Pith. sign in

REVIEW 4 major objections 5 minor 50 references

The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read A lightweight convolutional autoencoder can reconstruct planetary images that have lost 90% or more of their pixels to destructive noise, and the paper argues the learned latent space can outrank raw images as a science product.

desk verdict Plausible proof-of-concept for lightweight CAE inpainting, but the 90% masking claim needs baseline comparisons before it means more than scene-averaging. read the letter →

arxiv 2507.11400 v1 pith:FSQ6SSDG submitted 2025-07-15 astro-ph.EP astro-ph.IMcs.LG

classification astro-ph.EPastro-ph.IMcs.LG
keywords ImageprocessingMarssurfaceDatareductiontechniquesconvolutionalautoencoderinpaintinglatentrepresentationdestructivenoiseplanetaryimaging
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 small, simple convolutional autoencoder — roughly $10^{5}$ parameters, trained on subimages from one Mars scene — can serve as a practical reconstruction tool for multi-color planetary images degraded by destructive random noise. It demonstrates useful inpainting even when 90% or more of pixels are zeroed out, and a variant trained on mixed luminance-plus-color noise still produces recognizable reconstructions when every pixel has been perturbed by up to ±100% of the channel range. The results point to two concrete uses: deliberate masking to cut downlink bandwidth, and recovery of scenes obscured by low light, atmospheric, or instrumental effects. The paper's further, bolder claim is that the model's latent space and representations can have more scientific utility than a large corpus of raw images — 'the model is the message.' The proof-of-concept is explicitly narrow: the model learns the feature statistics of a single Mastcam image, and the sunflower 'out of bounds' test shows that out-of-distribution inputs are forcibly mapped into that known latent space rather than reconstructed honestly.

What carries the argument

The central object is the convolutional autoencoder's latent space: four convolutional layers (16, 32, 64, then 128 filters of 3x3 size with ReLU activations and 2x2 max pooling) compress a 128x128x3 image into a low-dimensional embedding, and a mirrored decoder with upsampling regenerates the image from that embedding. The mechanism that drives everything is masked training: inputs are deliberately destroyed — pixels randomly zeroed at fixed or varying areal coverage, or random additive perturbations across all color channels — and mean-squared-error backpropagation with the Adam optimizer forces the network's internal representation to encode the contextual features needed to fill in what is missing. The variant that carries the main results is gt-CAE, a single model trained across masking levels from 50% to 99%, whose near-equivalence to per-level fixed models is the paper's evidence that a generally trainable lightweight model is sufficient for the full range of destructive noise tested.

What would settle it

Retrain or reuse the gt-CAE and apply it to a second Perseverance Mastcam image of a different terrain unit under different lighting, measuring reconstruction error on an unseen strip as the paper does; if the error rises sharply relative to the in-scene validation strip, the claim that this lightweight approach transfers across planetary scenes fails. A cleaner scientific test: run a downstream analysis (for example, identifying or classifying surface features, or measuring color fidelity of geological units) on original images versus images reconstructed after 90% masking; the paper reports only pixel statistics, so if the scientific measurement degrades substantially while pixel errors stay small, the claim of 'useful reconstruction' for real science is not established.

Watch

Extended reading notes

Core claim

On its own terms the paper establishes that a deliberately simple convolutional autoencoder, trained on 5,000 random 128x128 RGB subimages cut from a single Perseverance Mastcam frame, learns a latent feature model of that scene sufficient to reconstruct unseen neighboring image strips whose pixels have been randomly zeroed at areal coverages from 50% to 99%. Reconstruction quality, measured as mean per-channel absolute pixel difference, stays in the range of about 3-5% even at 99% masking, with qualitative degradation becoming significant only around 90% coverage and above. The paper further claims that a single generally trained model (gt-CAE), built by presenting randomly varying masking levels to the same architecture, matches the per-level fixed-noise models, and that a gt-CAE trained on luminance plus additive color noise tolerates perturbations of ±100% of the channel range applied to every pixel, still yielding recognizable reconstructions. On that basis the paper argues that deliberate random masking could shrink returned image data by roughly 90% while retaining reconstructable scenes, and that for some scientific purposes the latent representation itself may be a more valuable returned product than the raw image corpus.

Load-bearing premise

The demonstration is built from subimages of a single Mastcam scene, so everything the paper concludes about planetary and astrobiological utility assumes that one scene's feature distribution — together with the two synthetic noise models — adequately represents the diverse terrains, lighting conditions, and degradations real missions will encounter.

Editorial extensions

If this is right

  • Deliberate, pre-determined random masking of image pixels could cut returned data volume by up to roughly 90% while still allowing a trained lightweight CAE to reconstruct scientifically usable multi-color scenes.
  • Because a single gt-CAE trained across 50-99% masking matches per-level fixed-noise models, a deployed system would not need to know the actual noise or mask level in advance.
  • Reconstruction quality stays nearly flat up to about 90% masking and degrades sharply only above it (the pixel-difference variance grows by a factor of roughly 3.5 between 50% and 99% coverage), defining a practical operating envelope for masked imaging.
  • For analysis workflows that are themselves machine-learning based, returning the model's latent representation instead of raw pixels could preserve equal or greater scientific value at a fraction of the data volume.
  • The sunflower test shows that any such model must be used within its training distribution or paired with explicit out-of-distribution or anomaly metrics, otherwise it will silently project unfamiliar scenes into the features it already knows.

Reading between the lines

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

  • A natural flight experiment follows: mask 90% of pixels in software during acquisition, reconstruct with the onboard model, and downlink the reconstruction plus a sparse set of unmasked reference pixels; comparing those reference pixels to the reconstruction on the ground would quantify 'scientific return per transmitted bit,' a metric the paper does not compute.
  • The single-scene training set is the obvious stress point: training the same architecture on a multi-scene Mars corpus would reveal whether gt-CAE acquires a genuinely planetary prior or simply memorizes one image's textures; the paper leaves this untested.
  • Because the paper measures only pixel statistics, the practical masking threshold probably depends on the scientific target: mapping large-scale terrain may tolerate 99% masking, while searching for small mineralogical or potential biosignature features may not survive even 90% — feature-weighted quality metrics would likely replace pixel-error metrics in actual mission use.
  • If the 'model is the message' idea is pushed to its limit, a variational autoencoder — which the paper sketches but does not test — would let the returned product be a probability distribution over scenes, likely far more compressible than even a 10% pixel sample and interrogable for statistically valid synthetic views of sites never fully imaged.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper presents a lightweight convolutional autoencoder (CAE) applied to a single Mars Mastcam image, with variants trained on fixed and variable random-luminance masking levels, and on an additive noise variant. The central claim is that, for certain use cases, multi-color planetary image reconstruction remains useful at 90% or higher destructive areal coverage, and that the learned latent representation may be more valuable than raw image corpora. The paper is explicitly framed as a proof of concept and includes qualitative reconstructions, pixel-level error metrics, loss curves, and an out-of-distribution 'sunflower test'.

Significance. If properly supported, the result would be a useful proof of concept for lightweight onboard reconstruction and bandwidth reduction for planetary imaging. The paper is clearly written, the architecture is simple and reproducible (code is linked), and it candidly discusses the narrow training set and the out-of-distribution failure mode. The main weakness is that the quantitative support for the central 90% claim consists only of pixel-level MAE and variance on a single scene, with no comparison to trivial baselines such as mean filling or classical inpainting. The significance of the bandwidth and latent-space claims therefore remains, at this stage, more suggestive than demonstrated.

major comments (4)
  1. [Section 2.2.1, Figure 4] The central claim that reconstruction is 'useful' at 90% and higher masking is not benchmarked against any baseline. The paper itself notes that the CAE, in the absence of informative input, will output values close to the statistical mean pixel values, so the reported 3-5% MAE is exactly what a trivial per-channel mean fill would also produce on this statistically homogeneous scene. Without comparisons to mean fill, bicubic interpolation, total-variation inpainting, or a maximum-entropy reconstruction, the low MAE and the ~3.5x variance increase in Figure 4b do not establish that the model is doing genuine learned reconstruction rather than mean filling. Please add such baselines and report error and variance relative to them, or substantially temper the 90% claim.
  2. [Section 2.1.1 and Section 2.3, Figure 8] The model is trained on subimages drawn from a single Mastcam master image, and the 'unseen' validation data are subimages from a contiguous strip of that same image. The generalization in the Discussion and abstract to planetary science and astrobiology therefore rests on a single-scene feature distribution. The sunflower test in Section 2.3 shows that the model projects out-of-distribution inputs onto the Mars training palette and fails to reconstruct them, which underscores how restrictive the current evidence is. To support the broader claims, the evaluation needs either multiple scenes or an explicit statement that all conclusions are limited to within-scene reconstruction for the single image tested.
  3. [Section 3.1 and Section 3.3] The bandwidth-reduction and 'model is the message' proposals depend on the reconstruction being genuinely informative for downstream scientific use. The paper's own Section 3.3 acknowledges that pixel-by-pixel error maps have limited utility for determining the value of reconstructed data. Yet no task-level evaluation is provided, such as whether geomorphic features, spectral classes, or mineralogical units are correctly recovered from 90%-masked inputs. Without such an evaluation, the claim that latent representations 'have equal, if not greater, value than any corpus of raw imaging data' is unsupported. Please either add a task-level assessment or reframe this as a speculative direction rather than a demonstrated capability.
  4. [Section 2.2.3, Figure 7] The experiment described as 'fully destructive luminance + color noise' actually applies the same random additive value to all three RGB channels for a given pixel. This is common-mode luminance noise, not independent per-channel color noise. The abstract specifically advertises 'color noise with random additive levels across pixel channels,' but that case is not tested. If the intent is to claim robustness to color noise, the experiment must use independent random offsets for each channel; otherwise the relevant claim should be narrowed to luminance noise only.
minor comments (5)
  1. [Figure 1 caption] The phrase 'Mastcam mage data' appears to be a typo for 'Mastcam image data'.
  2. [Section 3.1, line citing Kerner] The citation 'Kerner et al. 2029' appears to be a typo for 'Kerner et al. 2019', based on the reference list.
  3. [Section 1.1] The statement that lossless JPEG2000 and LZW 'typically offer lossless compression ratios of 2:1' is a broad generalization with no citation; consider softening or providing a reference.
  4. [References] Reference formatting is inconsistent in places, e.g., 'LeCun' is spelled 'Lecun' in the text and the Devlin et al. reference lists 2019 while the citation text says 2018; a final copyedit pass would help.
  5. [Section 2.1] The hyperparameter choices (filter counts 16/32/64/128, 40 initial epochs, data augmentation parameters) are described but not justified; a sentence explaining why these are 'generic levels' would make the baseline claim clearer.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the CAE reconstruction results are not forced by construction; same-image validation and mean-fill MAE caveat are explicit limitations, not circular reductions.

full rationale

The paper's central claims are empirical demonstrations of a lightweight CAE on Mars Mastcam data, not derivations that reduce to their own inputs. Training uses 5,000 subimages from one master image with random masking, and evaluation is on held-out subimages from an unseen strip of the same scene; while this limits scene-level generalization, it is a standard in-distribution holdout and not a fitted parameter renamed as a prediction. The paper explicitly acknowledges in Section 2.2.1 that low MAE is expected because the CAE 'in the absence of informative input, will output values close to the statistical mean pixel values across all training images,' and therefore relies on a variance metric and qualitative error maps for its conclusions, so the MAE is not used as a forced success metric. The sunflower out-of-distribution test (Figure 8) is presented as a caution, further showing the authors do not claim universal reconstruction. The only self-citation (Scharf et al. 2024) is a contextual remark about community standards and is not load-bearing. No uniqueness theorem, ansatz, or known result is imported from the author's prior work. Hence there is no circular step.

Assumptions & free parameters 4 free parameters · 3 assumptions · 0 invented entities

No new physical entities are postulated. The main assumptions are that a single Mastcam scene and synthetic random noise models suffice to support the proposed planetary science use cases; these are stated domain assumptions, and the paper's own sunflower test demonstrates the limits of its latent space.

free parameters (4)
  • Encoder filter counts = [16, 32, 64, 128]
    Chosen by hand following standard practice; affects model capacity and reconstruction quality.
  • Training epochs = 10 (fixed-noise), 12-20 (gt-CAE)
    Chosen from validation loss crossover (Figure 5) to avoid overfitting; the exact stopping point affects reported reconstruction quality.
  • Noise coverage levels = 50-99% (luminance); 10-100% (color)
    Selected to span the claimed operating range; the 90% threshold is a result of scanning these levels.
  • Number of training subimages = 5000
    Chosen to balance compute and dataset size; larger or smaller sets could change the quantitative results.
assumptions (3)
  • domain assumption A single Mastcam image's subimages represent the feature distribution of planetary imaging data relevant to the proposed applications.
    Training and validation are drawn from one image (Figure 1), yet the discussion generalizes to planetary exploration; the sunflower test shows the model cannot handle out-of-distribution data.
  • domain assumption Random pixel masking and additive color noise approximate real destructive noise from instruments, low light, or transmission errors.
    Section 2.2 introduces this noise model; real noise may be spatially correlated, non-uniform, or include sensor artifacts.
  • domain assumption The validation subimages from the unused strip of the same master image are sufficiently out-of-sample to test generalization.
    Section 2.1.1 describes the split; because the strip is from the same scene, it tests spatial generalization but not scene generalization.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology." pith.science (2026). https://pith.science/paper/FSQ6SSDG

@misc{pith2026250711400,
  author       = {Pith},
  title        = {Pith review of: The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FSQ6SSDG}},
  note         = {Machine review of arXiv:2507.11400}
}
read the original abstract

The application of convolutional autoencoder deep learning to imaging data for planetary science and astrobiological use is briefly reviewed and explored with a focus on the need to understand algorithmic rationale, process, and results when machine learning is utilized. Successful autoencoders train to build a model that captures the features of data in a dimensionally reduced form (the latent representation) that can then be used to recreate the original input. One application is the reconstruction of incomplete or noisy data. Here a baseline, lightweight convolutional autoencoder is used to examine the utility for planetary image reconstruction or inpainting in situations where there is destructive random noise (i.e., either luminance noise with zero returned data in some image pixels, or color noise with random additive levels across pixel channels). It is shown that, in certain use cases, multi-color image reconstruction can be usefully applied even with extensive random destructive noise with 90% areal coverage and higher. This capability is discussed in the context of intentional masking to reduce data bandwidth, or situations with low-illumination levels and other factors that obscure image data (e.g., sensor degradation or atmospheric conditions). It is further suggested that for some scientific use cases the model latent space and representations have more utility than large raw imaging datasets.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 27 canonical work pages

  1. [1]

    1 The model is the message: Lightweight convolutional autoencoders applied to noisy imaging data for planetary science and astrobiology Caleb Scharf, NASA Ames Research Center (caleb.a.scharf@nasa.gov) (Accepted 7/14/25 for publication in Icarus) Abstract: The application of convolutional autoencoder deep learning to imaging data for planetary science and...

  2. [3]

    The code utilizes the common libraries: cv2, tensorflow, and keras (see Github project page for associated code, https://github.com/caleb-nasa/gt-CAE)

    The decoder mirrors the encoder with 4 layers and filter numbers in sequence as 128, 64, 32, 16 using upsampling with 2 x 2 upscaling to finally output 128 x 128, 3-channel reconstructed images. The code utilizes the common libraries: cv2, tensorflow, and keras (see Github project page for associated code, https://github.com/caleb-nasa/gt-CAE). The traini...

  3. [6]

    and models that employ attention-based transformer techniques (e.g., Hou et al., 2022; Bachmann et al., 2022). In this present work, the utility of a baseline, lightweight masked CAE is explored for reconstructing common planetary imaging data subject to destructive noise that could be a consequence of instrument behavior, deliberate sparse sampling/maski...

  4. [8]

    in-painting

    was chosen as an example of a complex, but relatively uniform terrain under non-extreme lighting conditions (i.e., the Sun was at moderate elevation). As noted previously, this present work is not intended to develop a generalized model for Mars’s highly diverse surface scenes, but to instead examine a proof-of-principle. The master image was divided into...

  5. [9]

    Error bars correspond to 1-sigma errors on the mean

    as a function of luminance noise/masking level. Error bars correspond to 1-sigma errors on the mean. (b) Right panel: Mean variance in pixel-by-pixel differences between original and reconstructed images, across all images, as a function of luminance noise/masking level for the 6 models trained here. Using this metric the loss/error ranges from about 3% t...

  6. [10]

    out of bounds

    Original unseen Noisy input CAE recon. Error +/-10% noise over 100% +/-50% noise over 100%+/-100% noise over 50% +/-100% noise over 90% +/-100% noise over 100% 18 While this success is significant, it does involve a relatively narrow and small training set. This type of denoising problem is also the basis of diffusion models (e.g., Ho et al. 2020), and fo...

  7. [11]

    best-guess

    the generalized luminance/masking noise trained gt-CAE (50-99% areal noise) presented with inputs subjected to 90% and 98% noise. The direct reconstruction (no noise) CAE recovers most of the primary image features (brightness and contrast, separation of forms) but produces a very different color palette that matches the color palette of the Mastcam train...

  8. [12]

    best possible

    However, the best strategy is highly dependent on the goal. If, for example, a CAE or similar modeling approach is to be used for novelty or anomaly detection on Mars, then pre-training with diverse terrestrial data may be counterproductive, as novel, interesting features on Mars may be common in terrestrial data. 3.0 Discussion 3.1 Applications The basel...

Show all 50 references
  1. [13]

    https://doi.org/10.48550/arXiv.2204.01678 Baggenstoss, P.M.,

    MultiMAE: Multi-modal Multi-task Masked Autoencoders. https://doi.org/10.48550/arXiv.2204.01678 Baggenstoss, P.M.,

  2. [16]

    Machine Learning with Applications 6, 100134

    Deep learning in computer vision: A critical review of emerging techniques and application scenarios. Machine Learning with Applications 6, 100134. https://doi.org/10.1016/j.mlwa.2021.100134 Chen, Y., Saygin, E.,

  3. [17]

    JGR Solid Earth 126, e2020JB021589

    Seismic Inversion by Hybrid Machine Learning. JGR Solid Earth 126, e2020JB021589. https://doi.org/10.1029/2020JB021589 Devlin, J., Chang, M.-W., Lee, K., Toutanova, K.,

  4. [19]

    https://doi.org/10.48550/arXiv.1603.07285 Etsebeth, V., Lochner, M., Walmsley, M., Grespan, M.,

    A guide to convolution arithmetic for deep learning. https://doi.org/10.48550/arXiv.1603.07285 Etsebeth, V., Lochner, M., Walmsley, M., Grespan, M.,

  5. [21]

    Masked Autoencoders As Spatiotemporal Learners. https://doi.org/10.48550/arXiv.2205.09113 Francis, R., Estlin, T., Doran, G., Johnstone, S., Gaines, D., Verma, V., Burl, M., Frydenvang, J., Montaño, S., Wiens, R.C., Schaffer, S., Gasnault, O., DeFlores, L., Blaney, D., Bornstein, B.,

  6. [24]

    https://doi.org/10.48550/arXiv.2205.03892 25 He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.,

    ConvMAE: Masked Convolution Meets Masked Autoencoders. https://doi.org/10.48550/arXiv.2205.03892 25 He, K., Chen, X., Xie, S., Li, Y., Dollar, P., Girshick, R.,

  7. [25]

    IEEE, New Orleans, LA, USA, pp

    Masked Autoencoders Are Scalable Vision Learners, in: 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). IEEE, New Orleans, LA, USA, pp. 15979–15988. https://doi.org/10.1109/CVPR52688.2022.01553 Hendrycks, D., Lee, K., Mazeika, M.,

  8. [26]

    https://doi.org/10.48550/arXiv.1901.09960 Hipperson, M., Waldmann, I., Grindrod, P., Nikolaou, N.,

    Using Pre-Training Can Improve Model Robustness and Uncertainty. https://doi.org/10.48550/arXiv.1901.09960 Hipperson, M., Waldmann, I., Grindrod, P., Nikolaou, N.,

  9. [28]

    https://doi.org/10.48550/arXiv.2006.11239 Holland, M

    Denoising Diffusion Probabilistic Models. https://doi.org/10.48550/arXiv.2006.11239 Holland, M. et al. (2022). Artificially enhancing titan’s SAR images using earth satellite images and deep learning technology. https://www.hou.usra.edu/meetings/lpsc2022/pdf/2592.pdf Hong, D.,...

  10. [29]

    https://doi.org/10.48550/arXiv.2311.07113 Hou, Q., Lu, C.-Z., Cheng, M.-M., Feng, J.,

    SpectralGPT: Spectral Remote Sensing Foundation Model. https://doi.org/10.48550/arXiv.2311.07113 Hou, Q., Lu, C.-Z., Cheng, M.-M., Feng, J.,

  11. [30]

    https://doi.org/10.48550/arXiv.2211.11943 Huang, B., Zhao, Z., Zhang, G., Qiao, Y., Wang, L.,

    Conv2Former: A Simple Transformer-Style ConvNet for Visual Recognition. https://doi.org/10.48550/arXiv.2211.11943 Huang, B., Zhao, Z., Zhang, G., Qiao, Y., Wang, L.,

  12. [32]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 2506–2517

    Contrastive Masked Autoencoders are Stronger Vision Learners. IEEE Transactions on Pattern Analysis and Machine Intelligence 46, 2506–2517. https://doi.org/10.1109/TPAMI.2023.3336525 Jam, J., Kendrick, C., Walker, K., Drouard, V., Hsu, J.G.-S., Yap, M.H.,

  13. [33]

    Computer Vision and Image Understanding 203, 103147

    A comprehensive review of past and present image inpainting methods. Computer Vision and Image Understanding 203, 103147. https://doi.org/10.1016/j.cviu.2020.103147 26 Kerner, H.R., Wagstaff, K.L., Bue, B.D., Gray, P.C., Bell, J.F., Ben Amor, H.,

  14. [34]

    IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12, 3900–3918

    Toward Generalized Change Detection on Planetary Surfaces With Convolutional Autoencoders and Transfer Learning. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 12, 3900–3918. https://doi.org/10.1109/JSTARS.2019.2936771 Kingma, D.P., Ba, J.,

  15. [35]

    https://doi.org/10.48550/arXiv.1412.6980 Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.,

    Adam: A Method for Stochastic Optimization. https://doi.org/10.48550/arXiv.1412.6980 Lecun, Y., Bottou, L., Bengio, Y., Haffner, P.,

  16. [37]

    https://doi.org/10.5334/cstp.740 Nodjoumi, G., Pozzobon, R., Sauro, F., Rossi, A.P.,

  17. [38]

    Earth and Space Science 10, e2022EA002278

    DeepLandforms: A Deep Learning Computer Vision Toolset Applied to a Prime Use Case for Mapping Planetary Skylights. Earth and Space Science 10, e2022EA002278. https://doi.org/10.1029/2022EA002278 Ronneberger, O., Fischer, P., Brox, T.,

  18. [40]

    Nat Astron 8, 8–9

    Using artificial intelligence to transform astrobiology. Nat Astron 8, 8–9. https://doi.org/10.1038/s41550-023-02159-7 Seelos, F.P., Seelos, K.D., Murchie, S.L., Novak, M.A.M., Hash, C.D., Morgan, M.F., Arvidson, R.E., Aiello, J., Bibring, J.-P., Bishop, J.L., Boldt, J.D., Boy...

  19. [41]

    Icarus 419, 115612

    The CRISM investigation in Mars orbit: Overview, history, and delivered data products. Icarus 419, 115612. https://doi.org/10.1016/j.icarus.2023.115612 Skilling, J., Bryan, R.K.,

  20. [43]

    https://doi.org/10.3389/frobt.2022.974397 Szepesvári, C.,

  21. [46]

    Attention Is All You Need. https://doi.org/10.48550/arXiv.1706.03762 Walmsley, M., Scaife, A.M.M., Lintott, C., Lochner, M., Etsebeth, V., Géron, T., Dickinson, H., Fortson, L., Kruk, S., Masters, K.L., Mantha, K.B., Simmons, B.D.,

  22. [47]

    Monthly Notices of the Royal Astronomical Society 513, 1581–1599

    Practical galaxy morphology tools from deep supervised representation learning. Monthly Notices of the Royal Astronomical Society 513, 1581–1599. https://doi.org/10.1093/mnras/stac525 Wei, C., Mangalam, K., Huang, P.-Y., Li, Y., Fan, H., Xu, H., Wang, H., Xie, C., Yuille, A., ...

  23. [48]

    https://doi.org/10.48550/arXiv.2304.03283 Xie, Z., Geng, Z., Hu, J., Zhang, Z., Hu, H., Cao, Y.,

    Diffusion Models as Masked Autoencoders. https://doi.org/10.48550/arXiv.2304.03283 Xie, Z., Geng, Z., Hu, J., Zhang, Z., Hu, H., Cao, Y.,

  24. [49]

    https://doi.org/10.48550/arXiv.2205.13543 Zeiler, M.D., Fergus, R.,

    Revealing the Dark Secrets of Masked Image Modeling. https://doi.org/10.48550/arXiv.2205.13543 Zeiler, M.D., Fergus, R.,

  25. [511]

    https://doi.org/10.1364/JOSA.62.000511 Gallinari, P., LeCun, Y., Thiria, S., & Fogelman-Soulie, F. (1987). Memoires associatives distribuees. Proceedings of COGNITIVA

  26. [1972]

    latent”, features in data as multi-dimensional vectors (a “latent representation

    for reconstructing sparsely occupied fields of view, or data that requires deconvolution of a blurring function, and can often achieve super resolution effects. A Shannon entropy is summed over pixels in an image to yield the most mathematically robust way to select a single (...

  27. [1984]

    Monthly Notices of the Royal Astronomical Society 211, 111–124

    Maximum entropy image reconstruction: general algorithm. Monthly Notices of the Royal Astronomical Society 211, 111–124. https://doi.org/10.1093/mnras/211.1.111 Stefanuk, B., Skonieczny, K.,

  28. [1998]

    Proceedings of the IEEE 86, 2278–2324

    Gradient-based learning applied to document recognition. Proceedings of the IEEE 86, 2278–2324. https://doi.org/10.1109/5.726791 Mantha, K.B., Roberts, H., Fortson, L., Lintott, C., Dickinson, H., Keel, W., Sankar, R., Krawczyk, C., Simmons, B., Walmsley, M., Garland, I., Make...

  29. [2000]

    Proceedings of the IEEE 88, 569–587

    Sampling-50 years after Shannon. Proceedings of the IEEE 88, 569–587. https://doi.org/10.1109/5.843002 Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.,

  30. [2011]

    Reinforcement Learning Algorithms for MDPs, in: Wiley Encyclopedia of Operations Research and Management Science. Wiley. https://doi.org/10.1002/9780470400531.eorms0714 Unser, M.,

  31. [2014]

    28 Springer International Publishing, Cham, pp. 818–833. https://doi.org/10.1007/978-3-319-10590-1_53

  32. [2015]

    Springer International Publishing, Cham, pp. 234–241. https://doi.org/10.1007/978-3-319-24574-4_28 Scharf, C.A., Mayer, M.H., Boston, P.J.,

  33. [2016]

    up sampling

    and “up sampling” to restore the original image size (e.g., by padding a higher resolution layer with repeated neighboring pixel values). The difference between that decoded output and the original input yields errors that can be backpropagated through the CAE layers to attemp...

  34. [2017]

    AEGIS autonomous targeting for ChemCam on Mars Science Laboratory: Deployment and results of initial science team use. Sci. Robot. 2, eaan4582. https://doi.org/10.1126/scirobotics.aan4582 Frieden, B.R.,

  35. [2018]

    2022, Xie et al

    has been shown to outperform earlier self-supervised methods (see e.g., Gao et al 2022, He et al. 2022, Xie et al. 2022). Furthermore, a variety of studies have shown that masked autoencoders perform particularly well on images (e.g., Huang et al. 2024, He et al

  36. [2019]

    https://doi.org/10.48550/arXiv.1810.04805 Dumoulin, V., Visin, F.,

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding. https://doi.org/10.48550/arXiv.1810.04805 Dumoulin, V., Visin, F.,

  37. [2020]

    EPSC2020-773)

    Mapping Mineralogical Distributions on Mars with Unsupervised Machine Learning (No. EPSC2020-773). Copernicus Meetings. https://doi.org/10.5194/epsc2020-773 Ho, J., Jain, A., Abbeel, P.,

  38. [2021]

    https://doi.org/10.48550/arXiv.2104.07448 24 Bickel, V.T., Moseley, B., Lopez-Francos, I., Shirley, M.,

    Maximum Entropy Auto-Encoding. https://doi.org/10.48550/arXiv.2104.07448 24 Bickel, V.T., Moseley, B., Lopez-Francos, I., Shirley, M.,

  39. [2022]

    2022, Huang et al

    and videos (e.g., Feichtenhofer et al. 2022, Huang et al. 2023). Not only is masking a valuable strategy for model architectures that include the CAEs discussed here, it appears to work well with more recent approaches such as diffusion-based modeling (e.g. Wei et al

  40. [2023]

    https://doi.org/10.48550/arXiv.2308.10794 Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.-M., Fu, D., Shen, X., Feng, J.,

    MGMAE: Motion Guided Masking for Video Masked Autoencoding. https://doi.org/10.48550/arXiv.2308.10794 Huang, Z., Jin, X., Lu, C., Hou, Q., Cheng, M.-M., Fu, D., Shen, X., Feng, J.,

  41. [2024]

    Monthly Notices of the Royal Astronomical Society 529, 732–747

    Astronomaly at scale: searching for anomalies amongst 4 million galaxies. Monthly Notices of the Royal Astronomical Society 529, 732–747. https://doi.org/10.1093/mnras/stae496 Feichtenhofer, C., Fan, H., Li, Y., He, K.,

  42. [5607]

    https://doi.org/10.1038/s41467-021-25882-z Chai, J., Zeng, H., Li, A., Ngai, E.W.T.,

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.