Pith. sign in

REVIEW 3 major objections 4 minor 51 references

Designing Practical Models for Isolated Word Visual Speech Recognition

T0 review · 3 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The submission's abstract describes lightweight visual speech recognition models, but its full text is a database tuning paper; the only evidence-based claim in the document is about automatic HNSW configuration tuning.

desk verdict The submission is a mismatch: the abstract promises a VSR paper, but the full text is a different arXiv paper on proximity-graph tuning, so there is nothing coherent to review. read the letter →

arxiv 2508.17894 v1 pith:FJWDTMLI submitted 2025-08-25 cs.CV cs.AIcs.CL

Iason Ioannis Panagos , Giorgos Sfikas , Christophoros Nikou This is my paper
classification cs.CVcs.AIcs.CL
keywords visualspeechrecognitionlightweightarchitecturesproximitygraphsHNSWconfigurationtuningqueryperformancepredictionreinforcementlearningmodeltransfersubmissionmismatch
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

Taken as a single document, this submission does not make one coherent claim: the abstract promises lightweight end-to-end models for isolated-word visual speech recognition, while the full text is a system paper for PGTuner, an automatic tuner for proximity-graph approximate nearest neighbor search. The body's claim is that PGTuner can learn to recommend HNSW construction and query parameters that reach a user-specified recall with maximum queries per second, and can transfer that skill to new or growing datasets without building many throwaway graphs. If true, vector databases could reach near-optimal accuracy-versus-speed settings with far less tuning, with reported speedups over the standard grid-search baseline of up to about 14.7 times. The abstract-level claim about VSR—that light image-classification backbones inside a temporal convolution network give strong word recognition at low hardware cost—appears nowhere in the full text, which contains no VSR architecture, dataset, training protocol, or recognition result.

What carries the argument

The machinery is a pair of learned components plus a transfer gate. The query performance prediction (QPP) model is a four-layer MLP that predicts recall and average distance computations from dataset features and a configuration, replacing the expensive step of actually building and querying a graph. The parameter configuration recommendation (PCR) model is a TD3 deep-reinforcement-learning agent that proposes configurations and is rewarded for satisfying the target recall and then reducing predicted cost. A data-similarity detector, based on nearest-neighbor distances in the predictor's embedding space, decides whether a new dataset is close enough to the training distribution; if not, an

What would settle it

Run the pretrained predictor/recommender on a held-out dataset, take its recommended configuration, build the actual HNSW graph, and measure recall and QPS: if the recall misses the target or the QPS trails grid search by a meaningful margin, the transferability claim fails. For the abstract, the test is simpler: locate the promised VSR dataset split, training protocol, or word-recognition accuracy in the full text—none exists.

Watch

Extended reading notes

Core claim

The full text proposes PGTuner, a two-model tuning pipeline for proximity graphs such as HNSW. A query performance prediction model maps a dataset description plus a parameter configuration to expected recall and a device-independent cost proxy (average number of distance computations); a deep reinforcement learning recommender is trained on that predictor's outputs and learns a policy for proposing configurations that first hit the target recall, then minimize cost. The paper reports that on six real-world datasets this pipeline matches or beats exhaustive grid search in achieved QPS at most target recalls while cutting tuning time by up to 14.69x, and that in a data-growth scenario it reac

Load-bearing premise

The body's central claim collapses if a predictor trained on a few small base datasets cannot estimate recall and cost accurately enough on new, larger datasets for the learned recommender to match exhaustive search; separately, the submission's abstract assumes the full text is the visual speech recognition paper, which it is not.

Editorial extensions

If this is right

  • HNSW indexes in vector databases could be tuned per dataset and per target recall without repeated graph construction, making near-optimal recall/throughput trade-offs practical on large collections.
  • When a dataset grows or its query workload changes, the framework reuses its pretrained predictor and stops retraining as soon as similarity is detected, which the paper reports as up to a 14.64x speedup in dynamic scenarios.
  • Changing the target recall after a dataset is transferred only requires fine-tuning the recommender, so recall preferences could be adjusted without re-running the whole tuning pipeline.
  • The same training loop transfers beyond HNSW to another proximity graph family (the paper tests NSG with six tunable parameters), suggesting the approach is not tied to one index.

Reading between the lines

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

  • If the intended submission was the visual speech recognition paper named in the abstract, then no conclusion about lightweight VSR models can be drawn from this document; the abstract's architecture, dataset, and results are not present in the full text.
  • The speedup numbers rest on using average distance computations as a cost proxy rather than measured latency; a reader transferring the method to new hardware should verify the proxy tracks actual queries per second.
  • A natural test of the transfer mechanism is to run the predictor/recommender loop on a dataset deliberately dissimilar from all base datasets and measure whether the active-retraining rounds recover grid-search-level QPS; the paper's early-termination design predicts that it should, but only for datasets the similarity detector can recognize.
  • The body's tuning approach could be applied to the resource trade-offs in the abstract's domain: visual speech recognition model families have few controllable cost/accuracy knobs, and the same two-model predictor/recommender loop could auto-select backbone depth and temporal-window size, though the paper does not attempt this.
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

3 major / 4 minor

Summary. The abstract and metadata describe a paper on isolated-word visual speech recognition (VSR), proposing lightweight end-to-end architectures based on efficient image-classification backbones and temporal convolution networks, with experiments on the largest public English word database. The supplied full text, however, is a different paper: it is titled "PGTuner: An Efficient Framework for Automatic and Transferable Configuration Tuning of Proximity Graphs," self-identifies as arXiv:2508.17886v1 [cs.DB], and is wholly concerned with tuning HNSW/NSG parameters for approximate nearest neighbor search. There is no VSR architecture, no dataset, no training protocol, and no recognition result anywhere in the body. The central claim of the submission is therefore unsupported by any content in the manuscript.

Significance. If a VSR paper with the abstract's claims existed, lightweight models for isolated-word VSR at low hardware cost would be a practically useful contribution, potentially enabling deployment on constrained devices. The claimed benchmarking of efficient image-classification backbones inside a temporal convolution network is also a reasonable research direction. However, as submitted, there is zero evidence for the central claim. The body's PGTuner content may be a legitimate ANNS tuning system, but it is not the VSR paper described by the abstract and cannot provide supporting evidence for that claim. The manuscript in its current form is untestable and does not advance the stated VSR research question.

major comments (3)
  1. [Manuscript header / Full Text] The submitted full text is not the paper described by the abstract. The body self-identifies as 'arXiv:2508.17886v1 [cs.DB] 25 Aug 2025' with title 'PGTuner', and every section concerns configuration tuning of proximity graphs. None of the abstract's VSR content appears: no architecture, no 'temporal convolution network backbone', no dataset, no training/evaluation protocol, and no recognition results. This is a load-bearing mismatch: the central claim of the submission is entirely unsupported by the manuscript text.
  2. [Section 5 / Tables 1–3] The experimental evaluation in the body reports QPS, tuning time, recall/ADCN predictions, and dynamic-scenario experiments for HNSW and NSG. Even if these results were entirely sound, they measure performance of a configuration-tuning framework for ANNS, not visual speech recognition accuracy. The abstract's promise of 'Experiments on the largest public database for English words' has no corresponding section, table, or figure in the full text. Thus the body's empirical content cannot be used to evaluate the claimed VSR contribution.
  3. [Whole manuscript] Because the manuscript body is a different paper, there is no way to assess the VSR claim's correctness, novelty, or reproducibility. This is not a fixable local issue such as a missing equation or a plot; the actual VSR paper (architecture, datasets, training details, results) is absent. The submission cannot be reviewed as a VSR paper without a complete rewrite/replacement of the manuscript.
minor comments (4)
  1. [Headers throughout] Many page headers contain garbled mixed-script characters, suggesting an encoding or rendering problem that should be corrected in any resubmission.
  2. [ACM Reference Format] The ACM reference block states 'Conference acronym ’XX, June 03–05, 2018' and '© 2018', while the arXiv header states 25 Aug 2025. These are inconsistent.
  3. [Section 5.3] Typo: 'Constrction Parameter Configurations Selection' should be 'Construction Parameter Configurations Selection'.
  4. [Table 3] The caption says 'The time overhead (s) of PGTuer' (missing 'n'), and the dataset name 'Ntimes' should likely be 'Nytimes'.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the PGTuner body validates recommendations against measured QPS; the metadata/body mismatch is an integrity issue, not a circular derivation.

full rationale

Under the circularity rubric, the provided text contains no derivation step that is equivalent to its own inputs by construction. The abstract describes a VSR paper, but the supplied full text is actually PGTuner, self-identified as arXiv:2508.17886v1 [cs.DB]; the VSR method, dataset, training protocol, and results are entirely absent. That is a fatal support/integrity failure, but it is not circularity—absence of a derivation is not a self-referential derivation. Within the PGTuner content that is present, the headline quantities are externally validated: Section 4.4 post-processes the recommended configuration by actually constructing the PG and executing queries, and Section 5 reports measured QPS and tuning times against baselines. The QPP model is used internally for training and for the Data Similarity Detector's early-stopping heuristic, which is a feedback loop that could be risky, but no reported result is defined as the QPP model's own output; a configuration that fails the target recall is corrected by real search-configuration adjustment. There are no load-bearing self-citations, no imported uniqueness theorems, and no renamed known result. Therefore the circularity score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 1 invented entities

The claimed VSR paper contributes no text to audit, so the ledger can only be filled from the body, which is a different paper (PGTuner). The free parameters are hand-set constants from the body's design sections: k in the feature extractor, k in the data similarity detector, selection rounds R, the 95th percentile threshold, the nine preset target recalls, and the 250-round cap. The body also leans on a pre-training phase of about 208 hours on five datasets whose collected data are reused to judge the similarity of new datasets, an internal feedback worth noting. If the VSR work exists separately, its own ledger (backbone choice, temporal window, feature dimension, loss weights) is entirely unknown to this review.

free parameters (6)
  • Feature-extractor neighborhood size k (DS/DR statistics) = 10
    Chosen according to preliminary experimental results (Section 5.3). DS and DR dataset features depend on k and directly shape QPP model inputs.
  • Data similarity detector k (nearest neighbor rank) = 1
    Set to 1 for the most stringent detection (Section 4.5). Controls whether model transfer is early-terminated, a hand choice with direct effect on tuning time.
  • CPCS selection rounds R = 7
    Set from MAPE saturation experiments (Section 5.3). Controls how many construction configurations are collected during transfer, trading tuning time against QPP accuracy.
  • Distance threshold percentile in Data Similarity Detector = 95th percentile
    Algorithm 1 line 3. Decides whether a new dataset is similar enough to stop transfer; a constant of the method chosen without external justification.
  • PCR preset target recalls = 9 values from 0.85 to 0.99
    Section 3.2. The TD3 agent is trained against these nine targets; generalization to arbitrary user recalls is assumed but only these are demonstrated.
  • Maximum PCR recommendation rounds = 250
    Section 5 settings. A stopping rule that caps fine-tuning cost of the recommendation model.
assumptions (4)
  • domain assumption HNSW query performance is fully determined by efC, M, efS and the dataset, so tuning these three parameters suffices to optimize the recall/QPS trade-off.
    Section 2.1. The entire tuning problem is defined over (efC, M, efS); other structural choices are fixed and never varied.
  • domain assumption ADCN, the average number of distance computations normalized by query set size, is a faithful device-independent proxy for QPS.
    Section 4.1. QPS is discarded as device-dependent; the surrogate ADCN replaces it in the tuning objective, so the reported QPS gains rest on ADCN being proportional to query time.
  • ad hoc to paper The 12-element dataset feature vector (cardinality, dimensionality, LID, DS and DR statistics) captures enough of a dataset's hardness for a pre-trained QPP model to transfer across datasets.
    Section 4.2, Eq. 2. The features are hand-selected; DS and DR are custom statistics justified only by the authors' observations about cliques and distance separability.
  • ad hoc to paper The body text is the manuscript under review.
    The provided text self-identifies as arXiv:2508.17886 (PGTuner), while the metadata identifies 2508.17894 (VSR). The claimed paper's content is absent, so any evaluation must operate on this mismatch.
invented entities (1)
  • QPP model, PCR model, Data Similarity Detector, CPCS (PGTuner components)
    purpose: Predict PG query performance from configurations and dataset features, recommend configurations via TD3, detect dataset similarity for early transfer termination, and select construction configurations for model transfer.
    These are new software components of the body text, not physical entities. Independent evidence would require shipped code with a commit hash or reproducible artifacts; the body links a GitHub URL but ships no commit hash, and the submission metadata describes a different paper. The VSR abstract introduces no new entities at all.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Designing Practical Models for Isolated Word Visual Speech Recognition." pith.science (2026). https://pith.science/paper/FJWDTMLI

@misc{pith2026250817894,
  author       = {Pith},
  title        = {Pith review of: Designing Practical Models for Isolated Word Visual Speech Recognition},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FJWDTMLI}},
  note         = {Machine review of arXiv:2508.17894}
}
read the original abstract

Visual speech recognition (VSR) systems decode spoken words from an input sequence using only the video data. Practical applications of such systems include medical assistance as well as human-machine interactions. A VSR system is typically employed in a complementary role in cases where the audio is corrupt or not available. In order to accurately predict the spoken words, these architectures often rely on deep neural networks in order to extract meaningful representations from the input sequence. While deep architectures achieve impressive recognition performance, relying on such models incurs significant computation costs which translates into increased resource demands in terms of hardware requirements and results in limited applicability in real-world scenarios where resources might be constrained. This factor prevents wider adoption and deployment of speech recognition systems in more practical applications. In this work, we aim to alleviate this issue by developing architectures for VSR that have low hardware costs. Following the standard two-network design paradigm, where one network handles visual feature extraction and another one utilizes the extracted features to classify the entire sequence, we develop lightweight end-to-end architectures by first benchmarking efficient models from the image classification literature, and then adopting lightweight block designs in a temporal convolution network backbone. We create several unified models with low resource requirements but strong recognition performance. Experiments on the largest public database for English words demonstrate the effectiveness and practicality of our developed models. Code and trained models will be made publicly available.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

51 extracted references · 31 canonical work pages

  1. [1]

    Martin Aumüller, Edgar Chavez, and Eric S Tellez. 2023. Overview of the SISAP 2023 Indexing Challenge. In SISAP. Springer Nature Switzerland, Cham, 255–264. https://doi.org/10.1007/978-3-031-46994-7_21

    work page doi:10.1007/978-3-031-46994-7_21 2023

  2. [2]

    Baoqing Cai, Yu Liu, Ce Zhang, Guangyu Zhang, Ke Zhou, Li Liu, Chunhua Li, Bin Cheng, Jie Yang, and Jiashu Xing. 2022. HUNTER: an online cloud database hybrid tuning system for personalized requirements. In SIGMOD. ACM, New York, NY, USA, 646–659. https://doi.org/10.1145/3514221.3517882

    arXiv 2022

  3. [3]

    Juan Elenter, Navid NaderiAlizadeh, and Alejandro Ribeiro. 2022. A lagrangian duality approach to active learning. In NeuIPS. Curran Associates Inc., Red Hook, NY, USA, 37575–37589. https://proceedings.neurips.cc/paper_files/paper/2022/ file/f475bdd151d8b5fa01215aeda925e75c-Paper-Conference.pdf

    work page 2022

  4. [4]

    Cong Fu, Chao Xiang, Changxu Wang, and Deng Cai. 2019. Fast approximate nearest neighbor search with the navigating spreading-out graph. VLDB 12, 5 (2019), 461–474. https://doi.org/10.14778/3303753.3303754

    work page arXiv 2019

  5. [5]

    Scott Fujimoto, Herke Hoof, and David Meger. 2018. Addressing function ap- proximation error in actor-critic methods. In ICML. PMLR, 1587–1596. https: //proceedings.mlr.press/v80/fujimoto18a.html

    work page 2018

  6. [6]

    Denis Gudovskiy, Alec Hodgkinson, Takuya Yamaguchi, and Sotaro Tsukizawa

    work page

  7. [7]

    Dan Hendrycks and Kevin Gimpel. 2016. A baseline for detecting misclassified and out-of-distribution examples in neural networks. https://doi.org/10.48550/ arXiv.1610.02136 arXiv:1610.02136

    arXiv doi:10.48550/arxiv.1610.02136 2016

  8. [8]

    Wenqi Jiang, Shigang Li, Yu Zhu, Johannes de Fine Licht, Zhenhao He, Runbin Shi, Cedric Renggli, Shuai Zhang, Theodoros Rekatsinas, Torsten Hoefler, et al

Show all 51 references
  1. [9]

    Ivan Kobyzev, Simon JD Prince, and Marcus A Brubaker. 2020. Normalizing flows: An introduction and review of current methods. TPAMI 43, 11 (2020), 3964–3979. https://doi.org/10.1109/TPAMI.2020.2992934

    2020

  2. [10]

    Suraj Kothawade, Saikat Ghosh, Sumit Shekhar, Yu Xiang, and Rishabh Iyer. 2022. Talisman: targeted active learning for object detection with rare classes and slices using submodular mutual information. In ECCV. Springer, Berlin, Heidelberg, 1–16. https://doi.org/10.1007/978-3-...

    2022 doi

  3. [11]

    Kimin Lee, Kibok Lee, Honglak Lee, and Jinwoo Shin. 2018. A sim- ple unified framework for detecting out-of-distribution samples and ad- versarial attacks. In NeuIPS. Curran Associates Inc., Red Hook, NY, USA, 7167–7177. https://proceedings.neurips.cc/paper_files/paper/2018/fi...

    2018

  4. [12]

    Yejin Lee, Hyunji Choi, Sunhong Min, Hyunseung Lee, Sangwon Beak, Dawoon Jeong, Jae W Lee, and Tae Jun Ham. 2022. Anna: Specialized architecture for approximate nearest neighbor search. In HPCA. IEEE, 169–183. https://doi.org/ 10.1109/HPCA53966.2022.00021

    2022

  5. [13]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, et al . 2020. Retrieval-augmented generation for knowledge- intensive nlp tasks. In NeurIPS. Curran Associates Inc., Red H...

    2020

  6. [14]

    Conglong Li, Minjia Zhang, David G Andersen, and Yuxiong He. 2020. Improving approximate nearest neighbor search through learned adaptive early termination. In SIGMOD. ACM, New York, NY, USA, 2539–2554. https://doi.org/10.1145/ 3318464.3380600

    2020

  7. [15]

    Dongyuan Li, Zhen Wang, Yankai Chen, Renhe Jiang, Weiping Ding, and Manabu Okumura. 2024. A Survey on Deep Active Learning: Recent Advances and New Frontiers. IEEE Transactions on Neural Networks and Learning Systems 36, 4 (2024), 5879 – 5899. https://doi.org/10.1109/TNNLS.202...

    2024

  8. [16]

    Guoliang Li, Xuanhe Zhou, Shifu Li, and Bo Gao. 2019. Qtune: A query-aware database tuning system with deep reinforcement learning. VLDB 12, 12 (2019), 2118–2130. https://doi.org/10.14778/3352063.3352129

    2019

  9. [17]

    Wen Li, Ying Zhang, Yifang Sun, Wei Wang, Mingjie Li, Wenjie Zhang, and Xuemin Lin. 2019. Approximate nearest neighbor search on high dimensional data—experiments, analyses, and improvement. TKDE 32, 8 (2019), 1475–1488. https://doi.org/10.1109/TKDE.2019.2909204

    2019

  10. [18]

    Ziqian Lin, Sreya Dutta Roy, and Yixuan Li. 2021. Mood: Multi-level out-of- distribution detection. In CVPR. 15313–15323. https://doi.org/10.1109/cvpr46437. 2021.01506

    2021

  11. [19]

    Ting Liu, Andrew Moore, Ke Yang, and Alexander Gray. 2004. An investigation of practical approximate nearest neighbor algorithms. In NeurIPS. MIT Press, Cambridge, MA, USA, 825–832. https://proceedings.neurips.cc/paper_files/ paper/2004/file/1102a326d5f7c9e04fc3c89d0ede88c9-Paper.pdf

    2004

  12. [20]

    Wanqi Liu, Hanchen Wang, Ying Zhang, Wei Wang, Lu Qin, and Xuemin Lin. 2021. EI-LSH: An early-termination driven I/O efficient incremental c-approximate nearest neighbor search. The VLDB Journal 30, 2 (2021), 215–235. https://doi. org/10.1007/s00778-020-00635-4

    2021 doi

  13. [21]

    Weitang Liu, Xiaoyun Wang, John Owens, and Yixuan Li. 2020. Energy-based out- of-distribution detection. In NeuIPS. 21464–21475. https://proceedings.neurips. cc/paper_files/paper/2020/file/f5496252609c43eb8a3d147ab9b9c006-Paper.pdf

    2020

  14. [22]

    Ying Liu, Dengsheng Zhang, Guojun Lu, and Wei-Ying Ma. 2007. A survey of content-based image retrieval with high-level semantics. Pattern recognition 40, 1 (2007), 262–282. https://doi.org/10.1016/j.patcog.2006.04.045

    2007 doi

  15. [23]

    Wei-Liem Loh. 1996. On Latin hypercube sampling. The annals of statistics 24, 5 (1996), 2058–2080. https://doi.org/10.1214/aos/1069362310

    1996

  16. [24]

    Yu A Malkov and Dmitry A Yashunin. 2018. Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs. TPAMI 42, 4 (2018), 824–836. https://doi.org/10.1109/TPAMI.2018.2889473

    2018

  17. [25]

    Yusuke Matsui, Yusuke Uchida, Hervé Jégou, and Shin’ichi Satoh. 2018. A survey of product quantization. ITE Transactions on Media Technology and Applications 6, 1 (2018), 2–10. https://doi.org/10.3169/mta.6.2

    2018 doi

  18. [26]

    Mohamad Amin Mohamadi, Wonho Bae, and Danica J Sutherland. 2022. Making look-ahead active learning strategies feasible with neural tan- gent kernels. In NeuIPS. Curran Associates Inc., Red Hook, NY, USA, �������� �� �������� ��������� ��� ��������� ��� ������������ �����������...

    2022

  19. [27]

    Rafael Seidi Oyamada, Larissa C Shimomura, Sylvio Barbon Junior, and Daniel S Kaster. 2020. Towards proximity graph auto-configuration: An approach based on meta-learning. In ADBIS. Springer, Cham, 93–107. https://doi.org/10.1007/978-3- 030-54832-2_9

    2020 doi

  20. [28]

    Yun Peng, Byron Choi, Tsz Nam Chan, Jianye Yang, and Jianliang Xu. 2023. Efficient approximate nearest neighbor search in multi-dimensional databases. PACMMOD 1, 1 (2023), 1–27. https://doi.org/10.1145/3588908

    2023 doi

  21. [29]

    Jie Ren, Stanislav Fort, Jeremiah Liu, Abhijit Guha Roy, Shreyas Padhy, and Balaji Lakshminarayanan. 2021. A simple fix to mahalanobis distance for improving near-ood detection. https://doi.org/10.48550/arXiv.2106.09022 arXiv:2106.09022

    2021 doi arXiv

  22. [30]

    Jie Ren, Peter J Liu, Emily Fertig, Jasper Snoek, Ryan Poplin, Mark Depristo, Joshua Dillon, and Balaji Lakshminarayanan. 2019. Likelihood ratios for out- of-distribution detection. In NeuIPS. Curran Associates Inc., Red Hook, NY, USA, 14680–14691. https://proceedings.neurips....

    2019

  23. [31]

    David E Rumelhart, Geoffrey E Hinton, and Ronald J Williams. 1986. Learning representations by back-propagating errors. nature 323, 6088 (1986), 533–536. https://doi.org/10.7551/mitpress/1888.003.0013

    1986 doi

  24. [32]

    Ozan Sener and Silvio Savarese. 2017. Active learning for convolutional neu- ral networks: A core-set approach. https://doi.org/10.48550/arXiv.1708.00489 arXiv:1708.00489

    2017 doi arXiv

  25. [33]

    Ján Suchal and Pavol Návrat. 2010. Full text search engine as scalable k-nearest neighbor recommendation system. In IFIP International Conference on Artificial Intelligence in Theory and Practice . Springer, Berlin, Heidelberg, 165–173. https: //doi.org/10.1007/978-3-642-15286-3_16

    2010 doi

  26. [34]

    Yiyou Sun, Yifei Ming, Xiaojin Zhu, and Yixuan Li. 2022. Out-of-distribution detection with deep nearest neighbors. In ICML. PMLR, 20827–20840. https: //proceedings.mlr.press/v162/sun22d.html

    2022

  27. [35]

    George Valkanas, Theodoros Lappas, and Dimitrios Gunopulos. 2017. Mining competitors from large unstructured datasets. TKDE 29, 9 (2017), 1971–1984. https://doi.org/10.1109/TKDE.2017.2705101

    2017

  28. [36]

    Jingdong Wang, Naiyan Wang, You Jia, Jian Li, Gang Zeng, Hongbin Zha, and Xian-Sheng Hua. 2013. Trinary-projection trees for approximate nearest neighbor search. TPAMI 36, 2 (2013), 388–403. https://doi.org/10.1109/TPAMI.2013.125

    2013 doi

  29. [37]

    Mengzhao Wang, Xiaoliang Xu, Qiang Yue, and Yuxiang Wang. 2021. A com- prehensive survey and experimental comparison of graph-based approximate nearest neighbor search. VLDB 14, 11 (2021), 1964–1978. https://doi.org/10. 14778/3476249.3476255

    2021

  30. [38]

    Xu Wang, Sen Wang, Xingxing Liang, Dawei Zhao, Jincai Huang, Xin Xu, Bin Dai, and Qiguang Miao. 2022. Deep reinforcement learning: A survey. IEEE Transactions on Neural Networks and Learning Systems 35, 4 (2022), 5064–5078. https://doi.org/10.1109/TNNLS.2022.3207346

    2022

  31. [39]

    Xiang Wu, Ruiqi Guo, Ananda Theertha Suresh, Sanjiv Kumar, Daniel N Holtmann-Rice, David Simcha, and Felix Yu. 2017. Multiscale quantization for fast similarity search. In NeurIPS. Curran Associates Inc., Red Hook, NY, USA, 5745–5755. https://proceedings.neurips.cc/paper_files...

    2017

  32. [40]

    Zhisheng Xiao, Qing Yan, and Yali Amit. 2020. Likelihood regret: An out-of- distribution detection score for variational auto-encoder. In NeuIPS. Curran As- sociates Inc., Red Hook, NY, USA, 20685–20696. https://proceedings.neurips.cc/ paper_files/paper/2020/file/eddea82ad2755...

    2020

  33. [41]

    Binhui Xie, Longhui Yuan, Shuang Li, Chi Harold Liu, Xinjing Cheng, and Guoren Wang. 2022. Active learning for domain adaptation: An energy-based approach. In AAAI. 8708–8716. https://doi.org/10.1609/aaai.v36i8.20850

    2022 doi

  34. [42]

    Jingkang Yang, Kaiyang Zhou, Yixuan Li, and Ziwei Liu. 2024. Generalized out-of-distribution detection: A survey. IJCV 132, 12 (2024), 5635–5662. https: //doi.org/10.1007/s11263-024-02117-4

    2024 doi

  35. [43]

    Tiannuo Yang, Wen Hu, Wangqi Peng, Yusen Li, Jianguo Li, Gang Wang, and Xiaoguang Liu. 2024. VDTuner: Automated Performance Tuning for Vector Data Management Systems. In ICDE. IEEE, 4357–4369. https://doi.org/10.1109/ ICDE60146.2024.00332

    2024

  36. [44]

    Yazhou Yang and Marco Loog. 2022. To actively initialize active learning. Pattern Recognition 131, C (2022), 108836. https://doi.org/10.1016/j.patcog.2022.108836

    2022

  37. [45]

    Ji Zhang, Yu Liu, Ke Zhou, Guoliang Li, Zhili Xiao, Bin Cheng, Jiashu Xing, Yangtao Wang, Tianheng Cheng, Li Liu, et al. 2019. An end-to-end automatic cloud database tuning system using deep reinforcement learning. In SIGMOD. Association for Computing Machinery, New York, NY, ...

    2019

  38. [46]

    Guang Zhao, Edward Dougherty, Byung-Jun Yoon, Francis Alexander, and Xi- aoning Qian. 2021. Efficient active learning for Gaussian process classifica- tion by error reduction. In NeuIPS. Curran Associates Inc., Red Hook, NY, USA, 9734–9746. https://proceedings.neurips.cc/paper...

    2021

  39. [47]

    Xi Zhao, Yao Tian, Kai Huang, Bolong Zheng, and Xiaofang Zhou. 2023. Towards efficient index construction and approximate nearest neighbor search in high- dimensional spaces. VLDB 16, 8 (2023), 1979–1991. https://doi.org/10.14778/ 3594512.3594527

    2023

  40. [48]

    Yuxin Zheng, Qi Guo, Anthony KH Tung, and Sai Wu. 2016. Lazylsh: Approximate nearest neighbor search for multiple distance functions with a single index. In SIGMOD. ACM, New York, NY, USA, 2023–2037. https://doi.org/10.1145/2882903. 2882930

    2016 doi

  41. [49]

    Wenyang Zhou, Yuzhi Jiang, Yingfan Liu, Xiaotian Qiao, Hui Zhang, Hui Li, and Jiangtao Cui. 2024. Auto-Tuning the Construction Parameters of Hierarchical Navigable Small World Graphs. https://ssrn.com/abstract=4925468 Received 20 February 2007; revised 12 March 2009; accepted ...

    2024

  42. [2020]

    Deep active learning for biased datasets via fisher kernel self-supervision. In CVPR. 9041–9049. https://doi.org/10.1109/cvpr42600.2020.00906

    2020

  43. [2023]

    Co-design hardware and algorithm for vector search. In SC. IEEE, 1–15. https://doi.org/10.1145/3581784.3607045

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.