Pith. sign in

REVIEW 3 major objections 4 minor 59 references

Task-Driven Discrete Representation Learning

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper proves that task-driven discrete representation learning can be formulated as an unconstrained objective combining task loss with Wasserstein codebook alignment, and quantifies the accuracy–sample-complexity tradeoff of…

desk verdict A well-organized framework with real experiments, but the two theoretical results are not actually proven as written. read the letter →

arxiv 2506.11511 v1 pith:NZK3O6SW submitted 2025-06-13 cs.LG

Tung-Long Vuong This is my paper
classification cs.LG
keywords discreterepresentationlearningcodebookquantizationtask-drivenWassersteindistancesamplecomplexityalignmentdomaingeneralizationstateabstraction
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 proposes and defends a unified, task-driven view of discrete representation learning: instead of judging a discrete code by how well it reconstructs or generates data, the code is judged by how well it serves a downstream task such as classification, regression, or control. It proves that the constrained problem of learning a codebook-valued encoder is equivalent, under infinite-capacity assumptions, to an unconstrained objective with a Wasserstein codebook-alignment term, which makes end-to-end gradient training a principled choice. The theory then identifies a quantitative tradeoff: smaller codebooks cap the best achievable accuracy but need fewer samples, while larger codebooks improve accuracy but drive sample complexity up as the square of the codebook size. The paper demonstrates the formulation on state abstraction for reinforcement learning and on domain generalization, where aligning fine-grained discrete representations improves over class-level alignment.

What carries the argument

The central object is the codebook-quantized encoder $\bar{f}_e = Q_C \circ f_e$ over a finite codebook $C=\{c_m\}_{m=1}^M$; the hard constraint $\bar{f}_e\#P_x = P_{c,\pi}$ is relaxed into the Wasserstein term $\lambda W_{d_z}(f_e\#P_x, P_{c,\pi})$. The analysis splits the total error into two parts: a clustering loss (distance from the label to its nearest decoded centroid) and an assignment-discrepancy loss (the surplus cost from assigning to the nearest code in latent space rather than the nearest centroid in output space). The sample-complexity proof controls the assignment-discrepancy term with a concentration inequality over the $M\times M$ product codebook.

What would settle it

The bound in Corollary 3.3 would be falsified if, on a finite training set, the empirically optimal discrete encoder showed a strictly positive assignment-discrepancy term at some training point, or if the empirical generalization gap did not decrease at the $O(N^{-1/2})$ rate with the stated $\Omega(M^2)$ dependence.

Watch

Extended reading notes

Core claim

In the paper's own terms, the constrained discrete representation learning problem in Eq. (4) is equivalent, for infinite-capacity encoder and decoder families, to minimizing $\mathbb{E}_{x\sim P_x}[\ell_y(f_d(Q_C(f_e(x))), f_y(x))] + \lambda W_{d_z}(f_e\#P_x, P_{c,\pi})$. From this equivalence the paper derives two consequences: the optimal discrete representation solves an optimal clustering problem in the output space, so the achievable loss decreases as the codebook grows and tends to zero as $M\to\infty$; and the assignment-discrepancy part of the generalization gap has sample complexity $\Omega(M^2)$, giving an explicit accuracy-versus-data tradeoff. Experiments apply the framework to state abstraction in reinforcement learning and to fine-grained domain alignment in domain generalization.

Load-bearing premise

The sample-complexity proof assumes that the empirically optimal model makes the assignment-discrepancy term exactly zero on every training example, which minimizing an average loss does not by itself guarantee.

Editorial extensions

If this is right

  • End-to-end gradient training of discrete task representations is justified: the unconstrained objective with the Wasserstein term converges to the hard-constrained problem in the limit of infinite-capacity function families.
  • Codebook size becomes a tunable bias-variance knob: small codebooks generalize from fewer samples at the price of lower ceiling accuracy, and large codebooks raise the ceiling while demanding $\Omega(M^2)$ samples.
  • The generalization rate of the discrete-representation loss remains $O(N^{-1/2})$, matching k-means clustering, so discretization does not asymptotically worsen learning rates.
  • Lemma 3.2 supports multi-level discrete representations: enlarging the codebook yields a more detailed description of the data, which the domain-generalization experiments exploit via modality-wise alignment.
  • For domain generalization, aligning representations through fine-grained discrete codes (FDANN) yields consistent gains over whole-distribution and class-level alignment baselines on the reported benchmarks.

Reading between the lines

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

  • The equivalence in Theorem 3.1 is stated for infinite-capacity families; for finite networks the Wasserstein term is a soft constraint, so a natural test is whether the hard constraint is actually satisfied at convergence on real tasks.
  • The $\Omega(M^2)$ sample-complexity dependence is driven by the assignment-discrepancy term; if the empirical optimum does not achieve zero assignment discrepancy on the training set, the true dependence may be weaker or stronger than the stated bound, and measuring this term directly would settle it.
  • The framework could be applied to other discrete-latent settings such as planning and reasoning, where the accuracy–sample tradeoff would predict when small codebooks are preferable; that is a direct but unstated extension of Remark 2.
  • The FDANN idea suggests that the number of discrete representations per class is a hyperparameter with an optimum, as the reported drop at $M=K\times 32$ indicates; ablating over more datasets and class-to-codebook ratios would tell whether the optimum is robust.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces a task-driven discrete representation learning framework. Eq. (4) formulates a constrained optimization problem over a codebook distribution and a discrete encoder, and Theorem 3.1 claims that this problem is equivalent to the unconstrained, gradient-trainable objective in Eq. (5), which uses a continuous encoder, a quantization operator, and a Wasserstein penalty. Section 3.1 then states Lemma 3.2, an accuracy trade-off showing that optimal loss decreases as codebook size grows, and Corollary 3.3, a sample-complexity bound of order O(M^2) for the assignment-discrepancy loss. The experiments apply the framework to state abstraction in reinforcement learning and to a fine-grained discrete domain alignment method (FDANN), reporting accuracy gains over DANN and CDANN on VLCS, PACS, and Office-Home.

Significance. If the theoretical claims were valid, the paper would provide a principled task-driven alternative to VQ-VAE-style discrete representation learning and a formal account of the capacity-versus-sample-complexity trade-off of discrete codebooks. The empirical FDANN results are a useful contribution: they are reproducible in principle, the comparisons with DANN and CDANN are standard, and the reported gains, while modest, are consistent across benchmarks. The RL state-abstraction experiments also illustrate the claimed trade-off. However, the central theoretical results are not established by the supplied proofs: the proof of Theorem 3.1 does not prove the advertised equivalence, the proof of Corollary 3.3 relies on a false pointwise optimality assertion, and the proof of the monotonicity part of Lemma 3.2 ignores the effect of adding a codebook entry on the quantization of other points. Because the theoretical analysis is the paper's main advertised contribution, these gaps are load-bearing.

major comments (3)
  1. [§6.1, Theorem 6.2 (Theorem 3.1)] The proof does not establish the claimed equivalence between Eq. (4)/(12) and Eq. (5)/(13). In the proof, a 'feasible solution C, π, f_d, and \bar f_e of the OP in (18)' is assumed to satisfy \bar f_e#P_x = P_{c,π} and Q_C(\bar f_e(x)) = \bar f_e(x), but Eq. (13) imposes no exact equality of this kind: the Wasserstein term is only penalized with a finite λ and may be positive. Those two properties hold for feasible solutions of the constrained problem (12), not for feasible solutions of (13). Consequently the inequality chain (15)-(17) compares Eq. (12) with a constrained variant rather than with the actual unconstrained objective (13), and the claimed equivalence is unsupported. This matters because Eq. (5)/(13) is the surrogate objective used in the experiments. The equivalence can be proved with a two-sided argument, but that argument is not present in the manuscript.
  2. [§6.3, Corollary 6.6 (Corollary 3.3)] The proof of Corollary 6.6 asserts that because C*, π*, f_e*, and f_d* are optimal for Eq. (12) with respect to the empirical measure P_N^x, the assignment-discrepancy term ℓ(f_d*(\bar f_e*(x)), f_y(x)) − ℓ(\bar f_y(x), f_y(x)) equals zero for every x with P_N(x) > 0. This is false: optimality of an expectation does not force a pointwise nonnegative discrepancy to vanish at every training point; the discrepancy can be positive at some points and zero at others while the empirical expectation is minimized. The subsequent step (1), the reduction of L_A(\bar f_y, P_N^x) to a product-codebook concentration problem, and the final bound ϵ = LM sqrt(log 1/δ + log|F_e| + 1)/sqrt(N) all depend on this pointwise-zero claim. Without it, Corollary 3.3 and the Ω(M^2) sample-complexity conclusion do not follow from the supplied proof.
  3. [§6.2, Lemma 6.4, proof of parts (2)-(3)] The monotonicity proof ϵ_M^* ≥ ϵ_{M+1}^* constructs a codebook C_{M+1} = C_M^* ∪ {c_{M+1}} and claims that for all x ≠ x_max the quantized loss is unchanged because f_e(x) = f_e^*(x). This ignores that adding a codeword changes the Voronoi cells of the quantization operator: Q_{C_{M+1}}(f_e^*(x)) need not equal Q_{C_M^*}(f_e^*(x)) for points whose nearest codeword becomes c_{M+1}. The equality of the integrals over X \ {x_max} is therefore unjustified. The monotonicity statement may be true, but this proof does not establish it.
minor comments (4)
  1. [§6.1] The proof of Theorem 6.2 refers to 'the OP in (18)', but the appendix section only defines Eqs. (12) and (13); the equation numbering should be corrected throughout the appendix.
  2. [§6.2, Proposition 6.3] The constraint in Eqs. (18)–(20) is written as '\bar f_e#f_y#P_x = P_{c,π}', which is dimensionally inconsistent; it should read '\bar f_e#P_x = P_{c,π}'.
  3. [§6.2, Lemma 6.4] The lemma statement contains typos: 'C* = {c*_m}_M_{k=1}' should have index m = 1, and 'cσ(n)' in the optimization problem should be 'cσ(x)'.
  4. [Figure 2 and Section 4.1] The caption of Figure 2 contains an incomplete clause ('for sample , the decoded centroid'), and several sentences in Sections 3.1.2 and 4.1 are ungrammatical; a careful copyedit is needed.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the flagged issues are proof gaps, not reductions by construction.

full rationale

Applying the Pith criteria, I find no circular step in the paper's derivation chain. The self-citations (Vuong et al. 2023; Zheng et al. 2022) appear only in the introductory list of VQ-VAE-style variants and are not load-bearing for Theorem 3.1, Lemma 3.2, or the FDANN objective. Theorem 3.1 is an attempted mathematical derivation rather than a definitional identification: Eq. (4) is a constrained optimization problem and Eq. (5) is an unconstrained penalized problem, and the paper supplies a proof. The appendix proof of Theorem 3.1 (Section 6.1) appears to analyze feasible solutions satisfying the exact constraint \bar f_e#P_x=P_{c,\pi}, whereas Eq. (5) only penalizes W_{d_z} with finite \lambda; if correct, this is an unsupported step or proof gap, not a self-definitional reduction, because the theorem states a substantive equivalence between different optimization problems. Similarly, Corollary 6.6 (Section 6.3) assumes that the assignment-discrepancy term vanishes on the empirical support for the optimal solution; this is a missing justification in the sample-complexity argument, not a fitted parameter renamed as a prediction or a quantity defined to equal the claimed bound. The empirical results are benchmarked externally: DG experiments use DomainBed datasets (VLCS, PACS, Office-Home) with ERM/DANN/CDANN baselines, and RL experiments compare against the continuous Markov-abstraction baseline of Allen et al. (2021). These external benchmarks give the empirical claims independent content. Hence the paper has no significant circularity, only a minor non-load-bearing self-citation, and the two weaknesses identified by the reviewer are correctness/rigor concerns rather than circular reasoning.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; the codebook C and the discrete distribution P_{c,π} are standard constructs. The main load-bearing assumptions are the infinite-capacity domain, the bounded-metric loss condition, and the unjustified pointwise vanishing of the assignment-discrepancy loss in the sample-complexity proof.

free parameters (2)
  • λ (Wasserstein regularization coefficient) = RL: 100; DG: searched over {0.01, 0.1, 1.0}
    Controls the strength of the codebook-alignment penalty in Eq. (5) and Eq. (11); selected by hand or grid search per task.
  • Codebook size M = RL: {50, 100, 500}; DG: fixed at 16×K
    The central knob of the framework; chosen per experiment, and the trade-off analysis is about varying M.
assumptions (4)
  • domain assumption The family of all measurable functions (infinite capacity) is the search space for f_e and f_d in Theorem 3.1
    Invoked in Theorem 3.1 and acknowledged in the main text as impractical; without it the claimed OP equivalence has no basis.
  • domain assumption The loss ℓ_y is a proper metric and is upper-bounded by a positive constant L
    Assumed in Corollary 3.3 to bound the assignment-discrepancy concentration; restricts applicability to bounded metric losses.
  • ad hoc to paper For the empirical optimal solution, the assignment-discrepancy loss vanishes pointwise on the training set: ℓ(f_d*(f̄_e*(x)), f_y(x)) − ℓ(f̄_y(x), f_y(x)) = 0 whenever P_N(x) > 0
    This is the load-bearing premise in the proof of Corollary 6.6; it does not follow from empirical optimality of the OP and is false in general.
  • standard math Standard results: optimal transport duality, k-means uniform deviation bounds (Bachem et al., 2017), and concentration for discrete distributions (Hsu et al., 2012)
    These are cited and used as building blocks for Lemma 3.2 and Corollary 3.3; they are not re-derived.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Driven Discrete Representation Learning." pith.science (2026). https://pith.science/paper/NZK3O6SW

@misc{pith2026250611511,
  author       = {Pith},
  title        = {Pith review of: Task-Driven Discrete Representation Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NZK3O6SW}},
  note         = {Machine review of arXiv:2506.11511}
}
read the original abstract

In recent years, deep discrete representation learning (DRL) has achieved significant success across various domains. Most DRL frameworks (e.g., the widely used VQ-VAE and its variants) have primarily focused on generative settings, where the quality of a representation is implicitly gauged by the fidelity of its generation. In fact, the goodness of a discrete representation remain ambiguously defined across the literature. In this work, we adopt a practical approach that examines DRL from a task-driven perspective. We propose a unified framework that explores the usefulness of discrete features in relation to downstream tasks, with generation naturally viewed as one possible application. In this context, the properties of discrete representations as well as the way they benefit certain tasks are also relatively understudied. We therefore provide an additional theoretical analysis of the trade-off between representational capacity and sample complexity, shedding light on how discrete representation utilization impacts task performance. Finally, we demonstrate the flexibility and effectiveness of our framework across diverse applications.

Figures

Figures reproduced from arXiv: 2506.11511 by the authors.

Figure 1
Figure 1. Graphical model: (a) continuous representa [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Mismatching between the nearest assign￾ments in the latent and output spaces. Assignment Discrepancy Loss: It can be seen that ¯f ∗ e (x) is the nearest assignment on latent space while ¯fy(x) = f ∗ d (argmin c ′′∈C ℓy(x, f ∗ d (c ′′))) is the near￾est assignment on output space. The second term i.e., the discrepancy between ℓ [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. (Top row) The visualization true underly￾ing state on the left, the observations which are noisy version of true state in the middle and the task of RL agent i.e., moving from random point to goal (red point) in the right. (From second row) The visual￾ization of the learned latent codebook; from second top to bottom using 3k, 10k, and 30k samples for train￾ing; from left to right: discrete representation with the co… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: b), the discrete-state representation outper [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 4
Figure 4. Figure 4: Results on 10 × 10 gridworld environment, tested with different codebook sizes and training samples. 4.2 Domain Generalization Setting. Given E training source domains {P e (x, y)} E e=1. The objective of DG is to exploit the ‘commonalities’ present in the training dom…
Figure 5
Figure 5. Figure 5: Fine-Grained Discrete Representation: Z¯2 corresponds to class-level weights, capturing informa￾tion at the class level, while Z¯1 corresponds to fine￾grained discrete representations, characterizing data at a more detailed, modality-specific level. the vector correspo…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

59 extracted references · 38 canonical work pages

  1. [1]

    Reinforcement learning: Theory and algorithms

    Alekh Agarwal, Nan Jiang, Sham M Kakade, and Wen Sun. Reinforcement learning: Theory and algorithms. CS Dept., UW Seattle, Seattle, WA, USA, Tech. Rep, 32: 0 96, 2019

    2019

  2. [2]

    Learning markov state abstractions for deep reinforcement learning

    Cameron Allen, Neev Parikh, Omer Gottesman, and George Konidaris. Learning markov state abstractions for deep reinforcement learning. Advances in Neural Information Processing Systems, 34: 0 8229--8241, 2021

    work page 2021

  3. [3]

    Invariant risk minimization, 2020

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization, 2020. URL https://arxiv.org/abs/1907.02893

    arXiv 2020

  4. [4]

    Uniform deviation bounds for k-means clustering

    Olivier Bachem, Mario Lucic, S Hamed Hassani, and Andreas Krause. Uniform deviation bounds for k-means clustering. In International conference on machine learning, pp.\ 283--291. PMLR, 2017

    work page 2017

  5. [5]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In Proceedings of the European conference on computer vision (ECCV), pp.\ 456--473, 2018

    work page 2018

  6. [6]

    Swad: Domain generalization by seeking flat minima

    Junbum Cha, Sanghyuk Chun, Kyungjae Lee, Han-Cheol Cho, Seunghyun Park, Yunsung Lee, and Sungrae Park. Swad: Domain generalization by seeking flat minima. Advances in Neural Information Processing Systems, 34: 0 22405--22418, 2021

    work page 2021

  7. [7]

    Mode regularized generative adversarial networks

    Tong Che, Yanran Li, Athul Paul Jacob, Yoshua Bengio, and Wenjie Li. Mode regularized generative adversarial networks. arXiv preprint arXiv:1612.02136, 2016

    arXiv 2016

  8. [8]

    A simple framework for contrastive learning of visual representations

    Ting Chen, Simon Kornblith, Mohammad Norouzi, and Geoffrey Hinton. A simple framework for contrastive learning of visual representations. In International conference on machine learning, pp.\ 1597--1607. PMLR, 2020

    2020

Show all 59 references
  1. [9]

    Exploring simple siamese representation learning

    Xinlei Chen and Kaiming He. Exploring simple siamese representation learning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 15750--15758, 2021

    2021

  2. [10]

    Jukebox: A generative model for music

    Prafulla Dhariwal, Heewoo Jun, Christine Payne, Jong Wook Kim, Alec Radford, and Ilya Sutskever. Jukebox: A generative model for music. arXiv preprint arXiv:2005.00341, 2020

    2005 arXiv

  3. [11]

    Taming transformers for high-resolution image synthesis

    Patrick Esser, Robin Rombach, and Bjorn Ommer. Taming transformers for high-resolution image synthesis. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 12873--12883, 2021

    2021

  4. [12]

    Domain-adversarial training of neural networks

    Yaroslav Ganin, Evgeniya Ustinova, Hana Ajakan, Pascal Germain, Hugo Larochelle, Fran c ois Laviolette, Mario Marchand, and Victor Lempitsky. Domain-adversarial training of neural networks. The Journal of Machine Learning Research, 17 0 (1): 0 2096--2030, 2016

    2016

  5. [13]

    Domain adaptation with conditional transferable components

    Mingming Gong, Kun Zhang, Tongliang Liu, Dacheng Tao, Clark Glymour, and Bernhard Sch \"o lkopf. Domain adaptation with conditional transferable components. In International conference on machine learning, pp.\ 2839--2848. PMLR, 2016

    2016

  6. [14]

    Generative adversarial nets

    Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial nets. Advances in neural information processing systems, 27, 2014

    2014

  7. [15]

    In search of lost domain generalization

    Ishaan Gulrajani and David Lopez-Paz. In search of lost domain generalization. In International Conference on Learning Representations, 2021

    2021

  8. [16]

    Mastering atari with discrete world models

    Danijar Hafner, Timothy Lillicrap, Mohammad Norouzi, and Jimmy Ba. Mastering atari with discrete world models. arXiv preprint arXiv:2010.02193, 2020

    2010 arXiv

  9. [17]

    A spectral algorithm for learning hidden markov models

    Daniel Hsu, Sham M Kakade, and Tong Zhang. A spectral algorithm for learning hidden markov models. Journal of Computer and System Sciences, 78 0 (5): 0 1460--1480, 2012

    2012

  10. [18]

    Learning discrete representations via information maximizing self-augmented training

    Weihua Hu, Takeru Miyato, Seiya Tokui, Eiichi Matsumoto, and Masashi Sugiyama. Learning discrete representations via information maximizing self-augmented training. In International conference on machine learning, pp.\ 1558--1567. PMLR, 2017

    2017

  11. [19]

    Representation learning in deep rl via discrete information bottleneck

    Riashat Islam, Hongyu Zang, Manan Tomar, Aniket Didolkar, Md Mofijul Islam, Samin Yeasar Arnob, Tariq Iqbal, Xin Li, Anirudh Goyal, Nicolas Heess, et al. Representation learning in deep rl via discrete information bottleneck. arXiv preprint arXiv:2212.13835, 2022

    2022 arXiv

  12. [20]

    Video pixel networks

    Nal Kalchbrenner, A \"a ron Oord, Karen Simonyan, Ivo Danihelka, Oriol Vinyals, Alex Graves, and Koray Kavukcuoglu. Video pixel networks. In International Conference on Machine Learning, pp.\ 1771--1779. PMLR, 2017

    2017

  13. [21]

    On a problem of monge

    LK Kantorovich. On a problem of monge. Journal of Mathematical Sciences, 133 0 (4), 2006

    2006

  14. [22]

    L-verse: Bidirectional generation between image and text

    Taehoon Kim, Gwangmo Song, Sihaeng Lee, Sangyun Kim, Yewon Seo, Soonyoung Lee, Seung Hwan Kim, Honglak Lee, and Kyunghoon Bae. L-verse: Bidirectional generation between image and text. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 1...

    2022

  15. [23]

    Auto-encoding variational bayes

    Diederik P Kingma and Max Welling. Auto-encoding variational bayes. arXiv preprint arXiv:1312.6114, 2013

    2013 arXiv

  16. [24]

    Improved variational inference with inverse autoregressive flow

    Durk P Kingma, Tim Salimans, Rafal Jozefowicz, Xi Chen, Ilya Sutskever, and Max Welling. Improved variational inference with inverse autoregressive flow. Advances in neural information processing systems, 29, 2016

    2016

  17. [25]

    Out-of-distribution generalization via risk extrapolation (rex)

    David Krueger, Ethan Caballero, Joern-Henrik Jacobsen, Amy Zhang, Jonathan Binas, Dinghuai Zhang, Remi Le Priol, and Aaron Courville. Out-of-distribution generalization via risk extrapolation (rex). In International Conference on Machine Learning, pp.\ 5815--5826. PMLR, 2021

    2021

  18. [26]

    Deeper, broader and artier domain generalization

    Da Li, Yongxin Yang, Yi-Zhe Song, and Timothy M Hospedales. Deeper, broader and artier domain generalization. In Proceedings of the IEEE international conference on computer vision, pp.\ 5542--5550, 2017

    2017

  19. [27]

    Domain generalization with adversarial feature learning

    Haoliang Li, Sinno Jialin Pan, Shiqi Wang, and Alex C Kot. Domain generalization with adversarial feature learning. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 5400--5409, 2018 a

    2018

  20. [28]

    Domain generalization via conditional invariant representations

    Ya Li, Mingming Gong, Xinmei Tian, Tongliang Liu, and Dacheng Tao. Domain generalization via conditional invariant representations. In Proceedings of the AAAI conference on artificial intelligence, volume 32, 2018 b

    2018

  21. [29]

    Deep domain generalization via conditional invariant adversarial networks

    Ya Li, Xinmei Tian, Mingming Gong, Yajing Liu, Tongliang Liu, Kun Zhang, and Dacheng Tao. Deep domain generalization via conditional invariant adversarial networks. In Proceedings of the European conference on computer vision (ECCV), pp.\ 624--639, 2018 c

    2018

  22. [30]

    Frequency-aware self-supervised long-tailed learning

    Ci-Siang Lin, Min-Hung Chen, and Yu-Chiang Frank Wang. Frequency-aware self-supervised long-tailed learning. In Proceedings of the IEEE/CVF International Conference on Computer Vision, pp.\ 963--972, 2023

    2023

  23. [31]

    Adaptive discrete communication bottlenecks with dynamic vector quantization

    Dianbo Liu, Alex Lamb, Xu Ji, Pascal Notsawo, Mike Mozer, Yoshua Bengio, and Kenji Kawaguchi. Adaptive discrete communication bottlenecks with dynamic vector quantization. arXiv preprint arXiv:2202.01334, 2022

    2022 arXiv

  24. [32]

    Conditional adversarial domain adaptation

    Mingsheng Long, Zhangjie Cao, Jianmin Wang, and Michael I Jordan. Conditional adversarial domain adaptation. arXiv preprint arXiv:1705.10667, 2017

    2017 arXiv

  25. [33]

    How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450, 2022

    Yifei Ming, Yiyou Sun, Ousmane Dia, and Yixuan Li. How to exploit hyperspherical embeddings for out-of-distribution detection? arXiv preprint arXiv:2203.04450, 2022

    2022 arXiv

  26. [34]

    Representation learning via invariant causal mechanisms

    Jovana Mitrovic, Brian McWilliams, Jacob Walker, Lars Buesing, and Charles Blundell. Representation learning via invariant causal mechanisms. arXiv preprint arXiv:2010.07922, 2020

    2010 arXiv

  27. [35]

    Wavenet: A generative model for raw audio

    Aaron van den Oord, Sander Dieleman, Heiga Zen, Karen Simonyan, Oriol Vinyals, Alex Graves, Nal Kalchbrenner, Andrew Senior, and Koray Kavukcuoglu. Wavenet: A generative model for raw audio. arXiv preprint arXiv:1609.03499, 2016

    2016 arXiv

  28. [36]

    Vector quantized models for planning

    Sherjil Ozair, Yazhe Li, Ali Razavi, Ioannis Antonoglou, Aaron Van Den Oord, and Oriol Vinyals. Vector quantized models for planning. In international conference on machine learning, pp.\ 8302--8313. PMLR, 2021

    2021

  29. [37]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 2536--2544, 2016

    2016

  30. [38]

    Moment matching for multi-source domain adaptation

    Xingchao Peng, Qinxun Bai, Xide Xia, Zijun Huang, Kate Saenko, and Bo Wang. Moment matching for multi-source domain adaptation. In Proceedings of the IEEE/CVF international conference on computer vision, pp.\ 1406--1415, 2019

    2019

  31. [39]

    Zero-shot text-to-image generation

    Aditya Ramesh, Mikhail Pavlov, Gabriel Goh, Scott Gray, Chelsea Voss, Alec Radford, Mark Chen, and Ilya Sutskever. Zero-shot text-to-image generation. In International conference on machine learning, pp.\ 8821--8831. Pmlr, 2021

    2021

  32. [40]

    Generating diverse high-fidelity images with vq-vae-2

    Ali Razavi, Aaron Van den Oord, and Oriol Vinyals. Generating diverse high-fidelity images with vq-vae-2. Advances in neural information processing systems, 32, 2019

    2019

  33. [41]

    Parallel multiscale autoregressive density estimation

    Scott Reed, A \"a ron Oord, Nal Kalchbrenner, Sergio G \'o mez Colmenarejo, Ziyu Wang, Yutian Chen, Dan Belov, and Nando Freitas. Parallel multiscale autoregressive density estimation. In International conference on machine learning, pp.\ 2912--2921. PMLR, 2017

    2017

  34. [42]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj \"o rn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 10684--10695, 2022

    2022

  35. [43]

    Theory and experiments on vector quantized autoencoders

    Aurko Roy, Ashish Vaswani, Arvind Neelakantan, and Niki Parmar. Theory and experiments on vector quantized autoencoders. arXiv preprint arXiv:1805.11063, 2018

    2018 arXiv

  36. [44]

    Optimal transport for applied mathematicians

    Filippo Santambrogio. Optimal transport for applied mathematicians. Birk \"a user, NY , 55 0 (58-63): 0 94, 2015

    2015

  37. [45]

    Domain adaptation with conditional distribution matching and generalized label shift

    Remi Tachet des Combes, Han Zhao, Yu-Xiang Wang, and Geoffrey J Gordon. Domain adaptation with conditional distribution matching and generalized label shift. Advances in Neural Information Processing Systems, 33: 0 19276--19289, 2020

    2020

  38. [46]

    SQ - VAE : Variational B ayes on discrete representation with self-annealed stochastic quantization

    Yuhta Takida, Takashi Shibuya, Weihsiang Liao, Chieh-Hsin Lai, Junki Ohmura, Toshimitsu Uesaka, Naoki Murata, Shusuke Takahashi, Toshiyuki Kumakura, and Yuki Mitsufuji. SQ - VAE : Variational B ayes on discrete representation with self-annealed stochastic quantization. In Kama...

    2022

  39. [47]

    Moment-based uniform deviation bounds for k -means and friends

    Matus J Telgarsky and Sanjoy Dasgupta. Moment-based uniform deviation bounds for k -means and friends. Advances in Neural Information Processing Systems, 26, 2013

    2013

  40. [48]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. In CVPR 2011, pp.\ 1521--1528. IEEE, 2011

    2011

  41. [49]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, et al. Neural discrete representation learning. Advances in neural information processing systems, 30, 2017

    2017

  42. [50]

    Deep hashing network for unsupervised domain adaptation

    Hemanth Venkateswara, Jose Eusebio, Shayok Chakraborty, and Sethuraman Panchanathan. Deep hashing network for unsupervised domain adaptation. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 5018--5027, 2017

    2017

  43. [51]

    Vector quantized wasserstein auto-encoder

    Tung-Long Vuong, Trung Le, He Zhao, Chuanxia Zheng, Mehrtash Harandi, Jianfei Cai, and Dinh Phung. Vector quantized wasserstein auto-encoder. arXiv preprint arXiv:2302.05917, 2023

    2023 arXiv

  44. [52]

    Out-of-distribution generalization with causal invariant transformations

    Ruoyu Wang, Mingyang Yi, Zhitang Chen, and Shengyu Zhu. Out-of-distribution generalization with causal invariant transformations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 375--385, 2022

    2022

  45. [53]

    Hierarchical quantized autoencoders

    Will Williams, Sam Ringer, Tom Ash, David MacLeod, Jamie Dougherty, and John Hughes. Hierarchical quantized autoencoders. Advances in Neural Information Processing Systems, 33: 0 4524--4535, 2020

    2020

  46. [54]

    Pcl: Proxy-based contrastive learning for domain generalization

    Xufeng Yao, Yang Bai, Xinyun Zhang, Yuechen Zhang, Qi Sun, Ran Chen, Ruiyu Li, and Bei Yu. Pcl: Proxy-based contrastive learning for domain generalization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 7097--7107, 2022

    2022

  47. [55]

    Reinforcement learning with prototypical representations

    Denis Yarats, Rob Fergus, Alessandro Lazaric, and Lerrel Pinto. Reinforcement learning with prototypical representations. In International Conference on Machine Learning, pp.\ 11920--11931. PMLR, 2021

    2021

  48. [56]

    Ptarl: Prototype-based tabular representation learning via space calibration

    Hangting Ye, Wei Fan, Xiaozhuang Song, Shun Zheng, He Zhao, Dandan Guo, and Yi Chang. Ptarl: Prototype-based tabular representation learning via space calibration. arXiv preprint arXiv:2407.05364, 2024

    2024 arXiv

  49. [57]

    Vector-quantized image modeling with improved vqgan

    Jiahui Yu, Xin Li, Jing Yu Koh, Han Zhang, Ruoming Pang, James Qin, Alexander Ku, Yuanzhong Xu, Jason Baldridge, and Yonghui Wu. Vector-quantized image modeling with improved vqgan. In International Conference on Learning Representations, 2021

    2021

  50. [58]

    Movq: Modulating quantized vectors for high-fidelity image generation

    Chuanxia Zheng, Long Tung Vuong, Jianfei Cai, and Dinh Phung. Movq: Modulating quantized vectors for high-fidelity image generation. Advances in Neural Information Processing Systems, 35, 2022

    2022

  51. [59]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.