Pith. sign in

REVIEW 1 major objections 5 minor 52 references

QASP: Query-Adaptive Robust Vector Search Policy

T0 review · 1 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

Pith's one-line read This paper claims that a single upfront regression predicts each query's full recall-progression curve, from which a target-agnostic probe policy meets any recall goal with materially less data access and much lower per-query variance than

desk verdict The experiments are worth your time; the theory section is not, as written. read the letter →

arxiv 2607.29606 v1 pith:ORDKRURU submitted 2026-07-31 cs.IR cs.LG

classification cs.IRcs.LG
keywords query-adaptivevectorsearchrecallprogressioncurveIVFindexingapproximatenearestneighborsupervisedregressionscale-invariantfeaturesqueryvariability-awareevaluationdataaccessreduction
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

Vector search systems normally probe a fixed number of clusters for every query, so easy queries over-read and hard queries under-read, and average-recall evaluations hide this inconsistency. QASP attempts to establish that a single supervised regression, run once before search, predicts each query's complete recall progression curve — recall as a function of clusters probed — so the right number of probes can be read off for any recall target. The paper argues this works with normalized, scale-invariant features that transfer across datasets and index configurations, and proves that the training sample size need not grow with the dataset, that QASP's loss is within a vanishing margin of the best fixed policy, and that savings over fixed probing grow exponentially in intrinsic dimensionality. If true, production vector databases can serve every query at its own difficulty level, cutting data access by roughly 80% at 99% recall while making recall consistency across queries much higher.

What carries the argument

The central object is the recall predictor f(q,l), which maps a query and a probe count to predicted cumulative recall, and the induced policy g[f](q,r) = min{l : f(q,l) ≥ r}. Asking the model to predict the entire recall curve — normalized into [0,1] — rather than a single probe count means one trained model serves every recall target and any number of clusters; this is recall-based regression. The features are scale-invariant ratios: cluster rank expressed as 1−exp(−l/L), query-to-centroid distances normalized by the nearest centroid, cumulative cluster size as a fraction of the dataset, and a local-intrinsic-dimensionality term capturing query difficulty. The theory uses pseudo-dimension

What would settle it

Construct or find a partition index and a query whose nearest neighbor lies in an early cluster while the remaining k−1 true neighbors sit in far clusters; if the observed number of probes needed to reach full recall exceeds (4(r(q)+D_I)/σ_I)^∆, the Lemma 12 bound — and the exponential-savings corollary — is false. A second falsifier: train QASP on a workload where all queries need exactly the same probe count; if the induced policy returns anything other than that constant, the fixed-probe realizability assumption in Theorem 11 fails.

Watch

Extended reading notes

Core claim

QASP's central claim: a query's behavior in a partitioned index is summarized by a recall progression curve, and one supervised model with scale-invariant features predicts that curve well enough to derive a probe policy for any recall target. Predicting the whole curve, not a single probe count, makes the policy target- and index-size-agnostic: the induced policy is the smallest l where predicted recall crosses the target. The paper proves a finite sample suffices independent of dataset size, that fixed probe policies carry irreducible loss when query difficulty varies, and that QASP's data access is strictly lower, with savings exponential in intrinsic dimensionality. It reports 57.7% lowe

Load-bearing premise

The exponential-savings claim rests on Lemma 12, which assumes the probes needed for full recall are controlled by the clusters inside the ball around the query's single nearest neighbor; the proof only bounds the cluster containing one of the k neighbors, so the k>1 case is unproved, and Theorem 11 separately assumes every fixed-probe policy is realizable by the learned model class.

Editorial extensions

If this is right

  • Systems can serve each query with its own probe count from a single model call, so easy queries stop early and hard queries search deeper, instead of paying the fixed cost of the hardest query on every query.
  • The finite-sample guarantee means a small labeled query set suffices; adding more vectors to the index does not force more training data.
  • At high recall targets, especially 99%, the per-query policy reduces data access by roughly 80% on hierarchical indexes compared with the best fixed policy.
  • Because the model predicts the whole curve, the same deployment covers multiple recall targets and multiple index configurations with zero-shot transfer; few-shot fine-tuning on 1% of target queries closes most residual gaps.
  • The per-partition recall estimates can be used as live hypotheses during search, enabling early or late termination rules that improve consistency without additional model inference.

Reading between the lines

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

  • If the curve predictions are accurate enough, the same mechanism could drive progressive disclosure in interactive retrieval — showing the first clusters' results immediately and deepening only as the user scrolls — which the paper does not discuss.
  • The local-intrinsic-dimensionality feature ties query difficulty to known concentration-of-distance phenomena; one testable extension would be to use QASP's predicted curves as a per-query difficulty label to calibrate when fixed-probe rules or graph-based search are preferable.
  • The 80% savings figure is relative to oracle fixed policies tuned on the training distribution; under workload drift the advantage could be larger if QASP retrains online, or smaller if the drift moves queries outside the learned feature ranges — the paper does not evaluate drift over time.
  • A repair of the Lemma 12 bound for k>1 would directly extend the exponential-savings theorem from 1-NN to k-NN recall, a natural next step given that real workloads query with k > 1.
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

1 major / 5 minor

Summary. The paper proposes QASP, a supervised-learning method for IVF-style vector search that predicts the full per-query recall-versus-probes curve from scale-invariant features in a single proactive inference. A probe policy is then derived for any recall target, and a lightweight reactive complement adjusts the probe depth from observed discovery rates. The paper claims theoretical guarantees: a finite training sample independent of dataset size and dimensionality, a near-optimality gap relative to any fixed probe policy, and data-access savings over fixed probing that grow exponentially in the intrinsic dimensionality. Experiments on seven datasets report lower recall variance, lower deviation from target, higher query satisfaction, successful transfer across datasets and index configurations, and an 80% data-access reduction on hierarchical indices at 99% recall.

Significance. The empirical contribution is substantial and well specified: the QVE evaluation metrics address a real limitation of average-recall reporting, the comparison to oracle fixed-policy baselines is informative, the transfer and hierarchical-scaling results are interesting, and the reported inference latencies (0.06--0.26 ms/batch) make the approach plausibly deployable. If the theoretical claims could be made correct, they would be a valuable addition. However, the current proof of the headline exponential-savings result has a load-bearing gap: Lemma 12 bounds only the probes needed to find the single nearest neighbor, whereas the experiments and the dominance claim concern recall@k for k=100. The realizability premise in Theorem 11 is also asserted without proof. These issues do not invalidate the experiments, but they do invalidate the paper's central theoretical contribution as written.

major comments (1)
  1. [§IV-C, Corollary 15] The 'matched recall (ε=δ)' condition conflates two different failure notions. δ is defined as the upper quantile of the single-nearest-neighbor distance r(q), not as a recall failure probability, while ε is the probability that QASP under-probes relative to ℓ*(q,r*). Even if Lemma 12 were fixed for k-NN, the matching ε=δ would require a separate argument. As written, the dominance condition (11) does not establish the claimed interpretation of the vertical gap in Figure 2 as data-access savings at matched recall.
minor comments (5)
  1. [§IV-B] The sentence 'Proposition 10 allows us to complement this result' refers to Lemma 10, not Proposition 10; this cross-reference should be corrected.
  2. [Abstract and §VI-H] The abstract's '80% less data access' should explicitly state that this is for hierarchical indices at 99% recall (§VI-H), since flat-index results in Table I show much smaller data-access differences.
  3. [Table III] The zero-shot SIFT1M→GIST1M transfer obtains r̄=26.80, which is far from target; the claim that QASP 'often generalizes well zero-shot' is therefore too broad. Please report the distribution of transfer failures or qualify the claim.
  4. [§VII-C] The SHAP feature-importance analysis is based only on QASP-GBDT; stating this in the caption would help readers interpret the ablation results in Table VI.
  5. [General] No code or dataset release is mentioned. Given the number of tables and the specificity of the proposed evaluation metrics, a reproducibility statement would strengthen the paper.

Circularity Check

1 steps flagged · score 5.0 of 10

One theoretical guarantee reduces to label-as-feature, but empirical recall predictions are held-out and self-citations are background.

  1. self definitional [Section IV-B, Remark after Theorem 11 (extension to eq. (4))]
    "The formal proof goes by augmenting the feature set F(q) with the additional query features {R_I(D)(q,l)}_l=1^L, which enables computing the contribution of q to the loss from F(q) while maintaining the size bound |F(q)|=O(L)."

    This is the only offered justification for extending the finite-sample/near-optimal-loss guarantee to the recall-regression objective that QASP actually trains on (eq. (4)). By adding the ground-truth recall values R(q,l) to the feature set, the policy class contains a predictor that outputs the target label directly, making the induced probe policy equal to the optimal ℓ*(q,r) by construction. The claimed guarantee then follows from reading the answer off the input rather than from learning query features. The experimental recall predictions are held-out and unaffected, so the circularity is confined to this theoretical assertion.

full rationale

Most of the derivation is not circular. The empirical core is self-contained: recall features and labels are collected once, an 80/20 split is used, and test queries are sampled from the remaining query set with features computed on-the-fly during search, so the reported recall predictions are genuine out-of-sample predictions. The dominance condition (Cor. 15) compares QASP's predictor overhead against the fixed-policy tail waste; it does not bake in the experimental constants. Self-citations [9,10,16,38] are background or baseline references (IVF indexing, vector approximation, SQUASH) and are not load-bearing: no uniqueness theorem or ansatz is imported from the authors' own prior work. The principal circularity I find is the Remark after Theorem 11: to carry the theoretical guarantee over to the recall-regression loss used in training, the proof sketch augments the model's feature set with the very recall labels being predicted, so the loss bound is a tautology for that objective. I do not treat the unproven premise that fixed-probe policies are realizable in G_QASP, or the Lemma 12/Prop. 13 k-NN gap, as circularity: those are non-circular but correctness/rigor concerns. Because the empirical claims rest on held-out evaluation, the score is moderate rather than high.

Assumptions & free parameters 3 free parameters · 7 assumptions · 0 invented entities

The formal guarantees rest on standard PAC/packing results plus several domain assumptions: approximate ground truth as labels, query-workload stationarity, realizability of fixed policies within G_QASP, and the geometric model of query difficulty. The empirical method additionally relies on hand-set reactive constants and validation-tuned model hyperparameters. No new physical or latent entities are introduced.

free parameters (3)
  • QASP model hyperparameters = DL: 18-d hidden, 3 layers, lr 5e-3; GBDT: 100×depth-3 trees; LITE: Lasso 1e-5, degree 3
    Selected via held-out validation set (§VI-A3); these values are tuned on the datasets rather than derived.
  • Reactive complement constants = α=0.3, threshold=0.25, Lδ=3
    Hand-set in §VI-G and Algorithm 1; no sensitivity analysis is provided.
  • QVE satisfaction margin = 5% relaxation of r*
    §VI-B subtracts 5% from the recall target to define query satisfaction; this changes S% results and is chosen by hand.
assumptions (7)
  • domain assumption Ground-truth nearest-neighbor labels obtained from configurations that approximate exhaustive search are exact enough for training targets
    §II-C1 states training signals come from search configurations that 'approximate exhaustive retrieval'; recall labels are therefore approximate, biasing the fitted recall curve.
  • ad hoc to paper The QASP policy class contains all fixed-probe policies g^(l)
    Invoked in the proof of Theorem 11 (§IV-B) to compare against every fixed policy; not proven for the specific architectures and feature set.
  • standard math Pseudo-dimension bound applies to the class of policies induced by thresholding recall predictors (eq. (3))
    Theorem 9 is invoked via Lemma 10; the composition with min{·} is asserted finite without a full argument.
  • domain assumption Doubling dimension Δ of the dataset union centroids is the right measure of intrinsic dimensionality and is finite
    Used in Lemma 12 and Prop 13 (§IV-C); real high-dimensional data has large effective doubling dimension, weakening the packing bound.
  • domain assumption The target-query distribution Q equals the training workload distribution in the sample-complexity theorems
    Theorems 5-11 assume q ~ Q; experiments sample training queries from the dataset, so the assumption holds only if the serving workload matches the training sample.
  • ad hoc to paper Fixed-ℓ policy with ℓ ≥ ℓ_fix(δ) is the relevant cost of fixed probing
    Prop 13 gives a sufficient condition; the dominance condition treats this upper bound as 'the cost every query pays,' which is not the minimum fixed probe count.
  • ad hoc to paper The learned recall predictor's failure rate ε can be matched to the fixed policy's recall failure δ (ε=δ)
    Cor 15 sets ε=δ to compare failure-matched policies; no argument shows this matching is achievable without additional tuning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QASP: Query-Adaptive Robust Vector Search Policy." pith.science (2026). https://pith.science/paper/ORDKRURU

@misc{pith2026260729606,
  author       = {Pith},
  title        = {Pith review of: QASP: Query-Adaptive Robust Vector Search Policy},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ORDKRURU}},
  note         = {Machine review of arXiv:2607.29606}
}
read the original abstract

A fundamental challenge of vector search is achieving consistently high recall while minimizing computational costs. Fixed search parameters cause significant performance variance across queries, and conventional evaluation on average recall masks these per-query disparities. We introduce QASP (Query-Adaptive robust vector Search Policy), which predicts the complete recall progression curve per query via a single upfront supervised regression, from which a search policy is derived for any recall target; this avoids iterative model invocations during search or separate predictors per target. By predicting normalized recall values with scale-invariant features and pre-search inference, QASP generalizes across recall targets, index configurations, and datasets. Its fine-grained progress predictions further enable a lightweight reactive complement that adjusts search depth based on predicted-versus-observed deviations without additional inference. We prove that QASP requires a finite training sample independent of dataset size and dimensionality, that its loss exceeds the irreducible lower bound of any fixed policy by a vanishing margin, and that its data access savings over fixed probing grow exponentially in intrinsic dimensionality. Experimentally, QASP achieves significantly lower recall variance and deviation from target, higher query satisfaction rate, and scales to large data and hierarchical indices without retraining, achieving 99% recall with 80% less data access.

Figures

Figures reproduced from arXiv: 2607.29606 by the authors.

Figure 1
Figure 1. Comparing fixed and adaptive policies: Query sat [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Expected probes vs. recall target for fixed probing [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. Search time (ms) vs. recall (in-memory) and data [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: A% for SIFT1B and Deep1B subsets (10M–100M) on a two-level hierarchical index at 99% recall target. QASP is trained on a 10M subset and applied at inference without retraining. QASP consistently yields ≥80% reduction in A%. (a) MNIST (b) GIST (c) SIFT1M (d) GLOVE￾200 (…
Figure 5
Figure 5. Figure 5: Training loss (MSE) over epochs B. Prediction Error Analysis [PITH_FULL_IMAGE:figures/full_fig_p012_5.png]
Figure 6
Figure 6. Figure 6: Prediction error analysis showing actual vs. predicted [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 1 linked inside Pith

  1. [1]

    IVF-PQ index,

    LanceDB, “IVF-PQ index,” LanceDB Documentation, https://lancedb. github.io/lancedb/concepts/index ivfpq/#query-the-index

  2. [2]

    Guidelines to choose an index,

    Facebook AI Research, “Guidelines to choose an index,” FAISS Wiki, https://github.com/facebookresearch/faiss/wiki/ Guidelines-to-choose-an-index

  3. [3]

    Choose the k-NN algorithm for your billion- scale use case with OpenSearch,

    OpenSearch, “Choose the k-NN algorithm for your billion- scale use case with OpenSearch,” Amazon OpenSearch Service Blog, 2022, https://aws.amazon.com/blogs/big-data/ choose-the-k-nn-algorithm-for-your-billion-scale-use-case-with-opensearch/

  4. [4]

    Improving approximate nearest neighbor search through learned adaptive early termination,

    C. Li, M. Zhang, D. G. Andersen, and Y . He, “Improving approximate nearest neighbor search through learned adaptive early termination,” in Proceedings of the 2020 ACM SIGMOD International Conference on Management of Data, 2020, pp. 2539–2554

  5. [5]

    Quake: Adaptive indexing for vector search,

    J. Mohoney, D. Sarda, M. Tang, S. R. Chowdhury, A. Pacaci, I. F. Ilyas, T. Rekatsinas, and S. Venkataraman, “Quake: Adaptive indexing for vector search,” 2025

  6. [6]

    Vista: Vector indexing and search for large-scale imbalanced datasets,

    Y . Fu, C. Chen, Y . Chen, W.-F. Wong, and B. He, “Vista: Vector indexing and search for large-scale imbalanced datasets,” in2025 IEEE 41st International Conference on Data Engineering (ICDE), 2025, pp. 543– 556

  7. [7]

    Boosting accuracy and efficiency for vector retrieval with local scaling graph,

    H. Wang, W. Wu, C. Luo, A. Bian, C. Meng, Y . Wu, and J. Sun, “Boosting accuracy and efficiency for vector retrieval with local scaling graph,” in2025 IEEE 41st International Conference on Data Engineer- ing (ICDE), 2025, pp. 336–348

  8. [8]

    Scaling laws for nearest neighbor search,

    P. Sun, F. Chern, Y . Akhremtsev, R. Guo, D. Simcha, and S. Kumar, “Scaling laws for nearest neighbor search,” inThe 1st Workshop on Vector Databases, 2025

Show all 52 references
  1. [9]

    Ap- proximate nearest neighbor searching in multimedia databases,

    H. Ferhatosmanoglu, E. Tuncel, D. Agrawal, and A. El Abbadi, “Ap- proximate nearest neighbor searching in multimedia databases,” inPro- ceedings 17th International Conference on Data Engineering (ICDE). IEEE, 2001, pp. 503–511

  2. [10]

    Vq-index: An index struc- ture for similarity searching in multimedia databases,

    E. Tuncel, H. Ferhatosmanoglu, and K. Rose, “Vq-index: An index struc- ture for similarity searching in multimedia databases,” inProceedings of the 10th ACM International Conference on Multimedia, 2002, pp. 543–552

  3. [11]

    Product quantization for nearest neighbor search,

    H. Jegou, M. Douze, and C. Schmid, “Product quantization for nearest neighbor search,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 33, no. 1, pp. 117–128, 2010

  4. [12]

    Spann: Highly-efficient billion-scale approximate nearest neighbor search,

    Q. Chen, B. Wang, Y . Guo, Y . Zheng, Y . Li, X. Chang, E. Y . Sun, J. Zhang, X. Li, and X. Zhang, “Spann: Highly-efficient billion-scale approximate nearest neighbor search,” inAdvances in Neural Informa- tion Processing Systems, vol. 34, 2021, pp. 10 337–10 349

  5. [13]

    Parlayann: Scalable and deterministic parallel graph-based approximate nearest neighbor search algorithms,

    M. D. Manohar, Z. Shen, G. Blelloch, L. Dhulipala, Y . Gu, H. V . Simhadri, and Y . Sun, “Parlayann: Scalable and deterministic parallel graph-based approximate nearest neighbor search algorithms,” inPro- ceedings of the 29th ACM SIGPLAN Annual Symposium on Principles and Prac...

  6. [14]

    Vexless: A serverless vector data management system using cloud functions,

    Y . Su, Y . Sun, M. Zhang, and J. Wang, “Vexless: A serverless vector data management system using cloud functions,”Proceedings of the ACM on Management of Data, vol. 2, no. 3, pp. 1–26, 2024

  7. [15]

    Diskann: Fast accurate billion-point nearest neighbor search on a single node,

    S. Jayaram Subramanya, F. Devvrit, H. V . Simhadri, R. Krishnawamy, and R. Kadekodi, “Diskann: Fast accurate billion-point nearest neighbor search on a single node,”Advances in Neural Information Processing Systems, vol. 32, 2019

  8. [16]

    SQUASH: Serverless and dis- tributed quantization-based attributed vector similarity search,

    J. Oakley and H. Ferhatosmanoglu, “SQUASH: Serverless and dis- tributed quantization-based attributed vector similarity search,”arXiv preprint arXiv:2502.01528, 2025

  9. [17]

    On the consistency of ordinal regression methods,

    F. Pedregosa, F. Bach, and A. Gramfort, “On the consistency of ordinal regression methods,”Journal of Machine Learning Research, vol. 18, no. 55, pp. 1–35, 2017

  10. [18]

    On embeddings for numer- ical features in tabular deep learning,

    Y . Gorishniy, I. Rubachev, and A. Babenko, “On embeddings for numer- ical features in tabular deep learning,”Advances in Neural Information Processing Systems, vol. 35, pp. 24 991–25 004, 2022

  11. [19]

    Tabular data: Deep learning is not all you need,

    R. Shwartz-Ziv and A. Armon, “Tabular data: Deep learning is not all you need,”Information Fusion, vol. 81, pp. 84–90, 2022

  12. [20]

    Darth: Declara- tive recall through early termination for approximate nearest neighbor search,

    M. Chatzakis, Y . Papakonstantinou, and T. Palpanas, “Darth: Declara- tive recall through early termination for approximate nearest neighbor search,”Proc. ACM Manag. Data, vol. 3, no. 4, Sep. 2025

  13. [21]

    Revisiting batch normal- ization for practical domain adaptation,

    Y . Li, N. Wang, J. Shi, J. Liu, and X. Hou, “Revisiting batch normal- ization for practical domain adaptation,” 2016

  14. [22]

    The role of local dimensionality mea- sures in benchmarking nearest neighbor search,

    M. Aum ¨uller and M. Ceccarello, “The role of local dimensionality mea- sures in benchmarking nearest neighbor search,”Information Systems, vol. 101, p. 101807, 2021

  15. [23]

    On self-organizing sequential search heuristics,

    R. Rivest, “On self-organizing sequential search heuristics,”Communi- cations of the ACM, vol. 19, no. 2, pp. 63–67, 1976

  16. [24]

    Convergence of stochastic processes,

    D. Pollard, “Convergence of stochastic processes,”Springer Series in Statistics, 1984

  17. [25]

    Institute of Mathematical Statistics, 1990

    ——,Empirical Processes: Theory and Applications. Institute of Mathematical Statistics, 1990

  18. [26]

    Learn- ability and the vapnik-chervonenkis dimension,

    A. Blumer, A. Ehrenfeucht, D. Haussler, and M. K. Warmuth, “Learn- ability and the vapnik-chervonenkis dimension,”Journal of the ACM (JACM), vol. 36, no. 4, pp. 929–965, 1989

  19. [27]

    A pac approach to application-specific algorithm selection,

    R. Gupta and T. Roughgarden, “A pac approach to application-specific algorithm selection,”SIAM Journal on Computing, vol. 46, no. 3, pp. 992–1017, 2017

  20. [28]

    Data-driven algorithm design,

    M.-F. Balcan, “Data-driven algorithm design,”Beyond Worst Case Analysis of Algorithms (Tim Roughgarden, ed.), 2020

  21. [29]

    On the uniform convergence of relative frequencies of events to their probabilities,

    V . N. Vapnik and A. Y . Chervonenkis, “On the uniform convergence of relative frequencies of events to their probabilities,” inMeasures of complexity: festschrift for alexey chervonenkis. Springer, 2015, pp. 11–30

  22. [30]

    Anthony and P

    M. Anthony and P. L. Bartlett,Neural network learning: Theoretical foundations. Cambridge University Press, 2009

  23. [31]

    Vapnik-chervonenkis dimension of neural nets,

    P. L. Bartlett and W. Maass, “Vapnik-chervonenkis dimension of neural nets,”The handbook of brain theory and neural networks, pp. 1188– 1192, 2003

  24. [32]

    Bounding the vapnik-chervonenkis dimension of concept classes parameterized by real numbers,

    P. W. Goldberg and M. R. Jerrum, “Bounding the vapnik-chervonenkis dimension of concept classes parameterized by real numbers,”Machine Learning, vol. 18, no. 2-3, pp. 131–148, 1995

  25. [33]

    Polynomial bounds for vc dimension of sigmoidal and general pfaffian neural networks,

    M. Karpinski and A. Macintyre, “Polynomial bounds for vc dimension of sigmoidal and general pfaffian neural networks,”Journal of Computer and System Sciences, vol. 54, no. 1, pp. 169–176, 1997

  26. [34]

    Cover trees for nearest neighbor,

    A. Beygelzimer, S. M. Kakade, and J. Langford, “Cover trees for nearest neighbor,” inProceedings of the 23rd International Conference on Machine Learning (ICML). ACM, 2006, pp. 97–104

  27. [35]

    Multidimensional binary search trees used for associative searching,

    J. L. Bentley, “Multidimensional binary search trees used for associative searching,”Communications of the ACM, vol. 18, no. 9, pp. 509–517, 1975

  28. [36]

    Five balltree construction algorithms,

    S. M. Omohundro, “Five balltree construction algorithms,”International Computer Science Institute Berkeley, 1989

  29. [37]

    Approximate similarity retrieval with m-trees,

    P. Zezula, P. Savino, G. Amato, and F. Rabitti, “Approximate similarity retrieval with m-trees,”The VLDB Journal, vol. 7, no. 4, pp. 275–293, 1998

  30. [38]

    Vector- approximation based indexing for non-uniform high dimensional data sets,

    H. Ferhatosmanoglu, E. Tuncel, D. Agrawal, and A. El Abbadi, “Vector- approximation based indexing for non-uniform high dimensional data sets,”Proceedings of the 9th International Conference on Information and Knowledge Management (CIKM), pp. 202–209, 2000

  31. [39]

    Configure index parameters,

    Milvus, “Configure index parameters,” Milvus Documentation v2.3.0, 2023, https://milvus.io/docs/v2.3.0/index.md

  32. [40]

    Autotune example,

    Facebook AI Research, “Autotune example,” FAISS Source Code, https: //github.com/facebookresearch/faiss/blob/main/tutorial/cpp/4-GPU.cpp

  33. [41]

    Efficient and robust approxi- mate nearest neighbor search using hierarchical navigable small world graphs,

    Y . A. Malkov and D. A. Yashunin, “Efficient and robust approxi- mate nearest neighbor search using hierarchical navigable small world graphs,”IEEE Transactions on Pattern Analysis and Machine Intelli- gence, vol. 42, no. 4, pp. 824–836, 2018

  34. [42]

    Distribution-aware exploration for adaptive HNSW search,

    C. Zhang and R. J. Miller, “Distribution-aware exploration for adaptive HNSW search,”Proceedings of the ACM on Management of Data, vol. 4, no. 1, 2026

  35. [43]

    Learned probing cardinality estimation for high-dimensional approximate NN search,

    B. Zheng, Z. Yue, Q. Hu, X. Yi, X. Luan, C. Xie, X. Zhou, and C. S. Jensen, “Learned probing cardinality estimation for high-dimensional approximate NN search,” in2023 IEEE 39th International Conference on Data Engineering (ICDE), 2023, pp. 3209–3221

  36. [44]

    Distinctive image features from scale-invariant keypoints,

    D. G. Lowe, “Distinctive image features from scale-invariant keypoints,” Int. J. Comput. Vision, vol. 60, no. 2, p. 91–110, Nov. 2004

  37. [45]

    Gradient-based learning applied to document recognition,

    Y . LeCun, L. Bottou, Y . Bengio, and P. Haffner, “Gradient-based learning applied to document recognition,”Proceedings of the IEEE, vol. 86, no. 11, pp. 2278–2324, 2002

  38. [46]

    Modeling the shape of the scene: A holistic representation of the spatial envelope,

    A. Oliva and A. Torralba, “Modeling the shape of the scene: A holistic representation of the spatial envelope,”International journal of computer vision, vol. 42, pp. 145–175, 2001

  39. [47]

    Efficient indexing of billion-scale datasets of deep descriptors,

    A. Babenko and V . Lempitsky, “Efficient indexing of billion-scale datasets of deep descriptors,” inProceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2016, pp. 2055–2063

  40. [48]

    Glove: Global vectors for word representation,

    J. Pennington, R. Socher, and C. D. Manning, “Glove: Global vectors for word representation,” inEmpirical Methods in Natural Language Processing (EMNLP), 2014, pp. 1532–1543

  41. [49]

    Microsoft coco: Common objects in context,

    T.-Y . Lin, M. Maire, S. Belongie, L. Bourdev, R. Girshick, J. Hays, P. Perona, D. Ramanan, C. L. Zitnick, and P. Doll ´ar, “Microsoft coco: Common objects in context,” 2015

  42. [50]

    Deep learning using rectified linear units (relu),

    A. F. Agarap, “Deep learning using rectified linear units (relu),” 2019

  43. [51]

    k-means++: the advantages of careful seeding,

    D. Arthur and S. Vassilvitskii, “k-means++: the advantages of careful seeding,” inProceedings of the Eighteenth Annual ACM-SIAM Sym- posium on Discrete Algorithms, ser. SODA ’07. USA: Society for Industrial and Applied Mathematics, 2007, p. 1027–1035

  44. [52]

    A unified approach to interpreting model predictions,

    S. Lundberg and S.-I. Lee, “A unified approach to interpreting model predictions,” 2017. IX. AI ASSISTANCESTATEMENT We used Claude (Anthropic, Opus 4.6) to assist with imple- mentation of the PCE-Net baseline directly from the original paper descriptions as the original implem...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.