Atomizer: Generalizing to new modalities by breaking satellite images down to a set of scalars

Diego Marcos; Hugo Riffaud de Turckheim; Roberto Interdonato; Sylvain Lobry

REVIEW 4 major objections 6 minor 17 references

Atomizer: Generalizing to new modalities by breaking satellite images down to a set of scalars

T0 review · 4 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 treating every spectral band of every pixel as a metadata-tagged scalar token lets one encoder generalize across satellite modalities without retraining or resampling.

desk verdict Clever per-pixel tokenization with big reported gains, but the 'new modality' evidence is limited to recombined Sentinel-2 settings. read the letter →

arxiv 2506.13542 v2 pith:WAXFBZAR submitted 2025-06-16 cs.CV

Hugo Riffaud de Turckheim , Sylvain Lobry , Roberto Interdonato , Diego Marcos This is my paper · ORCID

classification cs.CV

keywords remotesensingmultimodallearningtokenizationmodalitygeneralizationmulti-spectralimagerycross-attentionFourierfeaturesBigEarthNet

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 claims that a satellite image can be processed as a flat set of scalars—one per spectral band per pixel—each carrying metadata about spatial resolution, position, central wavelength, and bandwidth. This lets a single Perceiver-style encoder handle arbitrary image sizes, ground distances, and spectral band combinations without resampling, interpolation, or modality-specific modules. The experiments use a modality-disjoint protocol in which training and test observations never share the same combination of size, resolution, and bands; Atomizer reports higher mean average precision than Perceiver, ViT, ResNet, and ScaleMAE on all six test modalities, and also on the standard BigEarthNet benchmark. If the protocol is a fair proxy for new sensors, this is a route to Earth-observation models that keep working as new satellites are launched.

What carries the argument

The load-bearing object is the atomic token $z_{xyb} = \mathrm{Concat}(\phi_I(I_{xyb}), \phi_{\mathrm{res}}(x,y,g), \phi_\lambda(\lambda_b, \Delta\lambda_b))$, where $\phi_I$ is a Fourier-feature encoding of the reflectance value, $\phi_{\mathrm{res}}$ is a resolution-scaled Fourier positional encoding using ground sampling distance, and $\phi_\lambda$ is a non-uniform radial-basis-function encoding of the band's central wavelength and bandwidth. These tokens are processed by Perceiver-style cross-attention into a compact set of learned latent tokens, followed by attention pooling for classification. The encoding is what lets the same architecture treat a 120-pixel, 12-band image and an 84-pixel, 2-band image as comparable sets of scalars.

What would settle it

Train Atomizer only on the Sentinel-2-derived training modalities described in the paper, then evaluate without fine-tuning on native imagery from a different satellite, such as Landsat 8 or MODIS, whose central wavelengths, bandwidths, or ground-sample distances fall outside the encoding ranges used in training. If mean average precision drops to the level of the resolution-invariant baselines instead of staying near the reported 42–48% range, the claim that the scalar tokenization generalizes to genuinely new modalities fails.

Watch

Extended reading notes

Core claim

Atomizer claims that the bottleneck to cross-modality generalization is not the attention machinery but the tokenization: representing each observed reflectance as a bare number discards the sensor context, so the model cannot align bands across sensors. Atomizer's token is the atomic measurement enriched with resolution-scaled Fourier position and RBF-encoded spectral support; feeding such tokens through cross-attention to learned latent tokens allows the encoder to remain fixed while the input modality varies. The reported results—42.50 to 47.74% mAP on the unseen-modality tests, versus 14.52 to 38.33% for baselines, and 48.66% versus 33.47% on BigEarthNet—are offered as evidence that this atomic representation is what carries generalization, since Perceiver with the same architecture but less informative tokens scores near 15%.

Load-bearing premise

The load-bearing premise is that the test protocol's recombinations of Sentinel-2 bands, pixel grids, and resolutions behave like genuinely new satellite modalities; if a real new sensor has spectral responses, band placements, or resolutions outside the ranges the encoding was designed for, the paper's generalization claim has not been tested.

Editorial extensions

If this is right

A single Atomizer encoder can ingest images of any pixel size, ground resolution, and spectral band composition without resizing, resampling, or retraining.
A new satellite's bands could be added to the token vocabulary simply by supplying wavelength, bandwidth, and resolution metadata, with no new encoder head required.
Random token pruning at 50% keeps memory manageable, so the representation can scale to high-resolution or spectrally rich images.
Because Atomizer's advantage over Perceiver with the same architecture is the token encoding, further gains should come from richer metadata encodings rather than new attention designs.
The standard BigEarthNet result of 48.66% AP shows the modality-agnostic design does not sacrifice performance on a conventional fixed-format benchmark.

Reading between the lines

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

This suggests a practical path to a single Earth-observation foundation model trained on mixed archives from many satellites, with new missions added by emitting tokens instead of retraining a patch-based encoder.
A stricter test than the paper's protocol would be training only on Sentinel-2-style combinations and evaluating on native Landsat or MODIS imagery; the paper's protocol draws all data from one sensor, so actual cross-sensor transfer remains unmeasured.
The larger drop at 80 m/px (44.09% versus 48.66% at 20 m/px) points to resolution coverage as the likely weak spot; pushing resolution far outside the training range should be the next stress test.
The same scalar-token principle could extend to non-optical sensors such as synthetic aperture radar if the spectral RBF encoding is replaced by a comparable metadata encoding for radar geometry, but the paper does not demonstrate this.

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.

Desk Editor's Note

read the letter

Short version: Atomizer is a well-designed tokenization scheme and the reported numbers are striking, but the paper's central claim—generalizing to arbitrary new modalities—is only supported by a protocol that recombines attributes from one sensor (Sentinel-2). The architecture is genuinely new: per-band-per-pixel tokens carrying wavelength, bandwidth, resolution, and position, encoded with Fourier features and non-uniform RBFs, processed through a Perceiver-style latent bottleneck. That is a clean departure from patch-based models and a meaningful extension of pixel-token ideas like Presto. The modality-disjoint evaluation is the right kind of stress test, and the gains over the baselines are consistent across all six tests.

The soft spots are real. Every test modality is derived from Sentinel-2: the GSDs lie between 10 and 15.38 m/px, the spatial sizes are 84–120 px, and the spectral bands are subsets of the same 12 Sentinel-2 bands. So the test sets probe interpolation over recombined configurations, not extrapolation to a genuinely different sensor with unknown spectral response functions or out-of-range bandwidths. The abstract says "arbitrary modalities without interpolation or resampling," but the experiments don't support that strong reading. The conclusion does acknowledge the limitation, which I respect. The RBF centers are hand-placed around 400–800 nm; if a sensor has bands at 2200 nm or with odd bandwidths, the encoding could become near-zero or ambiguous. We are not shown the actual RBF layouts, so this is a genuine unknown. Also, no error bars, no code or data release, and the Perceiver baseline at 15.16 AP looks suspiciously low, suggesting it may not have been tuned fairly.

That said, the idea is worth pursuing. It is a clean architecture and the internal logic is sound. The paper deserves a serious referee, but the referee should insist on at least one true cross-sensor experiment (e.g., Landsat or MODIS vs Sentinel-2) or simulated out-of-range attributes, plus variance and code release. I would not cite the generalization claim as established, but I would cite the tokenization design as a useful building block. Bring it to reading group if you want a discussion of how far synthetic disjoint protocols can carry a generalization argument.

Referee Report

4 major / 6 minor

Summary. The paper proposes Atomizer, a Perceiver-style architecture that represents a multi-spectral satellite image as an unordered set of scalar tokens, one per spectral band per pixel. Each token concatenates a Fourier encoding of the band value with Fourier-based, resolution-scaled positional encodings and an RBF-based spectral encoding of central wavelength and bandwidth. The model is trained end-to-end on multilabel classification and is evaluated on the BigEarthNet benchmark under a 'modality-disjoint' protocol in which training and test images are assigned different combinations of spatial size, ground sampling distance, and Sentinel-2 band subsets. The authors report large mAP improvements over Perceiver, ViT, ResNet, and ScaleMAE on all six modality-disjoint tests (e.g., 47.74% vs. 26.90% AP on Test 6) and on the standard BigEarthNet benchmark (48.66% vs. 33.47% AP for the best baseline).

Significance. If the generalization claim is substantiated, Atomizer would be a valuable step toward sensor-agnostic Earth observation models, since it removes the need for resampling or retraining when encountering new sensor configurations. The paper's strengths are its clear and simple tokenization idea, the explicit encoding of physical metadata, and the consistency of the reported gains across several disjoint test configurations and ablations. However, the central contribution — 'generalizing to new modalities' — is currently supported only by recombinations of attributes drawn from a single sensor (Sentinel-2), so the significance of the result is not yet established at the level claimed in the abstract and introduction.

major comments (4)

[Section 4, Table 1, Figure 2] The modality-disjoint protocol does not evaluate generalization to genuinely new modalities. All training and test modalities are constructed from BigEarthNet/Sentinel-2: spectral variation is limited to subsets of the same 12 bands, GSD values lie between 10 and 15.38 m/px, and spatial sizes range only from 72 to 120 pixels. Several test GSD values (e.g., 12.5 m/px in Test 2) appear in the training set, albeit paired with different image sizes. The results therefore demonstrate interpolation over resampled and recombined Sentinel-2 configurations, not the 'entirely different satellites' or 'arbitrary modalities' claimed in the Introduction. A cross-sensor experiment with, for example, Landsat 8/9 or MODIS bands, or at minimum out-of-range GSD and band placements, is needed to support the central claim.
[Section 3.1, spectral and resolution encoders] The hand-placed non-uniform RBF centers and widths in the spectral encoder are concentrated around 400–800 nm, and the resolution encoder uses Fourier features with a reference normalization constant G calibrated to the training GSD range. No experiment shows how the model behaves when a band falls outside the RBF support (e.g., a SWIR band near 2200 nm) or when the GSD lies outside the calibrated range. If the encoding saturates or yields near-zero vectors for such inputs, the model would fail precisely in the 'new modality' regime the paper claims to handle. The authors should report a sensitivity analysis or a cross-sensor test that exercises out-of-range attribute values.
[Section 5, Tables 1–3] All performance numbers are reported from single runs with no error bars, multiple seeds, or statistical significance tests. Given that the central claim is about generalization and that some differences are large while others (e.g., the resolution scaling trend in Table 2) are more nuanced, the absence of variance estimates makes it difficult to assess robustness. The authors should report mean and standard deviation over at least three seeds for the main experiments.
[Sections 2 and 5] The related-work section identifies AnySat and Presto as methods that also handle variable resolutions, scales, and modalities, yet neither is included in the experimental comparison. Because the paper's contribution is precisely modality generalization, these are the most relevant baselines, and their absence weakens the claim that Atomizer outperforms 'standard models' in the cross-modality setting. Adding at least one of these baselines, or explaining why a direct comparison is infeasible, is necessary for a fair evaluation.

minor comments (6)

[Figure 2] The figure is difficult to interpret: the legend says the shade of blue indicates the number of bands, but the rows for Test 3 and Test 5 appear to have missing or empty cells, and the exact band list per modality is not given. Please label each row with the explicit band set or provide the full configuration table.
[Table 1] The caption does not identify which modality configuration each Test 1–6 corresponds to. Please add a sentence or cross-reference to Figure 2 or a supplementary table so the reader can map the test columns to spatial size, GSD, and band subsets.
[Section 3.1] The notation is inconsistent: the bandwidth is denoted µ in the token definition and in the RBF integral, but µ is also used for the Gaussian means (µ_1,...,µ_k), and the resolution encoding text says the vector size is '4L+2' while also stating that original pixel coordinates are concatenated. Please clarify both the notation and the exact concatenation order.
[Section 3.2] The token pruning strategy is described as random removal of 50% of tokens before each cross-attention block during training. It is unclear whether pruning is applied at inference time and how test-time performance depends on the pruned fraction. Please specify the inference behavior and report a sensitivity check if pruning is used at test time.
[Section 1] There is a typo in the first paragraph: 'Y et traditional vision models' should read 'Yet traditional vision models.'
[Section 4] The text repeatedly refers to 'BigEarthNet' but cites the refined reben dataset [16]. Please clarify which exact version of BigEarthNet is used and whether the class set and splits follow reben.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: Atomizer's empirical claims are self-contained; the Sentinel-2-only test protocol is a limitation, not a circular derivation.

full rationale

The paper's central claim is empirical: a tokenization scheme with explicit metadata (resolution, wavelength, bandwidth, position) is trained on BigEarthNet and evaluated under a held-out modality-disjoint protocol. No equation is defined in terms of the result it is said to predict, and no fitted parameter is renamed as a prediction. The RBF and Fourier encoders are fixed, hand-specified encodings; they are not fit to the test modalities, and the test modalities are not used in training. The paper is also not dependent on self-citation: it cites Perceiver, Fourier features, ScaleMAE, and other external works, but the load-bearing evidence is the reported Table 1 and the standard BigEarthNet benchmark. The reviewer concern that all test modalities are recombinations of Sentinel-2 attributes, so the system is only interpolating within one sensor family, is a valid external-validity limitation, and the paper partly acknowledges it ('Although we have focused on single-time multi-spectral imagery derived from Sentinel-2'). But this is not circularity: the derivation of the architecture and the reported measurements do not reduce to the conclusion by construction. Therefore the appropriate circularity score is 0.

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

The encoding and architecture rely on a set of hand-chosen hyperparameters (Fourier frequency count, maximum frequency, RBF placement, pruning ratio) and on the assumption that resampled Sentinel-2 recombinations proxy for new sensors. No new physical entities are introduced.

free parameters (5)

reference normalization constant G = not reported
Divides the physical coordinate scaling in the resolution-aware positional encoding (Section 3.1); chosen by hand and not ablated.
Fourier frequency count L = not reported
Number of frequency components in Fourier features for reflectance and position; a hand-chosen hyperparameter.
maximum Fourier frequency f_max = not reported
Controls the frequency range of Fourier features; hand-chosen.
RBF center positions and widths = not reported (non-uniform, denser in visible range)
The k Gaussian centers and widths for spectral encoding are strategically hand-placed to match known sensor bands (Section 3.1); no ablation or sensitivity analysis.
token pruning proportion p = 0.5
Randomly removes half of input tokens before each cross-attention during training to reduce memory; chosen without reported sensitivity study.

assumptions (3)

domain assumption Perceiver-style cross-attention with a compact latent set can capture sufficient information for multilabel classification from per-scalar tokens.
The architecture relies on this to compress large token sets; supported by Perceiver results but not established for this tokenization.
ad hoc to paper Resampled and band-subset Sentinel-2 images are a valid proxy for unseen satellite modalities.
The modality-disjoint protocol in Section 4 uses only combinations of bands, sizes, and resolutions from one sensor, so cross-sensor generalization is assumed rather than measured.
domain assumption Spatial structure can be recovered purely from Fourier positional encodings attached to each scalar token.
The model treats tokens as a set; the positional encoding must inject enough geometric information for the classification task.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Atomizer: Generalizing to new modalities by breaking satellite images down to a set of scalars." pith.science (2026). https://pith.science/paper/WAXFBZAR

@misc{pith2026250613542,
  author       = {Pith},
  title        = {Pith review of: Atomizer: Generalizing to new modalities by breaking satellite images down to a set of scalars},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WAXFBZAR}},
  note         = {Machine review of arXiv:2506.13542}
}

read the original abstract

The growing number of Earth observation satellites has led to increasingly diverse remote sensing data, with varying spatial, spectral, and temporal configurations. Most existing models rely on fixed input formats and modality-specific encoders, which require retraining when new configurations are introduced, limiting their ability to generalize across modalities. We introduce Atomizer, a flexible architecture that represents remote sensing images as sets of scalars, each corresponding to a spectral band value of a pixel. Each scalar is enriched with contextual metadata (acquisition time, spatial resolution, wavelength, and bandwidth), producing an atomic representation that allows a single encoder to process arbitrary modalities without interpolation or resampling. Atomizer uses structured tokenization with Fourier features and non-uniform radial basis functions to encode content and context, and maps tokens into a latent space via cross-attention. Under modality-disjoint evaluations, Atomizer outperforms standard models and demonstrates robust performance across varying resolutions and spatial sizes.

Figures

Figures reproduced from arXiv: 2506.13542 by the authors.

**Figure 1.** Top: Atomizer’s token-based representation. Each spectral band of each pixel is decomposed into a token that encodes multiple attributes: reflectance (band value), wavelength, bandwidth, resolution, and position. Bottom: Diagram of the used architecture. 3.1 Token Construction To represent the highly heterogeneous nature of remote sensing data, we build a token for every spectral band of every pixel. Each token is d… view at source ↗

**Figure 2.** Modalities used for training and testing. Each row represents a distinct modality configuration with its spatial dimensions (in pixels), ground sampling distance (meters per pixels), and spectral bands. The shade of blue represents the total number of bands (2-12). 5. Results This section presents experimental results comparing Atomizer with baseline models (Perceiver [8], ViT [10], ResNet [17], and ScaleMAE [6]) ac… view at source ↗

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

17 extracted references · 10 canonical work pages

[1]

Kropuenske et al.Earth Observing Sensing Satellites Online Compendium: U.S

T. Kropuenske et al.Earth Observing Sensing Satellites Online Compendium: U.S. Geological Survey digital data. 2024.URL: https://calval.cr.usgs.gov/apps/compendium (visited on 05/15/2025)

work page 2024
[2]

Improved field-scale drought monitoring using MODIS and Sentinel-2 data for vegetation temperature condition index generation through a fusion framework

Mingqi Li et al. “Improved field-scale drought monitoring using MODIS and Sentinel-2 data for vegetation temperature condition index generation through a fusion framework”. In:Computers and Electronics in Agriculture234 (2025), p. 110256

work page 2025
[3]

Planted: a dataset for planted forest identification from multi-satellite time series

Luis Miguel Pazos-Outón et al. “Planted: a dataset for planted forest identification from multi-satellite time series”. In: IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE. 2024, pp. 7066–7070

work page 2024
[4]

S Skakun et al.Combined Use of Landsat-8 and Sentinel-2A Images for Winter Crop Mapping and Winter Wheat Yield Assessment at Regional Scale, AIMS Geosciences, 3, 163–186. 2017

work page 2017
[5]

AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities

Guillaume Astruc et al. “AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities”. In:arXiv preprint arXiv:2412.14123(2024)

arXiv 2024
[6]

Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning

Colorado J Reed et al. “Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4088–4099

work page 2023
[7]

FlexiMo: A Flexible Remote Sensing Foundation Model

Xuyang Li et al. “FlexiMo: A Flexible Remote Sensing Foundation Model”. In:arXiv preprint arXiv:2503.23844(2025)

arXiv 2025
[8]

Perceiver: General perception with iterative attention

Andrew Jaegle et al. “Perceiver: General perception with iterative attention”. In:International conference on machine learning. PMLR. 2021, pp. 4651–4664

work page 2021

Show all 17 references

[9]

Fourier features let networks learn high frequency functions in low dimensional domains

Matthew Tancik et al. “Fourier features let networks learn high frequency functions in low dimensional domains”. In: Advances in neural information processing systems33 (2020), pp. 7537–7547

2020
[10]

An image is worth 16x16 words: Transformers for image recognition at scale

Alexey Dosovitskiy et al. “An image is worth 16x16 words: Transformers for image recognition at scale”. In:arXiv preprint arXiv:2010.11929(2020)

2020 arXiv
[11]

Senpa-mae: Sensor parameter aware masked autoencoder for multi-satellite self-supervised pretraining

Jonathan Prexl and Michael Schmitt. “Senpa-mae: Sensor parameter aware masked autoencoder for multi-satellite self-supervised pretraining”. In:DAGM German Conference on Pattern Recognition. Springer. 2024, pp. 317–331

2024
[12]

Galileo: Learning Global and Local Features in Pretrained Remote Sensing Models

Gabriel Tseng et al. “Galileo: Learning Global and Local Features in Pretrained Remote Sensing Models”. In:arXiv preprint arXiv:2502.09356(2025)

2025 arXiv
[13]

A survey of remote-sensing big data

Peng Liu. “A survey of remote-sensing big data”. In:frontiers in Environmental Science3 (2015), p. 45

2015
[14]

Lightweight, pre-trained transformers for remote sensing timeseries

Gabriel Tseng et al. “Lightweight, pre-trained transformers for remote sensing timeseries”. In:arXiv preprint arXiv:2304.14065 (2023)

2023 arXiv
[15]

Self-attentive pooling for efficient deep learning

Fang Chen et al. “Self-attentive pooling for efficient deep learning”. In:Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision. 2023, pp. 3974–3983

2023
[16]

reben: Refined bigearthnet dataset for remote sensing image analysis

Kai Norman Clasen et al. “reben: Refined bigearthnet dataset for remote sensing image analysis”. In:arXiv preprint arXiv:2407.03653(2024)

2024 arXiv
[17]

Deep residual learning for image recognition

Kaiming He et al. “Deep residual learning for image recognition”. In:Proceedings of the IEEE conference on computer vision and pattern recognition. 2016, pp. 770–778

2016

Pith tools

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

[1] [1]

Kropuenske et al.Earth Observing Sensing Satellites Online Compendium: U.S

T. Kropuenske et al.Earth Observing Sensing Satellites Online Compendium: U.S. Geological Survey digital data. 2024.URL: https://calval.cr.usgs.gov/apps/compendium (visited on 05/15/2025)

work page 2024

[2] [2]

Improved field-scale drought monitoring using MODIS and Sentinel-2 data for vegetation temperature condition index generation through a fusion framework

Mingqi Li et al. “Improved field-scale drought monitoring using MODIS and Sentinel-2 data for vegetation temperature condition index generation through a fusion framework”. In:Computers and Electronics in Agriculture234 (2025), p. 110256

work page 2025

[3] [3]

Planted: a dataset for planted forest identification from multi-satellite time series

Luis Miguel Pazos-Outón et al. “Planted: a dataset for planted forest identification from multi-satellite time series”. In: IGARSS 2024-2024 IEEE International Geoscience and Remote Sensing Symposium. IEEE. 2024, pp. 7066–7070

work page 2024

[4] [4]

S Skakun et al.Combined Use of Landsat-8 and Sentinel-2A Images for Winter Crop Mapping and Winter Wheat Yield Assessment at Regional Scale, AIMS Geosciences, 3, 163–186. 2017

work page 2017

[5] [5]

AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities

Guillaume Astruc et al. “AnySat: An Earth Observation Model for Any Resolutions, Scales, and Modalities”. In:arXiv preprint arXiv:2412.14123(2024)

arXiv 2024

[6] [6]

Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning

Colorado J Reed et al. “Scale-mae: A scale-aware masked autoencoder for multiscale geospatial representation learning”. In:Proceedings of the IEEE/CVF International Conference on Computer Vision. 2023, pp. 4088–4099

work page 2023

[7] [7]

FlexiMo: A Flexible Remote Sensing Foundation Model

Xuyang Li et al. “FlexiMo: A Flexible Remote Sensing Foundation Model”. In:arXiv preprint arXiv:2503.23844(2025)

arXiv 2025

[8] [8]

Perceiver: General perception with iterative attention

Andrew Jaegle et al. “Perceiver: General perception with iterative attention”. In:International conference on machine learning. PMLR. 2021, pp. 4651–4664

work page 2021