Pith. sign in

REVIEW 4 major objections 5 minor 1 cited by

From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs

T0 review · 4 major / 5 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 A vision-language model that replaces both convolutions and attention with learned frequency atoms matches larger transformer baselines on captioning and VQA.

desk verdict A promising architecture idea undermined by a wrong complexity analysis and mixed-protocol benchmarks; not ready for peer review. read the letter →

arxiv 2506.18943 v1 pith:IOFN2F3P submitted 2025-06-22 cs.CV

Andrew Kiruluta , Priscilla Burity This is my paper · ORCID
classification cs.CV
keywords spectraldictionaryvision-languagemodeltokenmixingFouriertransformattention-freeconvolution-freeimagecaptioningvisualquestionanswering
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 sets out to show that a spectral dictionary token mixer—representing every image patch and text token as a sparse combination of learnable frequency atoms—can replace both the convolutional vision encoder and the quadratic self-attention used for multimodal fusion. It reports a 1.1-billion-parameter prototype that reaches BLEU-4 39.2, CIDEr 127.5, and SPICE 27.0 on MS-COCO captioning and 50.3 percent accuracy on VQAv2, closing about 85 percent of the gap to a much larger BLIP-2 model. If the claims hold, the result matters because it would be the first vision-language model that is both attention-free and convolution-free at competitive accuracy, with O(L log L) scaling and linear memory that make long multimodal contexts practical.

What carries the argument

The load-bearing object is the spectral dictionary token mixer in Equation (3), an analysis-synthesis operator built from a bank of $K$ complex exponentials whose frequency, amplitude, and phase are learned per layer. The paper treats the analysis step $\Phi^{\dagger} X$ and the synthesis step $\Phi S$ as FFT and inverse FFT, yielding $O(L \log L)$ compute and $O(L)$ memory, and adds a learnable phase-bias tensor $\Delta$ to restore locality. The dictionary is shared across image patches and wordpieces, giving cross-modal alignment in the frequency domain.

What would settle it

Record the learned atom frequencies of a trained SDict-VLM and check whether they coincide with the discrete Fourier transform grid; if any $f_k$ differs from a multiple of $1/L$, compute the actual cost of forming $\Phi^{\dagger} X$ and show it scales as $O(K L d)$ rather than $O(L \log L)$, which would refute the complexity claim while leaving the benchmark scores untouched.

Watch

Extended reading notes

Core claim

The central claim is that a single learnable frequency dictionary can do the work of both spatial convolutions and softmax self-attention in a vision-language model. The paper defines the spectral dictionary operator $\mathrm{SDict}(X)=\mathrm{Re}(\Phi(\Phi^{\dagger} X))$ with atoms $\phi_k[n]=A_k e^{i(2\pi f_k n+\phi_k)}$, analyzes each token sequence into coefficients over $K=128$ atoms, mixes them in the frequency domain, and synthesizes the output; the same dictionary is shared across vision and text streams, with fusion layers using the union of both atom sets. The authors report that SDict-VLM-1.1B matches or outperforms several mid-scale transformer baselines on COCO captioning and VQAv2 while using fewer parameters, less memory, and 2.2x faster inference than one of them, and claim this is the first VLM to remove both convolutions and self-attention.

Load-bearing premise

The compute advantage in the paper rests on treating the learned frequency atoms as if they always form a uniform Fourier grid; if the learned frequencies drift off that grid, the analysis step is not a fast Fourier transform and the advertised O(L log L) cost is not established.

Editorial extensions

If this is right

  • If the central claim is correct, vision-language models no longer need convolutions or self-attention; a single spectral mixer can carry both modalities.
  • The $O(L \log L)$ scaling and linear activation memory imply that contexts of tens of thousands of tokens, such as thousands of image patches plus text, fit on the same GPU budget that would require activation checkpointing for attention.
  • The shared frequency dictionary provides a built-in interpretability route: atom spectra can be inspected to see which frequency bands drive particular answers or caption phrases.
  • Because dictionary width $K$ is independent of sequence length, the same pretrained atoms can be reused across input resolutions without adding weights.
  • The accuracy-compute trade-off is tunable: reducing $K$ from 128 to 64 cuts FLOPs and lowers CIDEr by 7.9 points, while increasing $K$ beyond 128 gives almost no gain.

Reading between the lines

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

  • The paper leaves implicit that its $O(L \log L)$ bound depends on the learned frequencies $f_k$ remaining on the uniform DFT grid; if they drift, the analysis step $\Phi^{\dagger} X$ is no longer a Fourier transform and would cost $O(K L d)$ to evaluate. A direct complexity test with recorded atom frequencies would settle this.
  • A natural testable extension is to turn the fixed dictionary width $K$ into a data-adaptive budget, allocating atoms to high-entropy spectral bands and pruning unused ones; the paper outlines this idea but does not test it.
  • The absence of causal masking in the global FFT suggests the design is better suited to bidirectional captioning and VQA than to autoregressive or streaming generation; a sliding-window or dilated spectral basis would be needed to extend it, and the paper lists this as future work.
  • Spectral dictionaries could be compared against state-space models as alternative attention replacements; the paper mentions this possibility but does not test it.
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

4 major / 5 minor

Summary. The manuscript proposes SDict-VLM, a vision-language model that replaces convolutional vision encoders and softmax self-attention with a learnable spectral dictionary mixer. It claims O(L log L) complexity and linear memory, and reports MS-COCO captioning scores (BLEU-4 39.2, CIDEr 127.5, SPICE 27.0) and VQAv2 accuracy (50.3%) for a 1.1B-parameter model, positioning it as the first VLM to eliminate both convolutions and self-attention while matching mid-scale transformer baselines. The paper includes an architectural formulation, complexity analysis, experiments against several baselines, ablations over dictionary width, and qualitative spectral visualizations.

Significance. If the complexity and accuracy claims held, this would be a notable contribution: a convolution-free and attention-free VLM at the 1.1B scale with competitive results on standard benchmarks, accompanied by a clean spectral-complexity narrative. The paper is clearly organized, contains explicit complexity equations, reports ablations over dictionary width, and includes a candid limitations section. However, the central complexity result depends on an FFT implementation that the operator as defined does not admit, and the benchmark comparisons mix fine-tuned and zero-shot protocols; as written, the efficiency and accuracy claims are not established.

major comments (4)
  1. [3.1, Eq. (4); Sec. 4, Eq. (6)] The claimed O(L log L) complexity does not follow from the operator as defined. With learned f_k, A_k, phi_k and the position-dependent phase-bias tensor Delta in R^{L x K}, Phi in Eq. (2) is a dense L x K matrix, and Phi^dagger is its dense Moore-Penrose inverse; applying Phi^dagger and Phi to d feature channels costs Theta(K L d), not two FFTs. The phase-bias e^{i Delta[n,k]} prevents folding the phase into per-bin terms, so the operation is not a DFT and no FFT implementation exists for this operator. Additionally, the implementation sketched in Fig. 1 (FFT -> multiply by atoms -> IFFT) is a different operator from Eq. (3), yet both are used to justify Eq. (6). Thus the advertised Theta(L log L) compute and linear-memory advantages are unsupported by the paper's own equations.
  2. [7.2, Tables 1-2; 7.1 protocol] The benchmark comparisons are not apples-to-apples. Table 2 explicitly marks Flamingo 9B and BLIP-2 as 0-shot, while SDict-VLM is fine-tuned on train+val; Table 1 similarly places OpenFlamingo-4B (16-shot) next to fine-tuned numbers. The abstract's claim of closing about 85% of the gap to BLIP-2 is therefore computed against zero-shot baselines, and the statement that SDict-VLM 'matches mid-scale transformer baselines' confounds architecture with training protocol. To support the claim, the authors need to compare against baselines fine-tuned under the identical data and evaluation protocol, or report SDict-VLM in the zero-shot setting alongside the zero-shot baselines.
  3. [7.1, 8.2; captioning generation] The captioning protocol appears incompatible with the stated architecture. Section 7.1 says captions are generated with nucleus sampling and length normalization, which presumes an autoregressive sequence model, yet Section 8.2 concedes that SDict's global FFT-iFFT pipeline lacks an inherent causal ordering. The paper does not describe how a noncausal global mixer is made causal during decoding, or how previously generated tokens are fed back into the model. This is not a peripheral limitation but a missing description of the inference procedure for the main benchmark, and it raises doubts about whether the reported MS-COCO captioning results can be obtained from the architecture as specified.
  4. [7.1, 9; reproducibility] No variance is reported and no artifacts are provided. The paper states that numbers are averaged over three random seeds, but Tables 1-3 give only point estimates with no standard deviations, confidence intervals, or per-seed values. The conclusion announces release of all code, checkpoints, and evaluation scripts, but the manuscript contains no URL or artifact identifier. Since the empirical contribution rests on a small number of table entries, the lack of error bars and the absence of any reproducible artifact make the results effectively unverifiable.
minor comments (5)
  1. [Section 7.2, text before Table 1] The sentence claiming 'outperforming PaLI-3 by a non-trivial –2.4 pp margin in BLEU-4' is self-contradictory; Table 1 shows SDict-VLM at BLEU-4 39.2 versus PaLI-3 at 40.5, so the difference is negative and the numerical value does not match the table.
  2. [Section 7.4, energy estimate] The CodeCarbon estimate refers to an 'equivalently-sized ViT-GPT2 hybrid' that is not defined or listed in any table; please specify the exact baseline model, data, hardware, and whether the estimate covers pretraining or fine-tuning.
  3. [Section 2.1 and References] The citation for GPT-4o appears as '[?]' in the text, and references [27] and [28] point to different versions of the same arXiv identifier; please resolve these reference issues.
  4. [Section 3.2 and Fig. 1] The paper repeatedly calls the dictionary 'shared' across modalities, but Figure 1 and Section 3.2 describe vision-specific slices, text-specific slices, and a fusion dictionary initialized from the union of the two; please clarify what exactly is shared.
  5. [Section 4, discussion of Eq. (6)] Equation (4) includes an O(K^3) least-squares-solve term, but the numerical discussion after Eq. (6) only mentions Kd and treats it as negligible for L >= 1024; with K=128, K^3 is about 2.1e6 and is not negligible at L=1024, so the negligibility statement should be revisited or qualified.

Circularity Check

0 steps flagged · score 1.0 of 10

No circular derivation chain: benchmark claims are empirical, and the O(L log L) issue is a derivation gap, not a circular loop.

full rationale

This is an empirical architecture paper. The central claims — competitive MS-COCO and VQAv2 results, first attention-free and convolution-free VLM, and O(L log L) scaling — are not obtained by fitting a parameter to the outcome and then renaming it a prediction. The spectral dictionary operator SDict(X)=Re(Phi(Phi^dagger X)) is defined from learnable atoms, and the reported scores are presented as measured results under a stated protocol (Sections 7.1-7.2). The main legitimacy concern is not circularity: the complexity bound in Eqs. (4) and (6) omits the O(K L d) cost of applying a dense learned dictionary and its pseudo-inverse when frequencies f_k and the phase-bias tensor Delta drift off the DFT grid, so the advertised Theta(L log L) count does not follow from the stated operator. That is an internal correctness/derivation gap, not a case of the conclusion being identical to its inputs. The paper does cite the same authors' prior spectral-dictionary work ([3], [4], [5], [27]) as motivation and precedent, and one reference is left as an unresolved '[?]' in Section 2.2; however, none of these citations defines the present benchmark numbers or reduces the VLM claim to the cited papers. The architecture and experiments stand on their own reported measurements, so no specific reduction by construction or fitted-input-as-prediction can be exhibited. Score 1 reflects the minor, non-load-bearing self-citation reliance rather than any circular loop.

Assumptions & free parameters 4 free parameters · 4 assumptions · 1 invented entities

The central argument rests on the FFT implementability of the learned-atom projection, the adequacy of a rank-128 linear mixer, and the comparability of benchmark tables. No independent physical entities are introduced. All model parameters and the dictionary width K are fitted, making the empirical claim hard to separate from model selection.

free parameters (4)
  • Dictionary width K = 128 (default; ablated over 64, 128, 256)
    Chosen by validation performance; Table 3 shows CIDEr 119.6 at K=64 versus 127.5 at K=128. This is a model-defining hyperparameter.
  • Learned atom parameters (A_k, f_k, phi_k) per layer = Trained values not reported
    Each of K frequencies per SDict layer has learnable amplitude, frequency, and phase; these are fitted on LAION, COCO, and VQAv2 and form the core of the mixer.
  • Phase-bias tensor Delta (L x K) = Learned; not reported
    Introduced in Section 3.1 to restore locality; the ablation shows freezing it to sinusoidal codes costs 2.3 CIDEr points, so it contributes to the empirical claim.
  • Number of SDict layers Nv, Nl, Nm and hidden width d = 24 layers, d=1024, per-branch layer counts unspecified
    Architecture hyperparameters; no sensitivity analysis is reported, and capacity is central to benchmark results.
assumptions (4)
  • ad hoc to paper The learned atom frequencies remain on a uniform DFT grid so that the analysis and synthesis steps can be computed with radix-2 FFTs at O(L log L).
    Section 3.1 allows A_k, f_k, phi_k to be learned by gradient descent and drift away from the rigid DFT grid, yet Section 4 claims the forward path is FFT plus IFFT with O(L log L) cost. For arbitrary f_k, Phi^H X costs O(K L d) rather than O(L log L).
  • domain assumption The operator SDict(X) = Re(Phi Phi^dagger X) is a meaningful token mixer with sufficient expressivity to replace self-attention.
    The paper assumes a rank-K linear projection followed by nonlinear feedforward layers can match attention-based models; no proof or expressivity analysis is provided.
  • domain assumption The reported training and evaluation setup produces numbers comparable to published baselines.
    The comparisons in Tables 1 and 2 mix fine-tuned and zero-shot baselines, and the paper does not report variance, hyperparameter search, or exact evaluation settings for the baselines.
  • standard math Phi^H Phi is invertible (atoms linearly independent) for every layer during training.
    The pseudo-inverse formula in Eq. (3) requires (Phi^H Phi)^{-1}; no regularization or rank-condition is stated.
invented entities (1)
  • Shared spectral dictionary of K learnable complex exponential atoms
    purpose: Used as the token mixer to replace convolutions and self-attention in the VLM.
    The atoms are internal learned parameters with no outside falsifiable handle. Their interpretability is presented qualitatively via Fig. 2, not through a measured external benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs." pith.science (2026). https://pith.science/paper/IOFN2F3P

@misc{pith2026250618943,
  author       = {Pith},
  title        = {Pith review of: From Pixels and Words to Waves: A Unified Framework for Spectral Dictionary vLLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IOFN2F3P}},
  note         = {Machine review of arXiv:2506.18943}
}
read the original abstract

Vision-language models (VLMs) unify computer vision and natural language processing in a single architecture capable of interpreting and describing images. Most state-of-the-art systems rely on two computationally intensive components: convolutions in the vision encoder and quadratic self-attention for multimodal fusion. This work removes both by introducing a spectral dictionary token mixer, which represents each image patch or wordpiece as a sparse combination of learnable frequency atoms. Our 1.1B-parameter prototype, SDict-VLM, achieves BLEU-4 of 39.2, CIDEr of 127.5, and SPICE of 27.0 on MS-COCO captioning, along with 50.3 percent accuracy on VQAv2. These results close approximately 85 percent of the performance gap to BLIP-2 while using 60 percent fewer parameters, 2.3 times less peak GPU memory, and 2.2 times faster inference than PaLI-3. To our knowledge, this is the first VLM to eliminate both convolutions and self-attention while matching mid-scale transformer baselines. In addition to its O(L log L) complexity, the shared frequency dictionary enables transparent cross-modal alignment and offers a tunable trade-off between accuracy and compute, paving the way for efficient and interpretable VLMs.

Figures

Figures reproduced from arXiv: 2506.18943 by the authors.

Figure 1
Figure 1. Signal flow in SDict-VLM. The architecture comprises three parallel stages whose data streams converge in a shared spectral decoder. (Left) The vision branch receives a raw RGB image and divides it into non-overlapping p × p patches. Each patch is linearly projected to a d-dimensional token and augmented with a learned 2-D positional code before entering a stack of Nv SDICT-VISION layers. Every layer executes an FFT… view at source ↗
Figure 2
Figure 2. Magnitude spectra of four representative SDict atoms. Each 128×128 panel shows the log-scaled magnitude of the 2-D FFT for a learned atom. Top-left: a near-DC (low-frequency) atom that captures global colour and illumination cues. Top-right: a mid-band horizontal/vertical atom sensitive to object boundaries and coarse edges. Bottom-left: an oblique mid-band atom that fires on diagonal struc￾tures such as surfboards … view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. Operator-Based Machine Intelligence: A Hilbert Space Framework for Spectral Learning and Symbolic Reasoning

    cs.LG 2025-07 reject novelty 2.0 of 10

    A review of RKHS, scattering, Koopman, and spectral dictionary models under a Hilbert-space operator framing, plus an unvalidated reasoning operator and literature-reported benchmark numbers.

Reference graph

Works this paper leans on

34 extracted references · 27 canonical work pages · cited by 1 Pith paper

  1. [1]

    FNet: Mixing Tokens with Fourier Transforms,

    J. Lee-Thorpet al., “FNet: Mixing Tokens with Fourier Transforms,” NAACL 2022

    work page 2022

  2. [2]

    Global Filter Networks for Image Classification,

    Y . Raoet al., “Global Filter Networks for Image Classification,” NeurIPS 2021

    work page 2021

  3. [3]

    Spectral Dictionary Learning for Generative Image Modeling

    A. Kiruluta, “Spectral Dictionary Learning for Generative Image Modeling,” arXiv 2504.17804, 2025

    work page Pith review arXiv 2025

  4. [4]

    From Attention to Atoms: Spectral Dictionary Learning for Fast Language Models,

    A. Kiruluta, “From Attention to Atoms: Spectral Dictionary Learning for Fast Language Models,” arXiv 2505.00033, 2025

    arXiv 2025

  5. [5]

    Graph Laplacian Wavelet Transformer via Learnable Spectral Decomposition

    A. Kiruluta, E. Lundy and P. Burity, “Graph Laplacian Wavelet Transformer via Learnable Spectral Decomposition,” arXiv 2505.07862, 2025

    work page Pith review arXiv 2025

  6. [6]

    Flamingo: A Visual Language Model for Few-Shot Learning,

    J. Alayracet al., “Flamingo: A Visual Language Model for Few-Shot Learning,” arXiv 2204.14198, 2022

    arXiv 2022

  7. [7]

    BLIP-2: Bootstrapping Language-Image Pre-training,

    J. Liet al., “BLIP-2: Bootstrapping Language-Image Pre-training,” PMLR 2023

    work page 2023

  8. [8]

    PaLI-3: Vision Language Models—Smaller, Faster, Stronger,

    X. Chenet al., “PaLI-3: Vision Language Models—Smaller, Faster, Stronger,” arXiv 2310.09199, 2023. 12

    arXiv 2023

Show all 34 references
  1. [9]

    Kosmos-2: Grounding Multimodal Large Language Models,

    Z. Penget al., “Kosmos-2: Grounding Multimodal Large Language Models,” arXiv 2306.14824, 2023

    2023 arXiv

  2. [10]

    IDEFICS: An Open Reproduction of SOTA Visual Language Models,

    R. Touvronet al., “IDEFICS: An Open Reproduction of SOTA Visual Language Models,” Hug- gingFace Blog 2023

    2023

  3. [11]

    Available athttps://chat.openai.com

    OpenAI.ChatGPT (o3) [Large language model]. Available athttps://chat.openai.com

  4. [12]

    Gemini 1.5: Next-Generation Model,

    Google DeepMind, “Gemini 1.5: Next-Generation Model,” Feb 2024

    2024

  5. [13]

    Improved Baselines with Visual Instruction Tuning,

    H. Liuet al., “Improved Baselines with Visual Instruction Tuning,” arXiv 2310.03744, 2023

    2023 arXiv

  6. [14]

    Learning Transferable Visual Models From Natural-Language Supervision,

    A. Radford, J. Witte, R. Kimet al., “Learning Transferable Visual Models From Natural-Language Supervision,”ICML, 2021

    2021

  7. [15]

    An Image Is Worth 16×16 Words: Transformers for Image Recognition at Scale,

    A. Dosovitskiy, L. Beyer, A. Kolesnikovet al., “An Image Is Worth 16×16 Words: Transformers for Image Recognition at Scale,”ICLR, 2021

    2021

  8. [16]

    Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows,

    Z. Liu, Y . Lin, Y . Caoet al., “Swin Transformer: Hierarchical Vision Transformer Using Shifted Windows,”ICCV, 2021

    2021

  9. [18]

    Rethinking Attention With Performers,

    K. Choromanski, V . Likhosherstov, D. Dohanet al., “Rethinking Attention With Performers,” ICLR, 2021

    2021

  10. [19]

    Perceiver IO: A General Architecture for Structured Inputs & Outputs,

    A. Jaegle, F. Gimeno, A. Brocket al., “Perceiver IO: A General Architecture for Structured Inputs & Outputs,”ICML, 2021

    2021

  11. [20]

    ALIGN: Large-Scale Image–Language Pre-training,

    C. Jia, Y . Yang, Y . Xionget al., “ALIGN: Large-Scale Image–Language Pre-training,”ICML, 2021

    2021

  12. [21]

    OSCAR: Object-Semantics Aligned Pre-training for Vision–Language Tasks,

    X. Li, X. Yin, C. Liet al., “OSCAR: Object-Semantics Aligned Pre-training for Vision–Language Tasks,”ECCV, 2020

    2020

  13. [22]

    ALBEF: Align Before Fuse for Multimodal Learning,

    J. Li, P. Zhou, C. Lianget al., “ALBEF: Align Before Fuse for Multimodal Learning,”NeurIPS, 2021

    2021

  14. [23]

    OFA: Unifying Architectures, Tasks, and Modali- ties Through a Simple Sequence-to-Sequence Learning Framework,

    Z. Wang, J. Yang, T. M. Hospedales, T. Xiang, “OFA: Unifying Architectures, Tasks, and Modali- ties Through a Simple Sequence-to-Sequence Learning Framework,”ICML, 2022

    2022

  15. [25]

    Hyena Hierarchy: Towards Larger Convolutional Language Models,

    S. Poli, D. Sorscher, S. Baconet al., “Hyena Hierarchy: Towards Larger Convolutional Language Models,”ICLR, 2024

    2024

  16. [26]

    A Wavelet Tour of Signal Processing,

    S. Mallat, “A Wavelet Tour of Signal Processing,” 3rd ed., Academic Press, 2008

    2008

  17. [28]

    Reducing Deep Network Complexity with Fourier Transform Methods,

    A. Kiruluta "Reducing Deep Network Complexity with Fourier Transform Methods," arXiv:1801.01451v1, Dec 2017

    2017 arXiv

  18. [29]

    RoFormer: Enhanced Transformer with Rotary Position Embedding,

    J. Su, X. Lu, Y . Panet al., “RoFormer: Enhanced Transformer with Rotary Position Embedding,” ACL, 2021. 13

    2021

  19. [30]

    Online Learning for Matrix Factorization and Sparse Coding,

    J. Mairal, F. Bach, J. Ponce, G. Sapiro, “Online Learning for Matrix Factorization and Sparse Coding,”Journal of Machine Learning Research, vol. 11, pp. 19–60, 2010

    2010

  20. [31]

    Linformer: Self-Attention with Linear Complexity,

    S. Wang, B. Li, M. Khabsa, H. Fang, H. Ma, “Linformer: Self-Attention with Linear Complexity,” arXiv:2006.04768, 2020

    2006 arXiv

  21. [32]

    FFCN: Fourier Fully Convolutional Networks for Semantic Segmenta- tion,

    Y . Luo, J. Li, B. Li,et al., “FFCN: Fourier Fully Convolutional Networks for Semantic Segmenta- tion,”ECCV, 2022

    2022

  22. [33]

    Fourier Neural Operator for Parametric PDEs,

    Z. Li, N. Kovachki, K. Azizzadenesheli,et al., “Fourier Neural Operator for Parametric PDEs,” ICLR, 2022

    2022

  23. [34]

    SpectFormer: Frequency and Channel Dynamic Spectrum Filtering Trans- former,

    S. Park, Y . Kim, J. Lee, “SpectFormer: Frequency and Channel Dynamic Spectrum Filtering Trans- former,”CVPR, 2023

    2023

  24. [35]

    Efficiently Modeling Long Sequences with Structured State Spaces,

    A. Gu, T. Goel, C. Ré, “Efficiently Modeling Long Sequences with Structured State Spaces,”ICLR, 2022

    2022

  25. [36]

    Hyena Hierarchy: Towards Larger Convolutional Language Models,

    S. Poli, D. Sorscher, S. Bacon,et al., “Hyena Hierarchy: Towards Larger Convolutional Language Models,”ICLR, 2024

    2024

  26. [37]

    Mamba: Linear-Time Sequence Modeling with Selective State Spaces,

    A. Sun, S. Bello, Z. Zeng,et al., “Mamba: Linear-Time Sequence Modeling with Selective State Spaces,” arXiv:2312.00752, 2023. 14

    2023 arXiv

Pith tools

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