Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

Geometry as a Missing Axis of Representation Quality: The Variational Geometric Information Bottleneck under Data Scarcity

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read This paper claims that adding curvature and intrinsic-dimension penalties to the information bottleneck makes latent geometry part of the generalization guarantee, so intrinsic dimension, not ambient dimension, controls sample complexity.

desk verdict The paper's central theoretical claim is asserted rather than proved, and the empirical evidence for the headline gains is missing; the framework is plausible but the manuscript is not yet a reliable contribution. read the letter →

arxiv 2511.02496 v2 pith:TLJENBLT submitted 2025-11-04 cs.LG

Ronald Katende This is my paper
classification cs.LG MSC 68T0562R4068Q3253C21
keywords informationbottleneckcurvatureregularizationintrinsicdimensionsamplecomplexitymutualinterpretiveefficiencyfew-shotlearningrepresentationgeometry
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that latent geometry is a first-class axis of representation quality, not just a post hoc diagnostic: how well a representation generalizes from little labeled data is governed by the intrinsic dimension and curvature of the manifold it lives on. It proposes the unified objective U(φ)=I(φ(X);Y)−βC(φ)−γd_int(φ), adding curvature and dimension penalties to the task-information term, and proves non-asymptotic bounds in which the generalization gap scales like sqrt((d log N)/N) plus a curvature term βκmax, with no dependence on ambient dimension. The operative consequence is that smooth, low-dimensional encoders should need fewer labels, and among encoders with tied empirical risk the geometrically simpler one should generalize better. The accompanying V-GIB estimator is reported to improve accuracy and reduce geometric complexity on FashionMNIST and CIFAR-10 subsets under low label fractions, while the paper also notes no single regularizer dominates everywhere.

What carries the argument

The object carrying the argument is the understanding functional U(φ)=I(φ(X);Y)−βC(φ), with C(φ) combining a Hessian-norm curvature penalty and an intrinsic-dimension penalty. The proof mechanism is the curvature-sensitive covering-number bound: for a compact d-dimensional submanifold of reach τ and sectional curvature at most κmax, the covering number grows like (1+cκmax/τ)^d ε^(−d). That multiplicative geometry factor converts into the d log(1+cκmax/τ) term in the generalization bound via standard covering-to-complexity arguments, while the Hessian penalty supplies the additive βκmax approximation error. The V-GIB implementation estimates these quantities with a variational surrogate for m

What would settle it

Compute the generalization gap for the same encoder family on one data manifold embedded in R^3, R^30, and R^300 while keeping the intrinsic dimension, labels, and training size fixed. If the gap rises with ambient dimension, Theorem 2.4's D-independence is wrong; if it stays flat, the intrinsic-dimension claim has direct support.

Watch

Extended reading notes

Core claim

The paper's central claim is Theorem 2.4: a minimizer of the unified objective has, with probability at least 1−δ, generalization gap bounded by C sqrt((d log N + log(1/δ))/N) + C′βκmax, with constants independent of the ambient dimension D. The first term is the paper's core thesis—once data are assumed to live on a compact smooth d-dimensional manifold, sample complexity is controlled by intrinsic dimension, not input dimensionality. The second term says curvature regularization itself costs an additive approximation error proportional to the regularizer weight β times the manifold's maximum sectional curvature κmax. From this, Corollary 2.7 orders encoders with equal empirical risk: the f

Load-bearing premise

The whole bound rests on Assumption 2.3—data supported on a compact smooth d-dimensional manifold with bounded reach and sectional curvature and a bi-Lipschitz encoder—and the paper's own Appendix A.11 concedes this may fail for high intrinsic dimension or non-manifold data.

Editorial extensions

If this is right

  • Intrinsic dimension becomes the right complexity certificate: if Theorem 2.4 holds, adding irrelevant ambient coordinates should not raise the number of labels needed, provided the data still sit on the same low-dimensional manifold.
  • Low-curvature bias substitutes for labels: the paper reports that curvature-regularized models reach the same accuracy with roughly 32% fewer examples on its synthetic setup, matching the theoretical reduction in sample complexity.
  • Model selection by geometry: when two encoders tie on empirical risk, Corollary 2.7 says the one with lower estimated intrinsic dimension and lower curvature is the one to keep.
  • A Pareto law for representation design: tuning the curvature weight β traces a monotone frontier between task information and geometric complexity, with marginal trade-off equal to β; the paper reports positive KL–curvature coupling and stable frontiers across datasets.
  • Diminishing returns under more data: CIFAR-10 fractional runs show interpretive efficiency rising with data while curvature saturates, implying that once the geometry equilibrates, extra labels buy less.

Reading between the lines

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

  • A direct test of the D-independence claim: embed the same synthetic manifold in spaces of increasing ambient dimension, keep d and the encoder's latent dimension fixed, and check whether the generalization gap stays flat as D grows; Theorem 2.4 predicts it should, and the paper's synthetic protocol could be extended to do precisely this.
  • If the additive βκmax term is real, curvature regularization should have a U-shaped effect on generalization for data with high curvature: raising β from zero first reduces the statistical term by smoothing the latent manifold, then eventually hurts by paying approximation cost. A sweep of β on a known high-curvature dataset would expose the predicted minimum.
  • Theorem 2.8's frontier could be used as a pre-training calibration tool: fit the information–curvature curve from a few values of β and extrapolate to predict the achievable task information at a target curvature before committing compute to full training.
  • The framework's separation of encoder geometry from predictor complexity suggests a transfer test: a V-GIB encoder trained on one label set should transfer to a new task better than an equally accurate non-geometric encoder, because the latent manifold, not the label map, carries the reusable structure.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The manuscript defines a structural-understanding functional U(φ)=I(φ(X);Y)−βC(φ), instantiates it as V-GIB, and claims non-asymptotic generalization bounds in which intrinsic dimension d and curvature κmax enter separately (Theorems 2.4 and 2.6). It also states an information–curvature Pareto frontier, estimator concentration results, and an alignment–efficiency synergy theorem. Experiments on Swiss roll, Fashion-MNIST, and CIFAR-10 with reduced label fractions are presented as validation.

Significance. If the theoretical results were established, the paper would make a useful contribution: quantifying how latent curvature enters sample-complexity bounds beyond standard intrinsic-dimension results, and showing curvature regularization as a substitute for labels. The paper also ships implementation details, a code link, and reproducibility diagnostics (Appendix A.10, A.12), which is commendable. However, the two load-bearing geometric results are not actually derived: the additive βκmax term is asserted, and the curvature covering factor in Lemma 2.5 is dimensionally inconsistent. The circularity in Theorem 2.17 further weakens the theoretical contribution. The empirical correlations are suggestive but are not presented with comparisons that would distinguish the claimed law from a generic accuracy-curvature correlation during training.

major comments (4)
  1. [Theorem 2.4 and Appendix A.2, proof sketches] The central new term C′βκmax in Theorem 2.4 is asserted, not derived. The proof sketch says 'imposing weight β on the curvature functional produces an additional approximation/stability error of order βκmax', and Appendix A.2 step (iii) repeats 'curvature-regularized encoders ... incur an additional approximation error bounded by C3βκmax' without specifying the constrained hypothesis class, the approximation error being measured, or a bound on it. This term is exactly what makes the claim 'geometry becomes part of the bottleneck criterion' non-trivial; without a derivation, Theorem 2.4 is only the standard intrinsic-dimension bound plus an unexplained additive constant. The authors should either prove this term from an explicit class of curvature-bounded encoders or state it as a conjecture and remove it from the theorem.
  2. [Lemma 2.5, Eq. (2)] Eq. (2) is dimensionally inconsistent. Sectional curvature κmax has units L^{-2} and reach τ has units L, so κmax/τ has units L^{-3}, yet it appears inside a dimensionless factor (1+c1 κmax/τ)^d. The Bishop–Gromov comparison stated in the appendix (Eq. (6)) gives a dimensionless factor 1+c_d κmax r^2, with r the ball radius; replacing r by τ does not yield κmax/τ. Since Theorem 2.6's d log(1+c1κmax/τ) term is obtained directly from Eq. (2), the curvature-aware deviation bound is unsupported. The paper needs a correct curvature-sensitive covering estimate (e.g., in terms of κmax τ^2) or a different mechanism by which curvature enters the rate.
  3. [Theorem 2.17] Assumption (iii) already asserts the conclusion of the theorem: 'if A(φ,ψ) ≥ A(φ′,ψ) then I(φ(X);Y) ≥ I(φ′(X);Y)' is exactly the claim that alignment increases predictive information. The proof derives I-monotonicity from this assumption and then defines μ* as the infimum where I becomes constant; existence of a finite μ* does not follow from boundedness and monotonicity, since a strictly increasing bounded function need not become constant. The theorem therefore does not establish an alignment–efficiency synergy; it re-labels an assumption as a result. Please state assumption (iii) as the substantive open condition, or provide a proof from more basic information-theoretic properties.
  4. [Proposition 2.11] The claim sup_φ |Î−I| = O((NK)^(-1/2)) is derived from Theorem 2.14 and Proposition 2.15, but neither result is uniform over φ. Theorem 2.14 bounds |bC_{N,K}−E[bC_{N,K}]| for fixed h, and Proposition 2.15 bounds the VIB surrogate for a fixed decoder class; no metric-entropy or union-bound argument over the encoder class is given. Pointwise concentration does not imply sup-norm concentration. Consequently, the uniform consistency of empirical maximizers does not follow. Please supply a uniform bound over a compact class with an explicit covering argument, or restrict Proposition 2.11 accordingly.
minor comments (6)
  1. [Table 3] Table 3 appears corrupted: rows contain overlapping entries, and the row for frac=0.20 seems to include an implausible epoch value '520' inconsistent with other rows. The table needs to be regenerated and checked.
  2. [Figure captions 1–3] Several figure captions contain raw '/uni...' escape sequences (e.g., Figure 1 and Figure 3 captions). These will render incorrectly in the published version and should be fixed.
  3. [Section 4.3, η_eff] The 'effective sample efficiency ratio' η_eff is used to support a quantitative claim ('roughly 32% fewer labeled examples') but is never formally defined in Section 3 or 4. Please provide its definition and estimation procedure.
  4. [§4.4 and §4.6, E(φ;N)] Interpretive efficiency is defined in Definition 2.2 as U(φ)/N, but in the CIFAR-10 tables it is reported as acc/align (e.g., 23.33). This is a different quantity; the notation should be reconciled or the tables relabeled.
  5. [References] Reference [6] is incomplete: it has no journal, arXiv identifier, or year. Also, citations [9] and [6] are used for covering-number estimates that are not standard consequences of those papers; please give precise statements.
  6. [Appendix A.11] The limitation note that 'geometric assumptions may fail for high intrinsic dimension or non-manifold data' is placed only in the appendix. Given that the main theory and the experiments on CIFAR-10 depend on Assumption 2.3, this caveat should be stated prominently in the main text and tested with the provided bdPR diagnostics.

Circularity Check

3 steps flagged · score 6.0 of 10

Alignment–efficiency 'synergy' is assumed in Theorem 2.17; intrinsic-dimension selection and efficiency ordering are definitional rewrites, so several advertised theoretical results reduce to their own inputs.

  1. self definitional [Theorem 2.17, assumption (iii) and its proof, Section 2.5]
    "Assume that ... (iii) if A(ϕ,ψ)≥ A(ϕ′,ψ) then I(ϕ(X);Y)≥I(ϕ ′(X);Y), with strict inequality in the nondegenerate case; (iv) E(ϕ;N) is nondecreasing in I(ϕ(X);Y). Then µ→E(ϕ µ;N) is nondecreasing ... Proof: ... Assumption (iii) implies I(ϕ µ2 (X);Y)≥I(ϕ µ1 (X);Y), hence by (iv) E(ϕµ2 ;N)≥E(ϕ µ1 ;N)."

    The advertised 'Alignment–efficiency synergy' is exactly the assumption that alignment order implies task-information order (iii), combined with the definitional monotonicity of E in I (iv). The proof's only real step (optimality ⇒ A is nondecreasing in µ) is a standard envelope argument; the claimed synergy itself is never derived. Thus the theorem's conclusion is its hypothesis, relabeled.

  2. self definitional [Proposition 2.9 and Theorem 2.12, Section 2.5]
    "Define V(d ′) = sup_{ϕ∈Φd′} {I(ϕ(X);Y)−βC(ϕ)}. Then ... (b) adding a dimension penalty γd ′ yields d∗ ∈ arg max_{d′≤D} {V(d ′)−γd ′}. ... Proof. Partition F=∪Φd′. For fixed d ′, inf_{ϕ∈Φd′} J(ϕ) = γd ′ − V(d ′)."

    V(d′) is defined as sup(I−βC) over the class Φd′, and J is defined as −I+βC+γd′. Therefore inf_{Φd′} J = γd′−V(d′) follows by substituting the definitions. The 'intrinsic-dimension selection' theorem and the 'curvature–information–dimension linkage' are algebraic restatements of J and V, not independently derived predictions.

1 more flagged steps
  1. self definitional [Corollary 2.7 proof, Section 2.4]
    "By definition of E(ϕ;N) as a decreasing function of both d int and κ, E(ϕ 1;N)>E(ϕ 2;N)."

    Interpretive efficiency is defined in Definition 2.2 as E(ϕ;N)=U(ϕ)/N with U=I−βC, so 'lower d_int and lower κ imply higher E' is true by construction. The corollary presents this definitional fact as a consequence of the sample-complexity bound, but the E-ordering half does not use Theorem 2.4 at all.

full rationale

The clearest constructional circularity is Theorem 2.17: the theorem's central claim—alignment improves efficiency—is assumed as condition (iii) (higher alignment ⇒ higher mutual information) and condition (iv) (higher information ⇒ higher efficiency), and the proof applies those assumptions verbatim. Similarly, Proposition 2.9 and Theorem 2.12 unfold the definitions of V(d′) and J; their conclusions are algebraic identities rather than derived selection laws. Corollary 2.7's efficiency-ordering component is a restatement of Definition 2.2. The curvature-dependent generalization theorems (2.4 and 2.6) are not circular in the same constructional sense—the additive Cβκmax term and the κmax/τ covering factor are asserted rather than derived, and the latter is dimensionally suspect—but those are unsupported-claim/correctness concerns, not reductions to inputs. There is no load-bearing self-citation chain; the external references are standard and not author-overlapping. Because several advertised theoretical results reduce by definition or by assuming their conclusion, but the paper's main generalization claim is not itself constructionally circular, the score is partial rather than total.

Assumptions & free parameters 6 free parameters · 9 assumptions · 4 invented entities

The quantitative payload rests on (i) a manifold-regularity assumption that may fail for the very datasets tested (conceded in A.11), (ii) an asserted additive curvature-bias term that is never derived, (iii) theorem hypotheses (Thm 2.17 iii) that encode the conclusion, and (iv) hyperparameters β, γ, λ set by hand (λ never specified) plus the undefined 'effective sample efficiency ratio' η_eff used to claim label savings. No new falsifiable entity is introduced with evidence independent of the paper's own pipeline; the empirical metrics either mismatch the theory (E, curvature proxy) or are undefined (η_eff, alignment MI).

free parameters (6)
  • β (bottleneck/curvature weight) = {10^-3, 5×10^-3, 10^-2}
    Chosen by hand; the 'Pareto frontier' in Fig 1(d) is traced by sweeping β, so the reported trade-off slope dI/dC = β is imposed by the sweep, not independently predicted.
  • γ (curvature/dimension penalty weight) = {0, 10^-4} synthetic; {0,10^-5,10^-4} CIFAR
    Chosen by hand; the dimension penalty γdim(S) never appears in the generalization bounds, so its claimed effect is asserted but not analyzed.
  • λ (predictor regularizer in Eq. 1)
    Appears in the unified objective (1) but is never given a value or ablation.
  • η_eff ('effective sample efficiency ratio') = 1.32 (average)
    Undefined in the paper; used to claim 'roughly 32% fewer labeled examples for the same accuracy'—the central data-efficiency claim rests on this unspecified quantity.
  • K (Hutchinson probes) = 2 (train), ≥4 (eval)
    Estimator setting; the concentration rate (NK)^(-1/2) in Theorem 2.14/Prop 2.11 depends on it.
  • z_dim (latent dimension) = {8,16}
    Chosen by hand; the 98.2% headline synthetic result uses z_dim=16 (best seed).
assumptions (9)
  • domain assumption Manifold hypothesis with regularity (Assumption 2.3): data supported on a compact smooth d-dimensional Riemannian manifold M ⊂ R^D with reach τ>0 and |κ|≤κmax
    Load-bearing for Theorems 2.4/2.6 and Corollary 2.7: the d-dependence of the bounds and the covering estimates presuppose this structure; the paper's own A.11 concedes it may fail for high-dimensional or non-manifold data.
  • domain assumption Encoder φ is bi-Lipschitz on M (Assumption 2.3 / A.1)
    Bi-Lipschitzness (0<Lmin≤Lmax<∞) is used to transport covering numbers from the data manifold to the latent space; it is not verified for the trained neural network encoders.
  • standard math Loss ℓ bounded in [0,B] and Lipschitz; predictor g is Lg-Lipschitz (Thm 2.4 conditions)
    Standard Rademacher hypotheses; uncontroversial.
  • ad hoc to paper Curvature-regularized encoders incur an additive approximation/stability error C′βκmax (Thm 2.4 sketch, Thm 2.6, App. A.2 step iii)
    Asserted in the proof sketch with no derivation; the excess risk of a regularizer is not generally linear in its weight. The Lemma 2.5 factor (1 + c κmax/τ)^d is also dimensionally inconsistent (κmax has units L^-2, τ has units L).
  • domain assumption Sub-Gaussianity of Hessian-vector products h_v(z)=v^T H_φ(z)v and per-sample MI contributions (Lemma 2.13)
    The concentration results (Thm 2.14, Prop 2.11) rest on these; not checked for the trained models.
  • ad hoc to paper Theorem 2.17 assumption (iii): if A(φ,ψ)≥A(φ′,ψ) then I(φ(X);Y)≥I(φ′(X);Y)
    This hypothesis IS the theorem's conclusion (alignment improves task information); the result is forced by the assumption, not derived.
  • standard math Existence/uniqueness of maximizers, C^1 dependence, positive-definite reduced Hessian (Thm 2.8/2.10)
    Regularity needed for the envelope-theorem argument; generic but unverified for neural network utility surfaces.
  • standard math Covering estimates for submanifolds with bounded curvature N(ε)≤C_0(1+c_1 κmax/τ)^d ε^(-d) (Lemma 2.5, App. A.3)
    Attributed to [6,9]; the specific curvature-inflated form is where the paper's contribution would live, but it is referenced rather than proved, and its dimensional inconsistency is unresolved.
  • domain assumption VIB surrogate consistency: decoder family {q_ω} has Rademacher complexity R_dec(N) and Ĥ(Y) error negligible (Prop 2.15)
    Controls the gap between the variational MI surrogate and true I(φ(X);Y); since the estimator is a lower bound, the claimed 'consistency' is one-sided and decoder-quality dependent.
invented entities (4)
  • Structural understanding functional U(φ)
    purpose: Postulated scalar that 'formalizes understanding' as I(φ(X);Y) − βC(φ), anchoring the entire framework.
    A definition, not a measurable entity; its empirical counterpart is never directly computed (experiments report accuracy and KL traces instead).
  • Interpretive efficiency E(φ;N)
    purpose: Claims to measure 'usable predictive structure per labeled example' and to be the quantity ordered by Corollary 2.7.
    Two inconsistent definitions: theory E=U/N, experiments E=acc/align-MI; dividing accuracy by a mutual information in nats has no defined semantics.
  • Effective sample efficiency ratio η_eff
    purpose: Quantifies the central claim that curvature regularization buys roughly 32% label savings.
    Never defined; no formula or code reference given anywhere in the manuscript.
  • Alignment mutual information (empirical)
    purpose: Drives the claimed inverse accuracy-alignment correlation and the efficiency numbers in Tables 1 and 3.
    The concept variables C_ψ and the MI estimator are not specified for CIFAR-10/FashionMNIST; a per-epoch MI decaying 0.0407→0.0291 is reported without a definition.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Geometry as a Missing Axis of Representation Quality: The Variational Geometric Information Bottleneck under Data Scarcity." pith.science (2026). https://pith.science/paper/TLJENBLT

@misc{pith2026251102496,
  author       = {Pith},
  title        = {Pith review of: Geometry as a Missing Axis of Representation Quality: The Variational Geometric Information Bottleneck under Data Scarcity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TLJENBLT}},
  note         = {Machine review of arXiv:2511.02496}
}
read the original abstract

We study latent geometry as an explicit component of representation quality in data-scarce learning. For an encoder (\phi), we define (Q_{\beta,\gamma}(\phi)=I(\phi(X);Y)-\beta\mathcal C(\phi)-\gamma d_{\mathrm{int}}(\phi)), combining task-relevant information with penalties for curvature and intrinsic latent dimension. Thus geometry becomes part of the bottleneck criterion, not only a post hoc diagnostic. Under smooth-manifold, loss-transfer, and estimator-concentration assumptions, we derive non-asymptotic low-label generalization bounds where intrinsic dimension and covering complexity enter explicitly. We characterize the information--geometry frontier and prove empirical-surrogate consistency. The analysis links encoder geometry to learning through latent covering numbers, loss-class entropy, and uniform deviation. We instantiate the theory as \texttt{V-GIB}, adding curvature and dimension penalties to variational bottleneck training. Real low-label benchmarks compare \texttt{V-GIB} with ERM, VIB, and ablations across (1%)--(20%) label fractions. Results show improved performance and reduced geometric complexity in several regimes, especially FashionMNIST and CIFAR-10, while confirming that no fixed regularizer is universally dominant.

Figures

Figures reproduced from arXiv: 2511.02496 by the authors.

Figure 1
Figure 1. Empirical characterization of V-GIB. (a) The energy landscape exhibits a positive coupling between curvature and information energy (ρKL,κ = 0.67). (b) Curvature regularization improves effective sample effi￾ciency across noise levels. (c) Structured encoders out￾perform random baselines, confirming that geometric reg￾ularization; not model size; drives performance. (d) The information–curvature Pareto frontier show… view at source ↗
Figure 2
Figure 2. CIFAR-10 learning dynamics. Accuracy in￾creases (blue) as alignment MI decreases (red), with equi￾librium near epoch 60. Shaded regions indicate ±1σ over seeds [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (3 more)
Figure 6
Figure 6. Figure 6: Correlation of metrics. KL, curvature, loss, accuracy across epochs/seeds. B.4 Extended CIFAR-10 Diagnostics B.4.1 Epoch-wise metrics [PITH_FULL_IMAGE:figures/full_fig_p014_6.png]
Figure 5
Figure 5. Figure 5: Estimator stability. MI and curvature trajec￾tories over epochs (mean ± std over 3 seeds) [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 7
Figure 7. Figure 7: Distribution of interpretive efficiency acc/align across epochs. The right-skewed tail indicates progressive tightening of geometry as training advances. B.4.4 Saturation analysis Efficiency slope d(acc)/d(align) flattens around epoch 60, marking the transition from ge…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Auditing Training-Free 3D Shape Retrieval with Diffused Geodesic Moments

    cs.CV 2026-05 unverdicted novelty 6.0 of 10

    DGM is presented as a seed-conditioned descriptor to audit protocols, showing that input field and aggregation dominate moment formulas, with GMSD-HKS achieving top scores on FAUST-Reg and TOSCA.

Reference graph

Works this paper leans on

22 extracted references · 3 linked inside Pith · cited by 1 Pith paper

  1. [1]

    Human-level concept learning through probabilis- tic program induction,

    B. M. Lake, R. Salakhutdinov, and J. B. Tenenbaum, “Human-level concept learning through probabilis- tic program induction,”Science, vol. 350, no. 6266, pp. 1332–1338, 2015

    2015

  2. [2]

    Matching 10 networks for one shot learning,

    O. Vinyals, C. Blundell, T. Lillicrap, K. Kavukcuoglu, and D. Wierstra, “Matching 10 networks for one shot learning,” inAdvances in Neural Information Processing Systems (NeurIPS), 2016

    2016

  3. [3]

    Model-agnostic meta-learning for fast adaptation of deep networks,

    C. Finn, P. Abbeel, and S. Levine, “Model-agnostic meta-learning for fast adaptation of deep networks,” inProceedings of the 34th International Conference on Machine Learning (ICML), vol. 70 ofProceedings of Machine Learning Research, pp. 1126–1135, 2017

    2017

  4. [4]

    Representa- tion learning: A review and new perspectives,

    Y. Bengio, Y. LeCun, and G. Hinton, “Representa- tion learning: A review and new perspectives,”IEEE Transactions on Pattern Analysis and Machine In- telligence, 2013. review; see also arXiv:1206.5538

    arXiv 2013

  5. [5]

    Pearl,Causality: Models, Reasoning and Inference

    J. Pearl,Causality: Models, Reasoning and Inference. Cambridge University Press, 2nd ed., 2009

    2009

  6. [6]

    Testing the manifold hypothesis,

    C. Fefferman, S. Mitter, and H. Narayanan, “Testing the manifold hypothesis,”

  7. [7]

    Geometric deep learning: go- ing beyond euclidean data,

    M. M. Bronstein, J. Bruna, Y. LeCun, A. Szlam, and P. Vandergheynst, “Geometric deep learning: go- ing beyond euclidean data,”IEEE Signal Processing Magazine, vol. 34, no. 4, pp. 18–42, 2017

    2017

  8. [8]

    A global geometric framework for nonlinear di- mensionality reduction,

    J. B. Tenenbaum, V. de Silva, and J. C. Lang- ford, “A global geometric framework for nonlinear di- mensionality reduction,”Science, vol. 290, no. 5500, pp. 2319–2323, 2000

    2000

Show all 22 references
  1. [9]

    Manifold regularization: A geometric framework for learning from labeled and unlabeled examples,

    M. Belkin, P. Niyogi, and V. Sindhwani, “Manifold regularization: A geometric framework for learning from labeled and unlabeled examples,”Journal of Machine Learning Research, vol. 7, pp. 2399–2434, 2006

    2006

  2. [10]

    The in- formation bottleneck method,

    N. Tishby, F. C. Pereira, and W. Bialek, “The in- formation bottleneck method,” inProceedings of the 37th Annual Allerton Conference on Communication, Control and Computing, pp. 368–377, 1999

    1999

  3. [11]

    Deep variational information bottleneck,

    A. A. Alemi, I. Fischer, J. V. Dillon, and K. Murphy, “Deep variational information bottleneck,”arXiv preprint arXiv:1612.00410, 2016. Presented at ICLR 2017

    2016 arXiv

  4. [12]

    T. M. Cover and J. A. Thomas,Elements of Informa- tion Theory. Hoboken, NJ: Wiley-Interscience, 2 ed., 2006

    2006

  5. [13]

    Maximum likelihood es- timation of intrinsic dimension,

    E. Levina and P. J. Bickel, “Maximum likelihood es- timation of intrinsic dimension,” inAdvances in Neu- ral Information Processing Systems 17 (NIPS 2004), pp. 777–784, 2005

    2004

  6. [14]

    A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines,

    M. F. Hutchinson, “A stochastic estimator of the trace of the influence matrix for laplacian smoothing splines,”Communications in Statistics – Simulation and Computation, vol. 18, no. 3, pp. 1059–1076, 1989

    1989

  7. [15]

    Towards a rigorous science of interpretable machine learning,

    F. Doshi-Velez and B. Kim, “Towards a rigorous science of interpretable machine learning,”arXiv preprint, vol. arXiv:1702.08608, 2017

    2017 arXiv

  8. [16]

    Explanation in artificial intelligence: In- sights from the social sciences,

    T. Miller, “Explanation in artificial intelligence: In- sights from the social sciences,”Artificial Intelli- gence, vol. 267, pp. 1–38, 2019

    2019

  9. [17]

    V. N. Vapnik,Statistical Learning Theory. Wiley- Interscience, 1998

    1998

  10. [18]

    Rademacher and gaussian complexities: Risk bounds and structural re- sults,

    P. L. Bartlett and S. Mendelson, “Rademacher and gaussian complexities: Risk bounds and structural re- sults,”Journal of Machine Learning Research, vol. 3, pp. 463–482, 2002

    2002

  11. [19]

    Chavel,Riemannian Geometry: A Modern Intro- duction (2nd ed.)

    I. Chavel,Riemannian Geometry: A Modern Intro- duction (2nd ed.). Cambridge University Press, 2006

    2006

  12. [20]

    Shalev-Shwartz and S

    S. Shalev-Shwartz and S. Ben-David,Understand- ing Machine Learning: From Theory to Algorithms. Cambridge University Press, 2014

    2014

  13. [21]

    A. W. van der Vaart and J. A. Wellner,Weak Con- vergence and Empirical Processes: With Applications to Statistics. Springer Series in Statistics, New York, NY, USA: Springer-Verlag, 1996

    1996

  14. [22]

    Mine: Mutual information neural estimation,

    M. I. Belghazi, A. Baratin, S. Rajeswar, S. Ozair, Y. Bengio, A. Courville, and R. D. Hjelm, “Mine: Mutual information neural estimation,” inProceed- ings of the 35th International Conference on Machine Learning (ICML 2018), pp. 531–540, PMLR, 2018. A Appendix: Technical De- t...

    2018

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.