Pith. sign in

REVIEW 3 major objections 6 minor 45 references

Disentangling Latent Embeddings with Sparse Linear Concept Subspaces (SLiCS)

T0 review · 3 major / 6 minor · reviewed 2026-08-15 · deepseek-v4-flash

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

Pith's one-line read The paper claims that a dense image embedding can be approximated as a sum of concept-specific positive-cone components, and that retrieving with one such component is more precise than retrieving with the whole embedding.

desk verdict Useful retrieval technique, but the stated disentanglement claim is not supported by the headline experiments. read the letter →

arxiv 2508.20322 v1 pith:CPIZIJVO submitted 2025-08-27 cs.CV

Zhi Li , Hau Phan , Matthew Emigh , Austin J. Brockmeier This is my paper · ORCID
classification cs.CV
keywords conceptdisentanglementdictionarylearningimageretrievalCLIPembeddingsDINOv2TiToknon-negativeleastsquareszero-shotlabels
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

SLiCS argues that a dense vision-language embedding is not an unanalyzable whole: it can be written, approximately, as the sum of one component per concept, each component lying in a positive cone spanned by atoms reserved for that concept. The paper shows how to learn those concept cones from multi-label supervision, or from pseudo-labels produced by zero-shot classification of concept text embeddings, using a dictionary-learning routine that alternates non-negative least-squares coefficient estimation with rank-1 singular-value atom updates. If the claim holds, a user can decide which aspects of a query image should drive retrieval, because only the component for the chosen concept is used for ranking. The paper reports that concept-filtered retrieval sharply improves mAP@20 over whole-embedding retrieval on CLIP, DINOv2, and TiTok embeddings, and that the learned cones carry consistent word descriptions linking the decomposition back to text space.

What carries the argument

The machinery is a group-structured semi-non-negative dictionary: a global dictionary $B$ is partitioned into blocks $B_j$, each block spanning the positive cone $\{B_j a : a \ge 0\}$ for one concept, and coefficients are constrained to be non-negative and group-sparse according to the label vector. The optimization alternates a non-negative least-squares step for the coefficients with a K-SVD-style atom update in which each atom is refreshed from the rank-1 truncated SVD of the residual restricted to samples where that atom is active; a majority-sign rule for the atom's polarity is proved optimal, and thresholding restores the non-negativity of the coefficients. The component used for retrieval is the partial projection $\tilde{v}^\star_j = B_j \alpha_j$ obtained by solving the non-negative least-squares problem $\min_{a \ge 0} \|x^\star - B_j a\|^2$, which keeps the component close to the query while masking out other concepts.

What would settle it

Build paired queries that are the same scene with and without one added concept (for example, a person in an otherwise identical room), then compare the difference of the two SLiCS decompositions to the person concept's component; if adding the person substantially changes the other concepts' components or the difference is not close to the person cone, the additive model fails.

Watch

Extended reading notes

Core claim

The paper's central claim is the linear synthesis model $x^\star \approx \sum_{j=1}^S B_j \alpha_j$, in which an embedding is decomposed into $S$ concept components, with $B_j$ the dictionary of atoms for concept $j$ and $\alpha_j$ a non-negative coefficient vector that is zero when the concept is absent. The discovery is that this group-structured, non-negative decomposition can be learned and is useful: the proposed SLiCS algorithm initializes each concept's atoms by truncated SVD on embeddings that contain that concept, then alternates non-negative least-squares coefficient updates and K-SVD-style atom updates, using a provably optimal sign choice followed by thresholding to keep coefficients non-negative. Given the learned dictionaries, approximating a query embedding with only one concept's cone yields a component that isolates that concept, and retrieval ranked by cosine similarity to this component outperforms retrieval with the whole embedding for concept-specific queries.

Load-bearing premise

The load-bearing premise is that a dense image embedding is well approximated by a non-negative sum of concept-specific components; if concepts compose non-additively, interact during encoding, or leave negative semantic traces, the learned components are fitting artifacts rather than true concept separation.

Editorial extensions

If this is right

  • Concept-filtered retrieval becomes a practical query mode: ranking candidates by cosine similarity to the query's concept component rather than to the whole embedding, with reported mAP@20 gains such as 0.895 versus 0.728 on MIRFlickr25K and 0.929 versus 0.791 on MS COCO for CLIP ViT-B/32.
  • Training labels can be replaced by pseudo-labels: zero-shot classification of concept text embeddings against training images supplies the group-sparsity pattern, so the same dictionary learning runs without manual annotation beyond a list of concept names.
  • The same decomposition applies across embedding families, including contrastive CLIP, self-supervised DINOv2, and highly compressed TiTok token embeddings, with DINOv2 yielding the highest filtered retrieval in the reported experiments.
  • Each concept cone can be assigned human-readable words: the concept atoms that best reconstruct text embeddings of frequent words yield subspace captions consistent across supervised and unsupervised training.
  • Disentangled concept components can drive conditional generation by converting a component into a prompt for a latent diffusion model after an orthogonal Procrustes alignment of the image and text spaces.

Reading between the lines

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

  • Editorial inference: the additive cone model implies a compositional test: if concept components truly sum, removing one concept's component from an embedding should erase that concept from retrieval while leaving other concepts intact, a property that is implied but not directly measured in the paper.
  • Editorial inference: because the non-negative cone can only add presence, concept-filtered retrieval for negated queries (images without a concept) would need an explicit complement subspace, which is a natural extension not addressed by SLiCS as described.
  • Editorial inference: applying SLiCS to multiple encoders for the same concept could expose whether different vision models build aligned or incompatible concept geometry, turning the method into a comparative probe of embedding spaces.
  • Editorial inference: the near-orthogonality of atoms within a concept group and the group-structured co-occurrence pattern suggest the learned cones behave like simplex codes; if that structure is consistent, concept arithmetic in the embedding space should support editing operations such as swapping one concept for another.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper proposes SLiCS (Sparse Linear Concept Subspaces), a supervised dictionary-learning method that decomposes a dense image embedding into a sum of concept-specific components, each a non-negative combination of atoms associated with one concept. The method is applied to CLIP, TiTok, and DINOv2 embeddings, and is evaluated primarily through concept-filtered image retrieval (mAP@20), qualitative word-caption interpretation of learned subspaces, and image-to-prompt generation. An unsupervised variant uses CLIP zero-shot classification to generate pseudo-labels. The central claims are that the learned decomposition is semantically meaningful, that concept-filtered retrieval improves over unfiltered CLIP and SpLiCE baselines, and that the proposed alternating optimization has guaranteed convergence.

Significance. If the linear additive model of Eq. (1) is valid, SLiCS would provide a practical tool for interpreting dense visual embeddings and for concept-focused retrieval without retraining the backbone. The paper's strengths are its breadth of empirical evaluation across three embedding families (CLIP, TiTok, DINOv2), consistent mAP@20 improvements over UF-CLIP, SpLiCE, and F-SpLiCE, the inclusion of a rank-1 sign-optimality proof in §3.3.1, and the practical extension to product-quantized candidate pools for TiTok. However, the current experiments mostly validate a weaker statement: per-concept positive-cone projections are useful retrieval filters. The full additive-decomposition claim, including support recovery and cross-concept isolation, is not directly tested, and the convergence guarantee in §3.3.2 does not apply to the algorithm variant that is actually run.

major comments (3)
  1. [§3.2 and §3.5, Table 2] The headline retrieval results are computed with the partial projection ṽ*_j from Eq. (3), not with the joint additive component v*_j = B_j α_j from Eq. (2). Because the concept cones can overlap, ṽ*_j is the point in one cone closest to the whole embedding and can carry information about concepts other than j; Fig. 7 confirms that this partial projection outperforms the joint component for retrieval. Consequently, the mAP gains reported in Tables 2–4 are consistent with the much weaker claim that supervised per-concept positive-cone projections are effective retrieval filters, and they do not establish Eq. (1) as a faithful decomposition, nor do they establish that α_j = 0 iff concept j is absent. Please report retrieval using v*_j from Eq. (2), report reconstruction error of Eq. (1) on held-out data, and evaluate support recovery (active/inactive concept classification) against ground-truth labels.
  2. [§3.3.2 and Abstract] The claimed convergence guarantee applies only to the cyclic-minimizer variant that includes the error-reduction check in Eq. (13), but the text immediately states: 'In practice, we let the algorithm proceed with an atom update without checking if it will increase the reconstruction error.' Therefore the algorithm actually run, and for which all experimental results are reported, has no proven convergence guarantee. The abstract's statement that the dictionary is 'optimized through a novel alternating optimization with guaranteed convergence' is thus an overstatement. Please either run the version with the check and report the effect on accuracy, or revise the convergence claim to describe the checked variant and provide empirical convergence evidence for the unchecked variant.
  3. [§3.1–§3.2 and §4.4] The load-bearing premise of the paper is the linear additive synthesis model x* ≈ Σ_j B_j α_j with non-negative group-sparse coefficients. This assumption is not validated directly anywhere in the manuscript. No quantitative reconstruction error for Eq. (1) is reported for any embedding, no comparison is made against unconstrained linear subspaces or a simple per-concept non-negative projection baseline of matched capacity, and no test is provided showing that the estimated α_j is zero exactly when the ground-truth label is absent. Without such evidence, the 'disentanglement' interpretation of the learned components remains an artifact hypothesis. Please add these direct measurements and ablations, or substantially soften the disentanglement claims.
minor comments (6)
  1. [Algorithm 1] In the atom-update step, the support set is defined as L_m ← {l ∈ {1,...,M} : A_ml ≠ 0}; this should be over training samples, i.e., l ∈ {1,...,N}, not over atoms.
  2. [§5] The dataset name is misspelled as 'MIRFlikr25K' in the conclusion; it should be 'MIRFlickr25K'.
  3. [Figure 4] The caption of Fig. 4 lists panels (a), (c), and (d) with d0 = 2, d0 = 5, and d0 = 20, but there is no panel (b) described, and the text refers to 'Fig. 4(b)' in §4.2. Please fix the panel labels and the corresponding in-text references.
  4. [§4.1] The sentence '6 concepts have have finer-grained labels' contains a duplicated word and should read '6 concepts have finer-grained labels.'
  5. [§3.4] The 'unsupervised' variant requires the set of concept words and the expected number of active concepts S̃ to be known in advance, and it uses CLIP text embeddings to generate pseudo-labels. This is more accurately described as zero-shot-supervised or pseudo-supervised; the current terminology may overstate the degree of unsupervised learning.
  6. [Table 1] Some unsupervised word captions are clearly not semantically consistent with their subspace (e.g., 'aircraft', 'plane', and 'airplanes' assigned to 'electronic' and 'appliance'). The text mentions this briefly but could analyze it more explicitly as evidence about pseudo-label noise and its effects on disentanglement quality.

Circularity Check

1 steps flagged · score 2.0 of 10

Partial self-definition in the unsupervised variant; the supervised retrieval evaluation is otherwise independent of the model inputs.

  1. self definitional [Section 3.4 (Unsupervised Dictionary Learning via Zero-shot Concept Classification), used with Eq. (4); Section 4.4 sets tilde-S from ground-truth label counts.]
    "we also implement an essentially unsupervised version by exploiting the image-text alignment of CLIP if the set of concepts and the expected number of active concepts are known. ... The cosine similarity between an image embedding and each text embedding serves as a measurement of how close the word is to the image. ... since there are multiple concepts presented in each image we pick the highest tilde-S concepts as active. Thus, a pseudo-label is formed to enable the unsupervised SLiCS with the same algorithm as the supervised one."

    In U-SLiCS, the multi-labels that supervise dictionary learning (Eq. 4) are generated by ranking CLIP cosine similarities between the same image embeddings and the CLIP text embeddings of the concept words, then taking the top tilde-S. The learned concept subspaces are therefore, by construction, directions in CLIP space that best reconstruct the very CLIP embeddings that CLIP's own text similarity already flagged for that concept; the 'unsupervised' discovery of concepts partly restates the pseudo-label construction. This is compounded by setting tilde-S from the dataset's ground-truth average number of concepts per image (2.467 for MIRFlickr25K and 2.303 for MS COCO), leaking label statistics into the 'unsupervised' pipeline.

full rationale

No significant circularity is found in the central derivation or in the supervised evaluation. The dictionary is fit on training images with either external labels or pseudo-labels, and concept-filtered retrieval is scored against held-out dataset ground-truth labels on a disjoint query/candidate split, so the reported mAP gains are not forced by the fitting procedure. The linear additive model (Eq. 1) is an explicit assumption that is tested indirectly through retrieval, not an input masquerading as a prediction. The paper's own admission that the convergence guarantee in Sec. 3.3.2 requires an error-reduction check that the authors skip in practice is a rigor/evidence gap, not a circular reduction. Similarly, the fact that the main retrieval uses the partial cone projection v-tilde*_j (Eq. 3) rather than the joint component v*_j (Eq. 2) is an overclaim/mismatch between the headline decomposition and the evaluated operation, but it does not make the result equivalent to the method's inputs. The only partial circularity is the unsupervised pseudo-label path described above, which does not determine the central supervised results.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper relies on the linear-additive, non-negative decomposition assumption as its core modeling premise. The only numerical quantities fitted to data are hyperparameters (d0, ~S, T) chosen by validation or dataset statistics; there are no invented physical or geometric entities beyond the learned dictionary atoms, which are standard model parameters.

free parameters (3)
  • d0 (atoms per concept) = Optimal values from validation, e.g., between 5 and 20 (chosen per dataset/model; see Fig. 6)
    Dictionary size per concept is selected on the validation set by finer-grained retrieval performance; it substantially affects the quality of disentanglement and retrieval.
  • ~S (number of active concepts per image, unsupervised SLiCS) = 2 for both MIRFlickr25K and MS COCO
    Set to the average number of concepts per image over the dataset (2.467 for MIRFlickr, 2.303 for COCO), used to generate pseudo-labels via top-~S zero-shot classification.
  • Training iterations T = 10
    Fixed number of dictionary updates in all experiments; not tuned, but a free hyperparameter.
assumptions (5)
  • domain assumption The embedding x* is approximately a non-negative sum of concept-specific components, each lying in a positive cone (Eq. 1).
    This is the linear additive model that the entire SLiCS pipeline assumes, introduced in Section 3.1.
  • domain assumption Non-negative coefficients correspond to semantic presence, and negative coefficients would imply semantic negation.
    Used to justify the non-negativity constraint in Section 3.1; it relies on properties of CLIP cosine similarity.
  • domain assumption CLIP zero-shot classification with the top ~S concept words yields acceptable pseudo-labels for unsupervised training.
    Section 3.4 builds pseudo-labels from the text embeddings of concept words; noisy pseudo-labels are acknowledged but assumed tolerable.
  • domain assumption The orthogonal Procrustes alignment (Section 3.7) maps image-space components into Stable Diffusion's text prompt space accurately enough for generation.
    Used for image-to-prompt visualization; an imperfect alignment could corrupt the generated images.
  • standard math Eckart-Young-Mirsky theorem and NNLS convexity.
    Used in Section 3.3 to justify rank-1 SVD updates and the coefficient optimization; standard results.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Disentangling Latent Embeddings with Sparse Linear Concept Subspaces (SLiCS)." pith.science (2026). https://pith.science/paper/CPIZIJVO

@misc{pith2026250820322,
  author       = {Pith},
  title        = {Pith review of: Disentangling Latent Embeddings with Sparse Linear Concept Subspaces (SLiCS)},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CPIZIJVO}},
  note         = {Machine review of arXiv:2508.20322}
}
read the original abstract

Vision-language co-embedding networks, such as CLIP, provide a latent embedding space with semantic information that is useful for downstream tasks. We hypothesize that the embedding space can be disentangled to separate the information on the content of complex scenes by decomposing the embedding into multiple concept-specific component vectors that lie in different subspaces. We propose a supervised dictionary learning approach to estimate a linear synthesis model consisting of sparse, non-negative combinations of groups of vectors in the dictionary (atoms), whose group-wise activity matches the multi-label information. Each concept-specific component is a non-negative combination of atoms associated to a label. The group-structured dictionary is optimized through a novel alternating optimization with guaranteed convergence. Exploiting the text co-embeddings, we detail how semantically meaningful descriptions can be found based on text embeddings of words best approximated by a concept's group of atoms, and unsupervised dictionary learning can exploit zero-shot classification of training set images using the text embeddings of concept labels to provide instance-wise multi-labels. We show that the disentangled embeddings provided by our sparse linear concept subspaces (SLiCS) enable concept-filtered image retrieval (and conditional generation using image-to-prompt) that is more precise. We also apply SLiCS to highly-compressed autoencoder embeddings from TiTok and the latent embedding from self-supervised DINOv2. Quantitative and qualitative results highlight the improved precision of the concept-filtered image retrieval for all embeddings.

Figures

Figures reproduced from arXiv: 2508.20322 by the authors.

Figure 1
Figure 1. The pipeline of SLiCS-aided concept-filtered image retrieval. SLiCS decomposes a query [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) On the left is an illustration of the “subspace” disentanglement of a three-dimensional [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Generated images from image-to-prompt [Ding et al., 2023] applied to disentanglement in [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Visualization of disentangled subspaces by supervised SLiCS, with points corresponding [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Relationship between the atoms across the concept dictionaries and their coefficients in [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: mAP @ 20 of general label retrieval and sub-label retrieval with various [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: mAP @ 20 of general label retrieval and sub-label retrieval with various [PITH_FULL_IMAGE:figures/full_fig_p012_7.png]
Figure 8
Figure 8. Figure 8: Qualitative results for on CLIP ViT-B/32 embeddings. The concepts within each method are [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: Qualitative results for on TiTok-L-32 embeddings. The concepts within each method are [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]
Figure 10
Figure 10. Figure 10: Qualitative results for on DINOv2 ViT-B/14 embeddings. The concepts within each [PITH_FULL_IMAGE:figures/full_fig_p016_10.png]
Figure 11
Figure 11. Figure 11: Image-to-prompt visualization for SLiCS model. The generated images in each row [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 27 canonical work pages

  1. [1]

    Aharon, M

    M. Aharon, M. Elad, and A. Bruckstein. K-SVD : An algorithm for designing overcomplete dictionaries for sparse representation. IEEE Transactions on Signal Processing, 54 0 (11): 0 4311--4322, 2006. doi:10.1109/TSP.2006.881199

    arXiv 2006

  2. [2]

    Effective conditioned and composed image retrieval combining CLIP -based features

    Alberto Baldrati, Marco Bertini, Tiberio Uricchio, and Alberto Del Bimbo. Effective conditioned and composed image retrieval combining CLIP -based features. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 21434--21442, 2022. doi:10.1109/CVPR52688.2022.02080

    arXiv 2022

  3. [3]

    Calmon, and Himabindu Lakkaraju

    Usha Bhalla, Alex Oesterling, Suraj Srinivas, Flavio P. Calmon, and Himabindu Lakkaraju. Interpreting CLIP with sparse linear concept embeddings ( SpLiCE ). In A. Globerson, L. Mackey, D. Belgrave, A. Fan, U. Paquet, J. Tomczak, and C. Zhang, editors, Advances in Neural Information Processing Systems, volume 37, pages 84298--84328. Curran Associates, Inc....

    work page 2024

  4. [4]

    A fast non-negativity-constrained least squares algorithm

    Rasmus Bro and Sijmen De Jong. A fast non-negativity-constrained least squares algorithm. Journal of Chemometrics, 11 0 (5): 0 393--401, 1997. doi:https://doi.org/10.1002/(SICI)1099-128X(199709/10)11:5<393::AID-CEM483>3.0.CO;2-L. URL https://analyticalsciencejournals.onlinelibrary.wiley.com/doi/abs/10.1002/

    doi:10.1002/(sici)1099-128x(199709/10)11:5 1997

  5. [5]

    Multidimensional independent component analysis

    J-F Cardoso. Multidimensional independent component analysis. In Proceedings of the 1998 IEEE International Conference on Acoustics, Speech and Signal Processing, ICASSP'98 (Cat. No. 98CH36181), volume 4, pages 1941--1944. IEEE, 1998

    work page 1998

  6. [6]

    Unsupervised learning of visual features by contrasting cluster assignments

    Mathilde Caron, Ishan Misra, Julien Mairal, Priya Goyal, Piotr Bojanowski, and Armand Joulin. Unsupervised learning of visual features by contrasting cluster assignments. Advances in neural information processing systems, 33: 0 9912--9924, 2020

    2020

  7. [7]

    Image retrieval for complex queries using knowledge embedding

    Chandramani Chaudhary, Poonam Goyal, Navneet Goyal, and Yi-Ping Phoebe Chen. Image retrieval for complex queries using knowledge embedding. ACM Trans. Multimedia Comput. Commun. Appl., 16 0 (1), mar 2020. ISSN 1551-6857. doi:10.1145/3375786. URL https://doi.org/10.1145/3375786

    work page doi:10.1145/3375786 2020

  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, pages 1597--1607. PMLR, 2020 a

    work page 2020

Show all 45 references
  1. [9]

    Improved baselines with momentum contrastive learning

    Xinlei Chen, Haoqi Fan, Ross Girshick, and Kaiming He. Improved baselines with momentum contrastive learning. arXiv preprint arXiv:2003.04297, 2020 b

    2003 arXiv

  2. [10]

    An empirical study of training self-supervised vision transformers

    Xinlei Chen, Saining Xie, and Kaiming He. An empirical study of training self-supervised vision transformers. In Proceedings of the IEEE/CVF international conference on computer vision, pages 9640--9649, 2021

    2021

  3. [11]

    Embedding arithmetic of multimodal queries for image retrieval

    Guillaume Couairon, Matthijs Douze, Matthieu Cord, and Holger Schwenk. Embedding arithmetic of multimodal queries for image retrieval. In 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), pages 4946--4954, 2022. doi:10.1109/CVPRW56347.2022.00542

    2022

  4. [12]

    Convex and semi-nonnegative matrix factorizations

    Chris HQ Ding, Tao Li, and Michael I Jordan. Convex and semi-nonnegative matrix factorizations. IEEE Transactions on Pattern Analysis and Machine Intelligence, 32 0 (1): 0 45--55, 2008

    2008

  5. [13]

    The CLIP model is secretly an image-to-prompt converter

    Yuxuan Ding, Chunna Tian, Haoxuan Ding, and Lingqiao Liu. The CLIP model is secretly an image-to-prompt converter. In Thirty-seventh Conference on Neural Information Processing Systems, 2023. URL https://openreview.net/forum?id=lHa7gFbmvS

    2023

  6. [14]

    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, pages 12873--12883, 2021

    2021

  7. [15]

    Efros, and Jacob Steinhardt

    Yossi Gandelsman, Alexei A. Efros, and Jacob Steinhardt. Interpreting CLIP 's image representation via text-based decomposition. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=5Ca9sSzuDp

    2024

  8. [16]

    Exact and heuristic algorithms for semi-nonnegative matrix factorization

    Nicolas Gillis and Abhishek Kumar. Exact and heuristic algorithms for semi-nonnegative matrix factorization. SIAM Journal on Matrix Analysis and Applications, 36 0 (4): 0 1404--1424, 2015

    2015

  9. [17]

    Clip won't learn object-attribute binding from natural data and here is why

    Bijay Gurung, David T Hoffmann, and Thomas Brox. Clip won't learn object-attribute binding from natural data and here is why. arXiv preprint arXiv:2507.07985, 2025

    2025 arXiv

  10. [18]

    Noise-contrastive estimation: A new estimation principle for unnormalized statistical models

    Michael Gutmann and Aapo Hyv \"a rinen. Noise-contrastive estimation: A new estimation principle for unnormalized statistical models. In Proceedings of the Thirteenth International Conference on Artificial Intelligence and Statistics, pages 297--304. JMLR Workshop and Conferen...

    2010

  11. [19]

    Ridge regression: Biased estimation for nonorthogonal problems

    Arthur E Hoerl and Robert W Kennard. Ridge regression: Biased estimation for nonorthogonal problems. Technometrics, 12 0 (1): 0 55--67, 1970

    1970

  12. [20]

    Sparse autoencoders find highly interpretable features in language models

    Robert Huben, Hoagy Cunningham, Logan Riggs Smith, Aidan Ewart, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=F76bwRSLeK

    2024

  13. [21]

    Huiskes and Michael S

    Mark J. Huiskes and Michael S. Lew. The MIR Flickr retrieval evaluation. In MIR '08: Proceedings of the 2008 ACM International Conference on Multimedia Information Retrieval, New York, NY, USA, 2008. ACM

    2008

  14. [22]

    Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces

    Aapo Hyv \"a rinen and Patrik Hoyer. Emergence of phase-and shift-invariant features by decomposition of natural images into independent feature subspaces. Neural Computation, 12 0 (7): 0 1705--1720, 2000

    2000

  15. [23]

    Product quantization for nearest neighbor search

    Herve Jegou, Matthijs Douze, and Cordelia Schmid. Product quantization for nearest neighbor search. IEEE Transactions on Pattern Analysis and Machine Intelligence, 33 0 (1): 0 117--128, 2010

    2010

  16. [24]

    Vo, Patrick Labatut, and Piotr Bojanowski

    Cijo Jose, Th \'e o Moutakanni, Dahyun Kang, Federico Baldassarre, Timoth \'e e Darcet, Hu Xu, Shang-Wen Li, Marc Szafraniec, Michael Ramamonjisoa, Maxime Oquab, Oriane Sim'eoni, Huy V. Vo, Patrick Labatut, and Piotr Bojanowski. DINOv2 meets text: A unified framework for image...

    2025

  17. [25]

    InDiReCT : Language-guided zero-shot deep metric learning for images

    Konstantin Kobs, Michael Steininger, and Andreas Hotho. InDiReCT : Language-guided zero-shot deep metric learning for images. In 2023 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV), pages 1063--1072, 2023. doi:10.1109/WACV56688.2023.00112

    2023

  18. [26]

    Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning

    Victor Weixin Liang, Yuhui Zhang, Yongchan Kwon, Serena Yeung, and James Y Zou. Mind the gap: Understanding the modality gap in multi-modal contrastive representation learning. In S. Koyejo, S. Mohamed, A. Agarwal, D. Belgrave, K. Cho, and A. Oh, editors, Advances in Neural In...

    2022

  19. [27]

    Lawrence Zitnick

    Tsung-Yi Lin, Michael Maire, Serge Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll \'a r, and C. Lawrence Zitnick. Microsoft COCO : Common objects in context. In David Fleet, Tomas Pajdla, Bernt Schiele, and Tinne Tuytelaars, editors, Computer Vision -- ECCV 2014...

    2014

  20. [28]

    Image retrieval on real-life images with pre-trained vision-and-language models

    Zheyuan Liu, Cristian Rodriguez-Opazo, Damien Teney, and Stephen Gould. Image retrieval on real-life images with pre-trained vision-and-language models. In 2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2105--2114, 2021. doi:10.1109/ICCV48922.2021.00213

    2021

  21. [29]

    Mallat and Zhifeng Zhang

    S.G. Mallat and Zhifeng Zhang. Matching pursuits with time-frequency dictionaries. IEEE Transactions on Signal Processing, 41 0 (12): 0 3397--3415, 1993. doi:10.1109/78.258082

    1993 doi

  22. [30]

    Dinov2: Learning robust visual features without supervision

    Maxime Oquab, Timoth \'e e Darcet, Th \'e o Moutakanni, Huy Vo, Marc Szafraniec, Vasil Khalidov, Pierre Fernandez, Daniel Haziza, Francisco Massa, Alaaeldin El-Nouby, et al. Dinov2: Learning robust visual features without supervision. arXiv preprint arXiv:2304.07193, 2023

    2023 arXiv

  23. [31]

    Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition

    Yagyensh Chandra Pati, Ramin Rezaiifar, and Perinkulam Sambamurthy Krishnaprasad. Orthogonal matching pursuit: Recursive function approximation with applications to wavelet decomposition. In Proceedings of 27th Asilomar Conference on Signals, Systems and Computers, pages 40--4...

    1993

  24. [32]

    On variational bounds of mutual information

    Ben Poole, Sherjil Ozair, Aaron Van Den Oord, Alex Alemi, and George Tucker. On variational bounds of mutual information. In International Conference on Machine Learning, pages 5171--5180. PMLR, 2019

    2019

  25. [33]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. In Marina Meila and ...

    2021

  26. [34]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Bj\"orn Ommer. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 10684--10695, June 2022

    2022

  27. [35]

    Efficient implementation of the K-SVD algorithm using batch orthogonal matching pursuit

    Ron Rubinstein, Michael Zibulevsky, and Michael Elad. Efficient implementation of the K-SVD algorithm using batch orthogonal matching pursuit. Technical Report CS-2008-08, Technion - Computer Science Department, 2008. URL https://csaws.cs.technion.ac.il/ ronrubin/Publications/...

    2008

  28. [36]

    A generalized solution of the orthogonal procrustes problem

    Peter H Sch \"o nemann. A generalized solution of the orthogonal procrustes problem. Psychometrika, 31 0 (1): 0 1--10, 1966

    1966

  29. [37]

    Non-negative least squares for high-dimensional linear models: Consistency and sparse recovery without regularization

    Martin Slawski and Matthias Hein. Non-negative least squares for high-dimensional linear models: Consistency and sparse recovery without regularization. Electronic Journal of Statistics, 7: 0 3004--3056, 2013

    2013

  30. [38]

    Stoica and Y

    P. Stoica and Y. Selen. Cyclic minimizers, majorization techniques, and the expectation-maximization algorithm: a refresher. IEEE Signal Processing Magazine, 21 0 (1): 0 112--114, 2004. doi:10.1109/MSP.2004.1267055

    2004 arXiv

  31. [39]

    Grouped orthogonal matching pursuit for variable selection and prediction

    Grzegorz Swirszcz, Naoki Abe, and Aurelie C Lozano. Grouped orthogonal matching pursuit for variable selection and prediction. Advances in Neural Information Processing Systems, 22, 2009

    2009

  32. [40]

    Neural discrete representation learning

    Aaron Van Den Oord, Oriol Vinyals, and koray kavukcuoglu. Neural discrete representation learning. Advances in Neural Information Processing Systems, 30, 2017

    2017

  33. [41]

    Visualizing data using t-SNE

    Laurens Van der Maaten and Geoffrey Hinton. Visualizing data using t-SNE . Journal of Machine Learning Research, 9 0 (11), 2008

    2008

  34. [42]

    Take and took, gaggle and goose, book and read: Evaluating the utility of vector differences for lexical relation learning

    Ekaterina Vylomova, Laura Rimell, Trevor Cohn, and Timothy Baldwin. Take and took, gaggle and goose, book and read: Evaluating the utility of vector differences for lexical relation learning. arXiv preprint arXiv:1509.01692, 2015

    2015 arXiv

  35. [43]

    From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions

    Peter Young, Alice Lai, Micah Hodosh, and Julia Hockenmaier. From image descriptions to visual denotations: New similarity metrics for semantic inference over event descriptions. Transactions of the Association for Computational Linguistics, 2: 0 67--78, 2014. doi:10.1162/tacl...

    2014 doi

  36. [44]

    An image is worth 32 tokens for reconstruction and generation

    Qihang Yu, Mark Weber, Xueqing Deng, Xiaohui Shen, Daniel Cremers, and Liang-Chieh Chen. An image is worth 32 tokens for reconstruction and generation. Advances in Neural Information Processing Systems, 37: 0 128940--128966, 2024

    2024

  37. [45]

    Post-hoc concept bottleneck models

    Mert Yuksekgonul, Maggie Wang, and James Zou. Post-hoc concept bottleneck models. In The Eleventh International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=nA5AZ8CEyow

    2023

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.