Pith. sign in

REVIEW 4 major objections 5 minor 13 references

Relighting can expose what visual encoders preserve about the physical world, and this paper shows that the strongest semantic encoders often degrade, rather than improve, relighting quality—a semantic–photometric trade-off that its ALI met

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 05:38 UTC pith:H4673YH4

load-bearing objection Solid relighting pipeline with an interesting probe idea, but the headline semantic-photometric trade-off is not cleanly supported by the paper's own Table 5. the 4 major comments →

arxiv 2602.01391 v2 pith:H4673YH4 submitted 2026-02-01 cs.CV

Relighting as a Probe of Visual Priors via Augmented Latent Intrinsics

Xiaoyan Xing , Xiao Zhang , Sezer Karaoglu , Theo Gevers , Anand Bhattad This is my paper
classification cs.CV
keywords relightingvisual priorslatent intrinsicsself-supervised learningsemantic-photometric trade-offdiffusion modelsmaterial-aware relightingvisual encoder probing
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tries to establish that image relighting can serve as a probe of what visual encoders preserve about the physical world, and that the encoders best at recognition are not the best at relighting: high-level semantic abstraction discards the pixel-aligned photometric detail needed to move specular highlights and shadows. The central discovery is a semantic–photometric trade-off, demonstrated with a controlled latent-intrinsic relighting model, and a method—Augmented Latent Intrinsics (ALI)—that fuses dense, pixel-aligned features from a distilled encoder with latent intrinsics and refines the result with self-supervision. If the paper is right, choosing a visual encoder for physical image-generation tasks should be guided by the encoder's training objective and pixel alignment, not by recognition benchmarks, and relighting can quantify what representations encode about light and materials.

Core claim

The paper reports that in a fixed latent-intrinsic diffusion relighting framework, swapping only the frozen visual encoder changes relighting quality in a way that is inverse to semantic strength: CLIP and DINOv2/DINOv3, optimized for invariance, systematically underperform MAE, trained to reconstruct pixels, and RADIO, a multi-teacher distilled encoder with pixel alignment, performs best. This is interpreted as evidence of a semantic–photometric trade-off: semantic abstraction discards dense geometry, material, and shading cues that light transfer requires. The proposed ALI method fuses RADIO features into latent intrinsics and, after a decoder-alignment and self-refinement stage, improves

What carries the argument

Latent intrinsics—lighting-invariant hierarchical scene features learned from multi-illumination image pairs without labels—form the base representation. ALI augments them by upsampling and concatenating intermediate feature maps from a frozen visual encoder into a pixel-aligned hypercolumn descriptor, then projecting that descriptor with a learnable layer and adding it to the intrinsic features. A three-stage schedule carries the argument: stage one trains the relighting encoder and projection layer under reconstruction, improved intrinsic-invariance, and hyperspherical-regularization losses; stage two freezes these and fine-tunes the diffusion decoder; stage three self-refines on pseudo-re

Load-bearing premise

The headline conclusion depends on attributing relighting differences to the encoder's training objective, but the compared encoders also differ in patch stride, resolution, and feature density, so if those confounds—not the objective—drive the gap, the semantic–photometric trade-off is overstated.

What would settle it

Take one encoder architecture and train two versions identical in patch size, input resolution, layer count, and feature dimension—one with a contrastive semantic objective, one with pixel reconstruction—then compare relighting RMSE under the same ALI pipeline; equal performance would falsify the claim that semantic objectives degrade relighting. A cheaper check is feeding DINOv3 features through the same hypercolumn alignment used for RADIO and matching feature stride before comparing.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If the central claim is correct, recognition scores are the wrong selection criterion for encoders used in physical image synthesis; an encoder's training objective and pixel alignment should be the primary criteria.
  • Relighting can serve as a reproducible probe: ranking encoders by their performance in a fixed latent-intrinsic relighting pipeline measures how much photometric structure they preserve.
  • Careful prior integration can substitute for scale and dense supervision: trained only on unlabeled real multi-illumination pairs, ALI matches or beats methods that use privileged inputs such as environment maps, G-buffers, or albedo.
  • The gains concentrate on glossy, metallic, and specular materials, implying that tasks involving view-dependent reflectance are where representation choice matters most.
  • The latent-intrinsic representation separates scene intrinsics from lighting codes well enough that interpolating the lighting code yields continuous, plausible lighting changes.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The trade-off likely extends to other inverse-graphics tasks—reflectance editing, material estimation, and view synthesis—since they share the need for pixel-aligned photometric cues; testing ALI-style fusion there would show whether the result is relighting-specific or general.
  • A controlled re-test that holds architecture and feature density fixed while varying only the pre-training objective (and, separately, augmentation strength) would isolate the causal variable behind the trade-off.
  • The authors' 'relightability' framing could be turned into a two-axis evaluation—recognition accuracy versus relighting RMSE/SSIM—giving model developers a cheap physical-fidelity check for any new encoder before building a generative decoder.
  • Because the self-refinement stage lowers MIIW scores while improving in-the-wild generalization, a dedicated out-of-distribution relighting benchmark would make the robustness gain visible instead of letting it appear as a regression.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 5 minor

Summary. The paper proposes Augmented Latent Intrinsics (ALI), a three-stage pipeline that fuses features from a frozen pretrained visual encoder into the LumiNet latent-intrinsic relighting framework. Stage I injects a hypercolumn descriptor from the encoder into the intrinsics encoder, Stage II fine-tunes the diffusion decoder to the augmented representation, and Stage III applies self-refinement on pseudo-relit pairs. The authors report state-of-the-art MIIW results among open-sourced diffusion methods and, more centrally, claim a 'semantic–photometric trade-off': encoders trained for semantic invariance (CLIP, DINO) degrade relighting quality, while reconstruction-trained or pixel-aligned encoders (MAE, RADIO) preserve photometric fidelity. The relighting task is thus proposed as a probe of what visual representations encode about the physical world.

Significance. If the trade-off claim were cleanly established, it would be a valuable finding for representation learning in physically grounded image generation, challenging the assumption that stronger semantic encoders uniformly benefit downstream tasks. The ALI method itself appears technically sound: the multi-stage ablations and material-wise evaluations support internal consistency, and the reported relighting improvements, if reproduced, would be a useful practical contribution. The independent concurrent work RAE (Zheng et al., 2025) also supports the direction that reconstruction-based representations can outperform DINO-style features under diffusion decoding, which reduces but does not eliminate concern about self-confirmation. However, the central probe conclusion is not supported by the current evidence, because the encoder comparison in Table 5 is confounded by decoder retraining and by uncontrolled architectural differences. The paper's conceptual contribution therefore needs substantial additional analysis before the headline claim can be accepted.

major comments (4)
  1. [§6, Table 5] Table 5's Stage I results contradict the central narrative. At Stage I, MAE has RMSE 0.2195, which is worse than DINOv3's 0.1794 and comparable to CLIP's 0.2189. MAE overtakes the semantic encoders only after Stage II, where the LumiNet diffusion decoder is fine-tuned for each encoder with a fixed 2-epoch budget. Since Stage II retrains the decoder separately for each feature type, the final ordering can reflect feature-conditioned optimization dynamics or decoder compatibility rather than the information content of the pretrained representation. To support the claim that the training objective is decisive, the authors should measure frozen-feature relightability with a single shared decoder and controlled conditioning, or otherwise show that the Stage II gains do not depend on feature-specific decoder adaptation.
  2. [§5, Table 5 and §6 'Stronger Semantic Encoders' paragraph] The encoder comparison conflates multiple factors: patch stride, input resolution, feature density, layer selection, and pretraining data. The DINOv2→DINOv3 transition only partially controls for feature density, and even that changes architecture and training scale. The claim that 'the training objective is decisive' is therefore not isolated by the presented experiments. The authors should add controlled comparisons, e.g., using matched patch sizes and channel widths, or explicitly ablating feature upsampling/density while holding the pretraining objective fixed. Without such controls, the semantic–photometric trade-off framing overstates what Table 5 can establish.
  3. [§5, Table 3] The user-study percentages in Table 3 are uninterpretable: each row sums to greater than 1. For example, in the 'Lighting alignment' row, 0.125 + 0.415 + 0.931 = 1.471, and the other rows similarly exceed 1. If the task was forced-choice among three methods, each row should sum to 1; if the task was 'select all that apply', the protocol and aggregation should be described. Since the paper invokes this table to support 'significantly better' human judgments, the correct percentages and a clear description of the forced-choice setup are needed. This may also affect the comparison with Latent-Intrinsics on identity preservation.
  4. [§4, §6; baselines in Tables 1–5] The probe is built entirely inside the authors' own LumiNet framework, and the strongest baselines (Latent-Intrinsic, LumiNet) are prior work by the same group. The observed encoder ranking could be specific to this latent-intrinsic architecture. The paper cites concurrent RAE as independent support, but RAE addresses reconstruction, not relighting. To reduce this circularity concern, at least one experiment should demonstrate that the same encoder ranking holds under a different diffusion-based relighting decoder, or the paper should explicitly frame the conclusion as 'within the LumiNet/ALI framework'. As written, the abstract and discussion generalize beyond the single architecture, and this generalization is load-bearing for the probe claim.
minor comments (5)
  1. [Title] The arXiv metadata title ('Relighting as a Probe of Visual Priors via Augmented Latent Intrinsics') differs from the title on the first page ('Stronger Semantic Encoders Can Harm Relighting Performance: Probing Visual Priors via Augmented Latent Intrinsics'). Please harmonize in the final version.
  2. [§4, Stage 1] The text states that exact layer indices and feature dimensions are provided in the supplementary material, but the submitted manuscript does not include this table. Please include the hypercolumn layer selection and dimensions, as they are needed to reproduce the Stage I fusion.
  3. [Table 1 abstract vs. table] The abstract states 'improving RMSE by 4.5% and SSIM by 4.9% over LumiNet'. Table 1 shows RMSE 0.294 vs 0.310 (raw) and 0.231 vs 0.240 (color-corrected); the 4.5% and 4.9% figures appear to be from the color-corrected block (3.75% RMSE and 4.9% SSIM). Please clarify which protocol the abstract refers to.
  4. [Table 4 and §5] The text says Stage III 'slightly reduces quantitative scores on the MIIW benchmark', but Table 4 shows mixed changes (e.g., Metallic PSNR increases from Stage I&II to All Stage, while Diffuse decreases). Specify the averaged protocol or note the exceptions to avoid confusion.
  5. [Appendix B.1] The material clustering relies on GPT-O3 to produce 'BRDF-style categories', but there is no validation of the clustering. Table S.1 also contains inconsistencies: Tile appears in both Glossy (27) and Uncertain/Mixed (27), and Granite/marble is placed in Uncertain. Please provide a manually verified mapping or a human-validated agreement score, as the material-wise claims depend on these categories.

Circularity Check

0 steps flagged

No circular derivation: ALI's central claims are empirical comparisons on MIIW; the trade-off is a measured correlation, not an equation that reduces to its inputs.

full rationale

The paper's derivation chain is empirical rather than definitional. ALI augments LumiNet's latent intrinsics with frozen visual features, fine-tunes the decoder, and evaluates on the MIIW test split. The 'semantic–photometric trade-off' is an observed inverse relation between ImageNet linear-probing accuracy and relighting PSNR—two externally measured quantities, not quantities defined in terms of each other. Table 5's encoder comparison is a standard representation-evaluation protocol: Stage II trains the same diffusion decoder for each feature type, so the final ranking is a measured outcome rather than a parameter fitted to the conclusion. The paper even reports Stage I numbers showing the ordering is not pre-wired (MAE trails DINOv3 at Stage I: RMSE 0.2195 vs 0.1794), which indicates a possible confound between encoder objective, spatial density, and decoder adaptation, but this is a correctness/interpretation concern, not circularity. Self-citations to Latent Intrinsics and LumiNet supply the base architecture, but the method is then validated against external baselines and the MIIW benchmark, and concurrent RAE (Zheng et al., 2025) is cited as independent support for the reconstruction-versus-semantic trend. No uniqueness theorem is imported, no ansatz is smuggled in via self-citation, and no prediction is equivalent to its fit by construction. The paper's own caveat that pixel metrics correlate weakly with perceived lighting quality further shows the authors do not treat the quantitative trade-off as a self-justifying definition.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 2 invented entities

The method relies on standard deep learning components (diffusion, latent intrinsics) and a few hand-chosen hyperparameters (λ, layer indices, mixing ratios). The main epistemic load is on the assumption that relighting metrics and ImageNet probing are valid proxies for physical fidelity and semantic strength, which the authors partially acknowledge.

free parameters (4)
  • Number of intermediate layers N for hypercolumn descriptor
    Chosen by hand; exact layer indices and feature dimensions deferred to supplementary material (Section 4). Affects fusion quality and is not justified independently.
  • Hyperspherical regularization temperature λ
    Appears in Eq. 4 (Lreg). Temperature parameter in the regularization loss, chosen without a stated criterion or sensitivity analysis.
  • Lighting zoo scenes sampled for Stage III = 1,000
    Randomly sampled from ~6,000 curated images (Appendix A.2); ad hoc choice for self-refinement, affects generalization claims.
  • Identity relighting mixing ratio in Stage III = 10%
    Introduced to enforce geometry preservation (Appendix A.1); chosen by hand without ablation in the main paper.
axioms (5)
  • domain assumption Diffusion denoising objective (DDPM) is a suitable generative model for image relighting.
    Used throughout (Eq. 6) without comparison to other generative families; the probe result is conditional on this choice.
  • domain assumption MIIW and BigTime datasets provide sufficiently diverse multi-illumination pairs to learn latent intrinsics.
    Training relies on these datasets; the method's generalization beyond them is claimed only qualitatively.
  • domain assumption Relighting metrics (PSNR/SSIM/RMSE) operationalize 'relightability' and physical fidelity.
    The authors themselves note these metrics correlate weakly with perceived lighting quality (Section 5), yet the trade-off plot uses PSNR.
  • domain assumption ImageNet-1K linear probing accuracy is a valid measure of semantic strength for the compared encoders.
    Used in Figure 1 (right) to support the trade-off; accuracies are taken from original papers and may not be directly comparable across training setups.
  • ad hoc to paper GPT-O3 clustering of material labels into BRDF-style categories is physically meaningful.
    Used to define material groups for Table 4; the grouping is tool-generated and not independently validated.
invented entities (2)
  • ALI (Augmented Latent Intrinsics) representation independent evidence
    purpose: Balances semantic context and dense photometric structure by fusing frozen visual encoder features with latent intrinsics.
    The representation's outputs are evaluated on public MIIW benchmarks and user studies, providing a falsifiable handle outside the paper.
  • Lighting Zoo (pseudo-relit image pairs) no independent evidence
    purpose: Provides self-supervision for Stage III refinement on in-the-wild images.
    Generated internally by the model; not publicly released, so its benefit is not independently verifiable beyond the paper's qualitative claims.

pith-pipeline@v1.3.0-alltime-deepseek · 15708 in / 11835 out tokens · 117149 ms · 2026-08-03T05:38:43.676713+00:00 · methodology

0 comments
read the original abstract

Image-to-image relighting requires representations that separate illumination from scene properties while preserving dense geometry, material, and photometric cues. We use this task as a probe of visual priors: unlike recognition tasks that reward invariance, relighting tests whether visual features retain the information needed for light transfer. Through a controlled generative relighting framework, we find that strong semantic encoders can degrade relighting quality, exposing a semantic--photometric trade-off between abstraction and physical fidelity. We introduce Augmented Latent Intrinsics (ALI), which balances this trade-off by fusing dense, pixel-aligned visual features into a latent-intrinsic relighting model and refining it with self-supervision on unlabeled real image pairs. ALI improves relighting quality, especially on glossy, metallic, and transparent materials, and demonstrates that generative relighting is an effective tool for quantifying what visual encoders encode about the physical world.

Figures

Figures reproduced from arXiv: 2602.01391 by Anand Bhattad, Sezer Karaoglu, Theo Gevers, Xiaoyan Xing, Xiao Zhang.

Figure 1
Figure 1. Figure 1: Stronger Semantic Encoders Can Harm Relighting Performance. Left: Visual comparison on a scene with complex specular materials. The task is to relight the input image (top-left) using the target illumination (bottom-left), which requires moving specular highlights from left to right, as indicated by the chrome sphere. While features from semantic encoders (CLIP, DINO) fail to reproduce realistic highlights… view at source ↗
Figure 2
Figure 2. Figure 2: Our Three-Stage Training Pipeline for Augmented Latent Intrinsics (ALI). Our method progressively adapts a pretrained visual encoder and fine-tunes a generative decoder for high-fidelity, unsupervised relighting. Stage I: Augmenting Latent Intrinsics. We inject semantic features from a frozen vision encoder into the intrinsics encoder. This creates our semantically-enriched ALI, which better disentangles s… view at source ↗
Figure 3
Figure 3. Figure 3: Qualitative comparison of relighting methods on challenging MIIW test scenes. Our image-to-image approach produces more physically plausible results than competing methods, many of which rely on privileged information like GT light maps, G-buffers, or albedo. The task is to relight the Input scene using the illumination from the Target lighting image. In the top row, competing methods render the metallic t… view at source ↗
Figure 4
Figure 4. Figure 4: Relighting comparison across two real-world images, each shown under two target illuminations (Light 1: lamp-dominated; Light 2: sunlight through windows). IC-Light (Zhang et al., 2025) produces stylized results with exaggerated glow and artifacts that diverge from the targets. Latent-intrinsics (Zhang et al., 2024a) captures some variation but yields low-contrast, flattened illumination with weak directio… view at source ↗
Figure 5
Figure 5. Figure 5: Multi-stage ablation. Top: Compared to LumiNet, our Stage I improves fine geometry details. Adding Stage II sharpens directional cues and specular effects, while the full pipeline (Stage I&II&III) produces the closest match to ground truth, with accurate shadows, highlights, and material fidelity. This progression illustrates how each stage contributes complementary improvements, consistent with the quanti… view at source ↗
Figure 6
Figure 6. Figure 6: Lighting interpolation and diversity. Top: Generated images showing a smooth interpolation between two lighting codes. Note the plausible evolution of directional lighting, including the progressive appearance of sharp specular highlights on the toaster and caustic effects from the bottle. Bottom: In-the-wild relighting results using lighting codes sampled from random, unpaired images. Our method produces … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

13 extracted references · 11 linked inside Pith

  1. [4]

    Radiov2.5: Improved baselines for agglomerative vision foundation models

    Heinrich, G., Ranzinger, M., Hongxu, Yin, Lu, Y ., Kautz, J., Tao, A., Catanzaro, B., and Molchanov, P. Radiov2.5: Improved baselines for agglomerative vision foundation models. InCVPR2025, 2024a. Heinrich, G., Ranzinger, M., Lu, Y ., Kautz, J., Tao, A., Catanzaro, B., Molchanov, P., et al. Radio amplified: Improved baselines for agglomerative vision foun...

    Pith/arXiv arXiv

  2. [5]

    Lightit: Illumination modeling and control for diffusion models.arXiv preprint arXiv:2403.10615, 2024a

    Kocsis, P., Philip, J., Sunkavalli, K., Nießner, M., and Hold- Geoffroy, Y . Lightit: Illumination modeling and control for diffusion models.arXiv preprint arXiv:2403.10615, 2024a. Kocsis, P., Sitzmann, V ., and Nießner, M. Intrinsic image diffusion for single-view material estimation. InCVPR, 2024b. Li, J., Li, H., and Matsushita, Y . Lighting, reflectan...

    Pith/arXiv arXiv

  3. [9]

    Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V ., Khungurn, P., and Suwajanakorn, S

    doi: 10.1145/3450626.3459872. Phongthawee, P., Chinchuthakun, W., Sinsunthithet, N., Raj, A., Jampani, V ., Khungurn, P., and Suwajanakorn, S. Dif- fusionlight: Light probes for free by painting a chrome ball. InCVPR,

    arXiv

  4. [12]

    Latent intrinsics emerge from training to relight

    Zhang, X., Gao, W., Jain, S., Maire, M., Forsyth, D., and Bhattad, A. Latent intrinsics emerge from training to relight. InNeurIPS, 2024a. Zhang, X., Jiang, R., Gao, W., Willett, R., and Maire, M. Residual connections harm generative representation learn- ing.arXiv preprint arXiv:2404.10947, 2024b. Zhang, X., Yunis, D., and Maire, M. Decipher- ing’what’an...

    Pith/arXiv arXiv

  5. [13]

    11 Stronger Semantic Encoders Can Harm Relighting Performance A

    URL https://arxiv.org/abs/ 2303.07634. 11 Stronger Semantic Encoders Can Harm Relighting Performance A. Implementation details A.1. Training All experiments are conducted on a node equipped with 8 NVIDIA A6000 Ada 48GB GPUs. The model is trained at a resolution of512ˆ512with an effective batch size of 64 (including gradient accumulation). For Stage I, we ...

    Pith/arXiv arXiv 2014

  6. [2014]

    F., and Black, M

    Bharadwaj, S., Feng, H., Becherini, G., Abrevaya, V . F., and Black, M. J. Genlit: Reformulating single-image relight- ing as video generation.arXiv preprint arXiv:2412.11224,

    arXiv

  7. [2018]

    Lightlab: Controlling light sources in images with diffusion models.arXiv preprint arXiv:2505.09608,

    Magar, N., Hertz, A., Tabellion, E., Pritch, Y ., Rav-Acha, A., Shamir, A., and Hoshen, Y . Lightlab: Controlling light sources in images with diffusion models.arXiv preprint arXiv:2505.09608,

    Pith/arXiv arXiv

  8. [2020]

    Masked autoencoders are scalable vision learners

    9 Stronger Semantic Encoders Can Harm Relighting Performance He, K., Chen, X., Xie, S., Li, Y ., Doll ´ar, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp. 16000–16009, 2022a. He, K., Chen, X., Xie, S., Li, Y ., Doll ´ar, P., and Girshick, R. Ma...

    Pith/arXiv arXiv

  9. [2021]

    URLhttps://arxiv.org/abs/2104.09886. Li, Z. and Snavely, N. Learning intrinsic image decomposi- tion from watching the world.CVPR,

    Pith/arXiv arXiv

  10. [2022]

    T., Metzen, J

    Xing, X., Hu, V . T., Metzen, J. H., Groh, K., Karaoglu, S., and Gevers, T. Retinex-diffusion: On controlling illumi- nation conditions in diffusion models via retinex theory. arXiv preprint arXiv:2407.20785,

    Pith/arXiv arXiv

  11. [2023]

    Liang, R., Gojcic, Z., Ling, H., Munkberg, J., Hasselgren, J., Lin, Z.-H., Gao, J., Keller, A., Vijaykumar, N., Fidler, S., and Wang, Z

    URL https://arxiv.org/abs/ 2211.10206. Liang, R., Gojcic, Z., Ling, H., Munkberg, J., Hasselgren, J., Lin, Z.-H., Gao, J., Keller, A., Vijaykumar, N., Fidler, S., and Wang, Z. Diffusionrenderer: Neural inverse and forward rendering with video diffusion models. InCVPR, June

    Pith/arXiv arXiv

  12. [2024]

    Generative models: What do they know? do they know things? let’s find out!arXiv preprint arXiv:2311.17137,

    Du, X., Kolkin, N., Shakhnarovich, G., and Bhattad, A. Generative models: What do they know? do they know things? let’s find out!arXiv preprint arXiv:2311.17137,

    Pith/arXiv arXiv

  13. [2025]

    Vincent, P., Larochelle, H., Bengio, Y ., and Manzagol, P.-A

    URLhttps://arxiv.org/abs/2508.10104. Vincent, P., Larochelle, H., Bengio, Y ., and Manzagol, P.-A. Extracting and composing robust features with denoising autoencoders. InProceedings of the 25th international conference on Machine learning, pp. 1096–1103,

    Pith/arXiv arXiv