Pith. sign in

REVIEW 2 major objections 4 minor 51 references

A single random EM initialization can recover the true value vector in one step for a decoupled query–value multiple instance model.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 18:03 UTC pith:OQ6WEC72

load-bearing objection The concentration theorem and EM-kappa framing are worth a serious look, but the abstract's O(1)-step convergence claim is not supported by the proofs—the missing margin argument is a real overreach, not a technical nit. the 2 major comments →

arxiv 2607.17404 v1 pith:OQ6WEC72 submitted 2026-07-19 math.ST stat.TH

EM-based iterations for multiple instance learning on a query-value model

Ethan Levien This is my paper
classification math.ST stat.TH MSC 68T0768Q3260F05
keywords multiple instance learningmultiple instance regressionEM algorithmextremal poolingquery-value modelattentioncovariance concentrationextreme value statistics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper studies multiple instance regression, where each bag is a set of feature vectors and the bag-level label is set by the single 'active' instance: the instance that maximizes its projection onto a query direction, with the label equal to its projection onto a separate value direction. Learning both directions from bag labels is a stripped-down version of what attention-based MIL architectures do, and it has resisted rigorous analysis. The paper derives a family of hardened EM iterations, EM_kappa, that generalize the classical EM-DD method, and then proves concentration theorems for the value and query updates under isotropic Gaussian instances. The central result is that a value vector estimated from a completely random assignment is concentrated around a positive multiple of the true value vector, so once the number of bags is roughly d n^2 (ln n)^6, a single random initialization produces the correct assignment after one step of the gEM1 iteration. The proof connects covariance concentration with extremal-statistics moments; the results cover the first step of the iteration, not the full trajectory.

Core claim

The paper's centerpiece is Theorem 4.6, which shows that for i.i.d. Gaussian instances, the value update v-hat(ξ) = Σ-hat(ξ,ξ)^{-1} Σ-hat(ξ,ξ*) v* concentrates around A-bar(f)^{-1} B-bar(f) v* with error O(ε), provided m ≳ d (ln n)^6 / ε^2 and the match fraction f ∈ (0,1) is fixed. When the query and value directions are aligned (q* ∝ v*), the limiting vector is a scalar multiple φ_n(f) v*, where φ_n(f) is an explicit rational function of the extremal moments s and w. For a completely random assignment, f = 1/n and φ_n(1/n) = 1/n, so the signal is small but positive; choosing ε ~ 1/n gives the sample complexity m ≳ d n^2 (ln n)^6. Because gEM1's assignment step depends only on the direction

What carries the argument

The load-bearing object is the conditional covariance structure of the assigned instances. Lemma 4.5 gives the exact averages A-bar(f) = Σ + ρ_n(f) Σ_{1,1}^{-1} Σ_1 Σ_1^T and B-bar(f) = fΣ + ρ_n(f) Σ_{1,1}^{-1} Σ_1 Σ_1^T, where ρ_n(f) = f(s-1) + (1-f)(w-1), with s the second moment of the extremal instance and w the second moment of a non-extremal instance. The identities v = (1-s)/(n-1) and w = v+1 close the system, and the Sherman–Morrison formula turns the rank-one perturbation into the scalar φ_n(f) that controls the value-map scaling. Concentration of empirical covariance matrices (Lemma 4.2) is applied conditionally on the matched/unmatched split, giving the m ≳ d (ln n)^6 / ε^2 bound.

Load-bearing premise

The one-step recovery claim rests on the unproven assumption that an estimated value vector within O(1/n) of a positive multiple of the true value still selects the correct active instance in every bag; the paper proves concentration of the estimate, not stability of the argmax assignment under that noise.

What would settle it

Simulate the isotropic Gaussian model with m = C d n^2 (ln n)^6 and a single uniformly random initial assignment, then check whether the gEM1 update recovers the true assignment in all m bags; if the match probability decays or fails to approach 1 as n grows, the missing margin step is false. A sharper check is to measure the component of v-hat − v*/n orthogonal to v* and verify it is O(1/n) rather than only O(1/√m).

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • For aligned query and value directions, one uniformly random initial assignment suffices to recover the true value direction in a single gEM1 step once m ≳ d n^2 (ln n)^6 bags are available.
  • The sample-complexity scaling matches the uniform-selection mechanism up to logarithmic factors, showing that the extremal selection rule does not fundamentally change the difficulty of the random-initialization problem.
  • The linearized EM-DD algorithm is shown to be a variant of the κ=1 gEM1 iteration, connecting a classical MIL heuristic to the query–value modeling framework.
  • The finite-bag angle formula of Section 4.5 predicts that the value estimate's angle to the truth shrinks with f even when query and value are misaligned, a prediction verified in simulation.
  • The query update converges to q* for any fixed match fraction f > 1/n, so the approximate averaged-instance query estimator is consistent in the large-m limit.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Beyond the paper: since one step of gEM1 yields a match fraction f > 1/n, subsequent steps operate in a regime where the signal is even stronger; a finite-time analysis of the full iteration may show O(1)-step convergence under the same polynomial bag count, not just for the first step.
  • Beyond the paper: the 1/n signal floor appears independent of the selection mechanism, suggesting that any pooling rule with a unique active instance will require roughly m ~ d n^2 bags to resolve the direction from random initialization; this is a testable design principle for attention-based MIL.
  • Beyond the paper: the non-monotonic angle formula implies that the optimal κ schedule in the staged method should depend on the unknown angle θ_{v*,q*}; a concrete next experiment is to measure how the optimal switch time varies with that angle.
  • Beyond the paper: the one-step recovery claim depends on the unproven uniform margin property; one can test it directly by computing, under the m ≳ d n^2 (ln n)^6 regime, the probability that the estimated value vector's argmax assignments coincide with the true assignments across all bags.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes a query-value model for multiple instance regression in which each bag's label is determined by the active instance maximizing x·q and the value vector v. It derives a family of hard-assignment EM iterations (EM_kappa) that interpolate between residual-based and extremal assignment rules, and identifies gEM1, the kappa=1 member with q proportional to v, as a linearized variant of EM-DD. For Gaussian instances, the main theoretical result (Theorem 4.6) states that, conditionally on a fixed match fraction f between a random assignment and the true assignment, the value update \hat v(\xi) concentrates around \bar A(f)^{-1}\bar B(f)v^* with error O(epsilon) once m \gtrsim d(\ln n)^6/\epsilon^2. Section 4.3 combines the f=1/n case, where the scaling factor is phi_n(1/n)=1/n, with epsilon ~ 1/n, and claims that m \gtrsim d n^2(\ln n)^6 suffices for one step of gEM1 from a uniformly random initialization to recover the true value vector / assignment. The paper also proves a query-mean consistency theorem, presents numerical experiments on synthetic data, and studies kappa schedules.

Significance. If the concentration theorem is correct, it is a useful contribution to the theory of EM-type algorithms in MIL: it identifies an extremal-statistics mechanism that gives a O(1/n) signal from a completely random assignment and provides an explicit polynomial sample-complexity bound. The derivation of the EM_kappa family and the connection to EM-DD are original and clearly presented. The proof of the extremal moment identities (Lemma 4.3) is elegant, and the proof of Theorem 4.6 is detailed. The authors are also transparent that the analysis covers only the first step and that the full finite-time dynamics are left open. The main weakness is that the abstract's headline claim that the EM algorithm converges in O(1) steps is not justified by the stated results; the missing uniform-perturbation argument is load-bearing.

major comments (2)
  1. [§4.3, Eqs. (52)–(57)] The claim that m \gtrsim d n^2(\ln n)^6 guarantees one-step recovery from a random initialization is not supported by Theorem 4.6. The theorem only provides \|\hat v(\xi)-\phi_n(f)v^*\|=O(epsilon) with \phi_n(1/n)=1/n. Writing \hat v=v^*/n+e with \|e\|\le C/n (the epsilon ~ 1/n choice), the gEM1 update selects argmax_j x_j·\hat v. The score difference protecting the true active instance is (X_(n)-X_(n-1))/n, and for i.i.d. Gaussian instances the top-two order statistic gap is of order 1/sqrt(2 ln n). The perturbation (x_{\xi^*}-x_j)·e has typical size sqrt(2)C/n. The signal-to-noise ratio is therefore O(1/(C sqrt(ln n))), which tends to 0. Hence an O(1/n) Euclidean error does not preserve the argmax; the sentence 'any positive scalar multiple of v* is sufficient' conflates exact proportionality with noisy estimation. A valid argument would require either a relative-error bound with epsil
  2. [Abstract; §4.3 (last paragraph); §5] The paper's own §4.3 states that 'the statements above concern only the first step' and §5 says the results 'do not by themselves characterize the full finite-time dynamics.' Nevertheless, the abstract claims that a polynomial number of bags is enough for the EM algorithm to converge in O(1) steps. Even if the one-step concentration were strengthened as in the previous comment, this would not establish convergence of the iteration: one would need to show that after the first update the assignment is, with high probability, the true assignment, and that subsequent value-map updates keep the iteration in the correct basin. No such argument is given. The abstract and the Section 1.1 bullet summary should be weakened to a statement about one step from a random initialization producing a directionally informative value estimate, or the missing convergence argument must be supplied.
minor comments (4)
  1. [§2.5, Eq. (37)] The surrogate q_hat_a(\xi) is introduced as an approximation to the query MLE without a derivation. Theorem 4.8 justifies it only for a random assignment at fixed f; its behavior inside the EM iteration is not analyzed. This limitation should be flagged where the surrogate is first used.
  2. [Abstract/§1.1] The phrase 'single random initialization of the value vector' is imprecise: the algorithm initializes an assignment, not a value vector; the value estimate comes from the M-step. Please rephrase to avoid confusion.
  3. [§3] The numerical experiments do not reach the sample sizes required by Eq. (57). For n=50, d=15, Eq. (57) would require m ~ 10^10, whereas Figures 2–4 use m <= 5000. Thus the simulations neither confirm nor refute the asymptotic one-step claim; this disconnect should be acknowledged.
  4. [§4.3] Theorem 4.6 is stated for fixed f in (0,1), while Section 4.3 uses f=1/n, which tends to 0 with n. The proof's conditions appear to allow this pointwise, but the transition should be made explicit.

Circularity Check

0 steps flagged

No significant circularity: the main concentration result is derived from the stated generative model and external inequalities; the unsupported O(1)-step margin step is a proof gap, not a circular reduction.

full rationale

The paper's derivation chain is self-contained. The EM-kappa family follows from the soft query-value model (Eqs. 17-28), and the value-map concentration result (Theorem 4.6, Eq. 52) is proved from the Gaussian generative model in Eq. 38, the external matrix concentration lemma (Lemma 4.2, Vershynin), and standard extremal-statistics estimates (Lemmas 4.3-4.4). No fitted parameter is renamed as a prediction: the specialization f=1/n and epsilon ~ 1/n is an explicit choice, not an estimate from data, and gives m ~ d n^2 (ln n)^6 as stated. The use of the q-hat_a heuristic is not load-bearing for the central value-vector result, and Section 4.6 only proves its consistency under the same model; this is a justification, not a circular definition. References to prior work are external and not used to force the conclusion. The one weakness is that Section 4.3 asserts that 'obtaining any positive scalar multiple of v* ... is sufficient for gEM1 to recover the correct assignment' and then moves from the O(1/n) value-error bound to a claimed single-step recovery, but no uniform-margin/argmax-stability argument is supplied; the paper itself limits the claim: 'the statements above concern only the first step.' This is an unproved gap (and the abstract's O(1)-step phrasing outruns the theorem), not a circular step. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

1 free parameters · 7 axioms · 0 invented entities

No data fitting occurs: kappa is an algorithmic knob, and the extremal moments are derived, not fitted. The main mathematical content rests on the Gaussian-instance assumption and the unproved assignment-stability step.

free parameters (1)
  • kappa = algorithmic parameter on [0,1]; experiments use 0, 0.5, 1, and schedules
    Interpolates between label-fit and query-alignment in the hard assignment update (Eq. 28). It is chosen by hand, but it does not enter the data-generating process in the noiseless limit and the main value-vector concentration result is independent of it.
axioms (7)
  • domain assumption Instances are i.i.d. Gaussian, x_j^k ~ Normal(0, Sigma), and labels are generated noiselessly via h_infinity (Eq. 38).
    The entire asymptotic analysis in Lemmas 4.3-4.5 and Theorems 4.6/4.8 assumes this distribution; no non-Gaussian robustness is shown.
  • domain assumption Random assignments xi are sampled uniformly at a fixed Hamming distance from xi*, independent of X.
    Used in Theorem 4.6 to condition on match indicators and restore conditional independence of rows.
  • standard math Sub-Gaussian covariance concentration (Lemma 4.2, from Vershynin) applies to conditionally extremal rows.
    The proof of Theorem 4.6 relies on this bound, requiring sub-Gaussian norm O(sqrt(s)) and spectral norm O(s) for matched rows.
  • standard math Extremal moment asymptotics mu ~ sqrt(2 ln n), s ~ 2 ln n, and identities v = (1-s)/(n-1), w = v + 1.
    Standard extreme-value facts for i.i.d. Gaussian variables; drive the signal size phi_n(1/n) = 1/n.
  • ad hoc to paper The normalized average of assigned instances, q-hat_a, is a sufficient surrogate for the query MLE.
    Section 2.5 Eq. (37) defines the analyzed algorithm with this approximation; the paper justifies it heuristically for large Gaussian n, not by proving q-hat_a maximizes Q*.
  • domain assumption For gEM1, the query and value directions are aligned: q proportional to v.
    Section 4.3's one-step recovery claim is only for this aligned special case.
  • ad hoc to paper Directional closeness of v-hat to v* transfers to correctness of all argmax assignments.
    Section 4.3 asserts that a positive scalar multiple or sufficiently close direction yields correct gEM1 assignment; no margin or extremal-gap argument is provided.

pith-pipeline@v1.3.0-alltime-deepseek · 19311 in / 19173 out tokens · 187850 ms · 2026-08-01T18:03:46.658440+00:00 · methodology

0 comments
read the original abstract

In multiple instance regression (MIR) data are organized into bags (collections of instances in feature space) and the goal is to learn a mapping that assigns labels to bags. A typical assumption is that there is a so-called concept point in feature space, the proximity to which dictates the bag label. Motivated by modern MIR architectures which are based on attention, we study a softmax model that decouples the concept point and the labeling scheme. The two are respectively determined by a query direction and a value direction value in feature space. This problem isolates a basic challenge of learning both the query and value vectors from bag-level supervision. From this model we derive a parametric family of iterations in the noiseless limit, which generalizes a method known as the EM-DD algorithm. We then derive concentration results for the MLE estimators of the query and value vectors obtained from a random selection of instances. Our result for the value vector shows that a single random initialization of the value vector already points in the correct direction on average, so that a polynomial (in the number of instances per bag and the feature dimension) number of bags is enough for the EM algorithm to converge in $O(1)$ steps with high probability. A key aspect of this analysis is the interplay between concentration of empirical covariance matrices and extremal statistics arising from the selection rule.

Figures

Figures reproduced from arXiv: 2607.17404 by Ethan Levien.

Figure 1
Figure 1. Figure 1: Comparison of standard supervised learning and MIL. Each row represents a bag of instances (circles). Instances live in feature space, which here is R 2 with coordinates (x1, x2) represented by a hue (edge) and lightness (filling). In standard supervised learning each bag contains a single instance. In MIL, we select an active instance via some selection rule, which takes the lightest circle (middle) or th… view at source ↗
Figure 2
Figure 2. Figure 2: Comparison of the EM0 iteration (30) (top row) and EMg1 (32) (bottom row) as a function of the number of instances per bag n at t = 100 steps from a uniformly random initial assignment. The image plots (right) show which bags have correctly identified active instances. For clarity, we only show a subset of the instances. The n values corresponding to these examples are circled (left). Parameter values: θq∗… view at source ↗
Figure 3
Figure 3. Figure 3: Trajectories of the match fraction f (t) over iterations. Each panel shows a different variant of the hard EM iteration. Faint colored curves are individual trajectories initialized from random assignments. The bold black curve is the mean over all trajectories. Parameter values: m = 1000, n = 50, d = 10, θq∗,v∗ = 0. 0 π/4 π/2 1 − f mean 1 − f κ = 0 κ = 0.5 κ=0→1 EMg1 staged θˆv, v ∗ θv, v ˆ ∗ θˆqa, q ∗ θq… view at source ↗
Figure 4
Figure 4. Figure 4: Performance of the EMκ schedules, EMg1, and the staged method after t = 100 iterations, as a function of the misalignment angle θv∗,q∗ . Columns show the match-fraction error 1−f (left), the value angle θvˆ,v∗ (middle), and the query angle θqˆa,q∗ (right). The top row shows the mean over the ensemble of random initial assignments (shaded bands: ± one standard error); the bottom row shows the minimum over t… view at source ↗
Figure 5
Figure 5. Figure 5: Value update (Theorem 4.6), conditioned on a given match fraction f. The angle θvˆ,v∗ is shown for several values of the alignment parameter u = cos(θv∗,q∗ ), (left) as simulation vs. theory across f and (right) as a contour over (f, u), with the u values shown on the left indicated by dashed lines. Parameter values: n = 20, d = 3, m = 50,000. 102 103 m 100 θˆqa, q ∗ f = 0.3 f = 0.5 f = 0.8 f = 1.0 theory … view at source ↗
Figure 6
Figure 6. Figure 6: Query update (Theorem 4.8). (left) the angle θqˆa,q∗ as a function of m for several match fractions f, comparing simulation to the theoretical prediction Eq. (61). (right) a contour plot of the same prediction over (f, m), with the f values shown on the left indicated by dashed lines. Parameter values: n = 20, d = 500. Theorem 4.8. Let x k j be i.i.d. Normal(0, Id), and let ξ be sampled uniformly from {ξ ∈… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

51 extracted references · 3 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume=

    A framework for multiple-instance learning , author=. Advances in Neural Information Processing Systems , volume=

  2. [2]

    The Annals of Statistics , volume=

    Statistical guarantees for the EM algorithm: From population to sample-based analysis , author=. The Annals of Statistics , volume=

  3. [3]

    Neural computation , volume=

    On convergence properties of the EM algorithm for Gaussian mixtures , author=. Neural computation , volume=. 1996 , publisher=

    1996

  4. [4]

    Journal of the American statistical association , volume=

    Probability inequalities for sums of bounded random variables , author=. Journal of the American statistical association , volume=. 1963 , publisher=

    1963

  5. [5]

    Journal of Computer and System Sciences , volume=

    Real-valued multiple-instance learning with queries , author=. Journal of Computer and System Sciences , volume=. 2006 , publisher=

    2006

  6. [6]

    Zhang, Qi and Goldman, Sally A , booktitle=

  7. [7]

    European conference on computer vision , pages=

    Weakly-supervised action localization with expectation-maximization multi-instance learning , author=. European conference on computer vision , pages=. 2020 , organization=

    2020

  8. [8]

    Wiley Interdisciplinary Reviews: Computational Molecular Science , volume=

    Chemical complexity challenge: Is multi-instance machine learning a solution? , author=. Wiley Interdisciplinary Reviews: Computational Molecular Science , volume=. 2024 , publisher=

    2024

  9. [9]

    Journal of Machine Learning Research , volume=

    Multiple-instance learning of real-valued data , author=. Journal of Machine Learning Research , volume=

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Multi-instance multi-label learning with application to scene classification , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    Pattern Recognition , volume=

    Multiple instance learning: A survey of problem characteristics and applications , author=. Pattern Recognition , volume=. 2018 , publisher=

    2018

  12. [12]

    arXiv preprint arXiv:1409.0473 , year=

    Neural machine translation by jointly learning to align and translate , author=. arXiv preprint arXiv:1409.0473 , year=

    Pith/arXiv arXiv

  13. [13]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    2018

  14. [14]

    IEEE Transactions on fuzzy systems , volume=

    Switching regression models and fuzzy clustering , author=. IEEE Transactions on fuzzy systems , volume=. 1993 , publisher=

    1993

  15. [15]

    2008 , publisher=

    The EM algorithm and extensions , author=. 2008 , publisher=

    2008

  16. [16]

    SIAM Journal on Optimization , volume=

    On the convergence of alternating minimization for convex programming with applications to iteratively reweighted least squares and decomposition schemes , author=. SIAM Journal on Optimization , volume=. 2015 , publisher=

    2015

  17. [17]

    Learning from ambiguity , author=

  18. [18]

    Annals of Mathematical Statistics , volume=

    Adjustment of an inverse matrix corresponding to a change in one element of a given matrix , author=. Annals of Mathematical Statistics , volume=

  19. [19]

    Information Geometry , volume=

    Geometry of EM and related iterative algorithms , author=. Information Geometry , volume=. 2024 , publisher=

    2024

  20. [20]

    2016 , publisher=

    Information geometry and its applications , author=. 2016 , publisher=

    2016

  21. [21]

    Proceedings of the IEEE/CVF international conference on computer vision , pages=

    Expectation-maximization attention networks for semantic segmentation , author=. Proceedings of the IEEE/CVF international conference on computer vision , pages=

  22. [22]

    International conference on machine learning , pages=

    Attention-based deep multiple instance learning , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  23. [23]

    The 41st Conference on Uncertainty in Artificial Intelligence , year=

    Learning Algorithms for Multiple Instance Regression , author=. The 41st Conference on Uncertainty in Artificial Intelligence , year=

  24. [24]

    Pharmacological reviews , volume=

    Computational methods in drug discovery , author=. Pharmacological reviews , volume=. 2014 , publisher=

    2014

  25. [25]

    Performance comparisons on a steroid benchmark , author=

    Compass: predicting biological activities from molecular surface properties. Performance comparisons on a steroid benchmark , author=. Journal of Medicinal Chemistry , volume=. 1994 , publisher=

    1994

  26. [26]

    Artificial intelligence , volume=

    Solving the multiple instance problem with axis-parallel rectangles , author=. Artificial intelligence , volume=. 1997 , publisher=

    1997

  27. [27]

    2024 , publisher=

    Applied Extreme Value Statistics , author=. 2024 , publisher=

    2024

  28. [28]

    Artificial Intelligence and Statistics , pages=

    Two-stage sampled learning theory on distributions , author=. Artificial Intelligence and Statistics , pages=. 2015 , organization=

    2015

  29. [29]

    Advances in neural information processing systems , volume=

    Modern hopfield networks and attention for immune repertoire classification , author=. Advances in neural information processing systems , volume=

  30. [30]

    International Conference on Learning Representations , year=

    Hopfield Networks is All You Need , author=. International Conference on Learning Representations , year=

  31. [31]

    Physical Review Letters , volume=

    Exponential capacity of dense associative memories , author=. Physical Review Letters , volume=. 2024 , publisher=

    2024

  32. [32]

    Journal of Machine Learning Research , volume=

    Multiple-instance learning from distributions , author=. Journal of Machine Learning Research , volume=

  33. [33]

    arXiv preprint arXiv:2207.10227 , year=

    The tropical geometry of causal inference for extremes , author=. arXiv preprint arXiv:2207.10227 , year=

    Pith/arXiv arXiv

  34. [34]

    arXiv preprint arXiv:1903.07745 , year=

    Learning with sets in multiple instance regression applied to remote sensing , author=. arXiv preprint arXiv:1903.07745 , year=

    Pith/arXiv arXiv 1903

  35. [35]

    Proceedings of the International Conference on Machine Learning (ICML) , year =

    Soumya Ray and David Page , title =. Proceedings of the International Conference on Machine Learning (ICML) , year =

  36. [36]

    arXiv preprint , year =

    Thomas Uriot , title =. arXiv preprint , year =

  37. [37]

    Information Processing Letters , volume=

    On multiple-instance learning of halfspaces , author=. Information Processing Letters , volume=. 2012 , publisher=

    2012

  38. [38]

    Proceedings of Uncertainty in Artificial Intelligence (UAI) , year =

    Kushal Chauhan and Rishi Saket and Lorne Applebaum and Ashwinkumar Badanidiyuru and Chandan Giri and Aravindan Raghuveer , title =. Proceedings of Uncertainty in Artificial Intelligence (UAI) , year =

  39. [39]

    Tran , title =

    Carlos Améndola and Claudia Klüppelberg and Steffen Lauritzen and Ngoc M. Tran , title =. The Annals of Applied Probability , volume =

  40. [40]

    and McDonald, Michael J

    Good, Benjamin H. and McDonald, Michael J. and Barrick, Jeffrey E. and Lenski, Richard E. and Desai, Michael M. , title =. Nature , year =

  41. [41]

    and Shraiman, Boris I

    Neher, Richard A. and Shraiman, Boris I. , title =. Reviews of Modern Physics , year =

  42. [42]

    and Kaiser,

    Vaswani, Ashish and Shazeer, Noam and Parmar, Niki and Uszkoreit, Jakob and Jones, Llion and Gomez, Aidan N. and Kaiser,. Attention Is All You Need , booktitle =. 2017 , publisher =

    2017

  43. [43]

    and Laird, Nan M

    Dempster, Arthur P. and Laird, Nan M. and Rubin, Donald B. , title =. Journal of the Royal Statistical Society: Series B (Methodological) , volume =. 1977 , doi =

    1977

  44. [44]

    Advances in Neural Information Processing Systems , volume =

    Locatello, Francesco and Weissenborn, Dirk and Unterthiner, Thomas and Mahendran, Aravindh and Heigold, Georg and Uszkoreit, Jakob and Dosovitskiy, Alexey and Kipf, Thomas , title =. Advances in Neural Information Processing Systems , volume =

  45. [45]

    , title =

    Jaynes, Edwin T. , title =. Physical Review , volume =. 1957 , publisher =

    1957

  46. [46]

    A classification

    Celeux, Gilles and Govaert, G. A classification. Computational Statistics & Data Analysis , volume =. 1992 , publisher =

    1992

  47. [47]

    Advances in Neural Information Processing Systems , volume =

    Andrews, Stuart and Tsochantaridis, Ioannis and Hofmann, Thomas , title =. Advances in Neural Information Processing Systems , volume =

  48. [48]

    Applied Intelligence , volume=

    Multi-instance clustering with applications to multi-instance prediction , author=. Applied Intelligence , volume=. 2009 , publisher=

    2009

  49. [49]

    Vishnu, Adit and Shastry, Abhay and Kashyap, Dhruva and Bhattacharyya, Chiranjib , journal=

  50. [50]

    Multi-Instance Kernels , booktitle =

    G. Multi-Instance Kernels , booktitle =

  51. [51]

    2015 , publisher=

    An Introduction to Matrix Concentration Inequalities , author=. 2015 , publisher=

    2015