sentence classification. arXiv preprint arXiv:1510.03820

  43. [51]

    Selectq: Calibration data selection for post-training quantization

    Zhang, Z., Gao, Y ., Fan, J., Zhao, Z., Yang, Y ., Yan, S., 2025b. Selectq: Calibration data selection for post-training quantization. Machine Intelligence Research , 1–12. 23

Pith tools

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