acceptodds
Under review as a conference paper at ICLR 2027

LATEC: Learned Axial TEnsor Compression for High-Dimensional Scientific Data

Abstract

With the growing scale of scientific simulation and infrastructure, part of the burden has shifted from producing scientific data to analyzing and storing it. Compression is a key step to expanding storage capacity, but requires dedicated techniques to enforce user-specified pointwise error bounds that ensure the validity of downstream scientific workflows. To further push the limits of compression, novel learned codecs have been developed, in particular for images and videos, but these methods often optimize for aggregated distortion and fail to meet error bound standards. Moreover, these methods often target fixed field dimensionality, failing to account for many scientific data use cases, or at the expense of large adaptation or retraining costs. In this work, we introduce the Learned Axial TEnsor Compression (LATEC) method, an error-bounded codec with a dimension-agnostic learned predictor based on a GNN architecture. LATEC is pretrained on synthetic 2D and 4D fields, and evaluated with and without domain fine-tuning on 13 scientific fields in 3D to 5D. Results show that LATEC is consistently competitive with SZ3, achieving a 10.8% average improvement and greater consistency across all domains than other compressors. These results demonstrate that a shared learned predictor can serve as a single error-bounded codec across tensor dimensionalities, competitive with specialized compressors, while retaining pointwise error guarantees.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
scientific data compression, learned compression, graph neural network, dimension-agnostic neural networks

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_RbtsOpenReview ↗
Rating
4(+1)
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

LATEC is a learned compression method for scientific data with varying dimensionality. It combines neural prediction with classical compression techniques to improve compression efficiency while maintaining pointwise error bounds. A shared model trained on synthetic data can be applied across different dimensionalities and scientific domains without retraining. Evaluation on diverse scientific datasets shows competitive compression performance, with further gains from domain-specific fine-tuning.

Critical strengths

Practical relevance and generality. I see scientific data compression as an important problem that deserves more attention from the deep learning community. LATEC uses a shared architecture across dimensionalities, while many existing compression frameworks that incorporate deep learning assume a fixed input dimensionality. This makes it easier to apply to different scientific datasets without redesigning the model. I see this work as a useful starting point for encouraging further research in this practically important area.

Synthetic pretraining and broad evaluation. I also like that a model trained on synthetic data works well on real scientific datasets, including unseen dimensionalities, without target-domain fine-tuning. I particularly appreciate the evaluation across diverse datasets and several established compressors. Bringing these comparisons together takes considerable effort and provides stronger evidence of the method’s generality and competitive performance under pointwise error constraints.

Critical weaknesses

Limited practical efficiency. The compression gains come with a substantial computational cost: LATEC is considerably slower than SZ3 for both encoding and decoding. Given that scientific compression often aims to reduce I/O costs, I would like to better understand when the bitrate savings outweigh this overhead. This limits the practical benefit demonstrated by the current results.

Modest gains on real high-dimensional data. While the shared model handles unseen dimensionalities, its gains over SZ3 on the two real 5D datasets are limited. This raises the question of whether training a model specifically for the input data at compression time could yield larger compression gains. The throughput implications are unclear, however, so comparing both compression performance and total encoding time would help clarify this trade-off.

Questions for authors
  1. Synthetic pretraining and data-specific adaptation. Could the limited gains on the real 5D datasets reflect a mismatch between the synthetic training distribution and the structure of these fields? How much does domain-specific fine-tuning close this gap? If the gains remain limited, could training a model specifically for the input data at compression time help? A comparison or discussion of approaches such as NeurLZ, considering both compression performance and throughput including training time, would help clarify this trade-off.

  2. Practical efficiency. In which use cases would LATEC’s bitrate savings outweigh its substantially lower encoding and decoding throughput? An estimate or an end-to-end example would help clarify when it would be preferable to SZ3.

  3. Benefit of gating. Figure E.5 shows how often each predictor is selected. Could you also clarify how much bitrate reduction scale-guided gating provides compared with using the GNN alone?

  4. Model capacity and efficiency. The reported results use a fixed GNN architecture. Have you explored how model size affects compression performance and throughput? In particular, could a smaller model retain most of the compression gains while reducing the encoding and decoding cost?

4 comments

We are grateful to the reviewer for the encouraging assessment and for very precise questions. They led to the new gating, model-width, per-input experiments and the scale shift improvement.

W1, Q2. Practical efficiency. With the Triton and Numba implementation (see W1 in the global response), the gap to SZ3 is much smaller than the 12-24× of the submission (Sec. 3, Fig. 9, App. E.4). On one H100, LATEC encodes at 220-270 MiB/s from 2-D to 4-D and 153 MiB/s in 5-D. SZ3 on 8 CPU threads encodes 1.4-1.5× faster in 3-D and 4-D, and decodes 4-5× faster. TTHRESH, the only baseline that runs natively in 5-D and 6-D, encodes 5-7× slower than LATEC there.

LATEC is therefore practical for the many workloads, in particular when data is compressed once and rarely read. Simulation checkpoints are a typical example: they are decompressed only if the simulation must be restarted, so rate matters much more than decoding time.

W2, Q1. Domain gap, fine-tuning, and per-input training.

  • GyroSwin. As explained in the global response (W2), the limited gain came from an evaluation error. Corrected, pretrained LATEC saves 27.4% against SZ3, and fine-tuning improves this to 30.8%.
  • Gysela. Gysela is genuinely out of distribution: 55% of its values pile up near the minimum (vs 5.9% in our synthetic data generator), and the pretrained scale head over-estimates its residuals by about 2.8×. Fine-tuning on spatially disjoint crops of the same run (no other run exists) raises LATEC's saving against SZ3 from 3.0% to 16.8% (Table 2, App. C.3).
  • Per-input training (NeurLZ-style). We thank the reviewer for the suggestion. We fine-tune for k steps on the input itself and count the weight delta (81-141 KiB) in the stream:
Field (points)total bpp change at k = 1000 (RME 1e-3 / 1e-4)training timecompression time
Gysela (134M)−7.2% / −5.9%29 min20 s
GyroSwin (11M)−6.5% / −5.9%15 min1.6 s
MHD-64 (17M)−1.8% / −1.0%2.7 min2.6 s

Per-input training helps only after many steps, and its training time is far larger than the compression time; after 50-200 steps, the stream is up to 20% larger. Fine-tuning once per dataset gives comparable or larger gains at no cost per input. This experiment is now in App. C.4.

Q3. Gating versus the GNN alone (new Figs. 6 and 8, App. E.2). We thank the reviewer for suggesting this ablation. We compare the GNN alone with LATEC (GNN + gate) on crops at RME 1e-3. Values are % bpp vs SZ3; lower is better:

HurricaneMirandaS3DMHD-64RT-Inst.GyroSwinSynth. 5D
LATEC−26.7−18.8−13.3−22.6−43.0−26.2−55.6
w/o gating−25.5+8.3+773.8−22.0−41.1−22.1−55.6
  • Smooth fields: gating is essential. On S3D the gate removes 90% of the bits of the GNN alone, and on Miranda 25%.
  • Other fields: there the GNN alone already beats SZ3, and the gate saves a further 0-5.3%, in particular in smooth areas of the field.
  • Spatial pattern (new gate maps, Fig. 6). The GNN is chosen where the field is nonlinear (Gray-Scott spots, the RT mixing interface), and cubic or linear interpolation in smooth regions.

Q4. Model capacity (new App. E.2.4, Table E.3). To evaluate the impact of model capacity, we retrained the model at widths 16 and 32 with the full recipe and ran the full benchmark, excluding Gysela. Speed-ups are relative to width 64, on the H100 setup of Fig. 9. The last column is the mean % bpp vs SZ3 over these 12 fields (width 64 averages −18.1% over all 13 fields of Table 2); per-field values are in Table E.3. Lower is better:

WidthEnc. / dec. speed-up, 3-DEnc. / dec. speed-up, 5-DMean % bpp vs SZ3 (12 fields, no Gysela)
161.06× / 1.42×1.57× / 1.82×+83.4
321.01× / 1.27×1.59× / 1.47×−16.1
64 (shipped)1× / 1×1× / 1×−19.4

Width 32 keeps most of the gain (4.5% more bits than width 64), but it is only 1.0-1.6× faster to encode and 1.3-1.5× faster to decode. Runtime is dominated by per-point work that does not depend on the width, such as building each point's causal context. Width 16 is far too small: it uses more bits than SZ3 on four fields, and 11× as many bits on S3D. Width 32 remains a useful trade-off when decoding speed matters most, and we will release it alongside the width-64 model in the LATEC code.

Thank you for the detailed response and the additional experiments. I appreciated that the per-input fine-tuning experiment accounted for both training time and the cost of storing weight updates.

I’d still be interested in a direct comparison with NeurLZ. It trains a small residual model for each input block, which can keep both training and inference fast. The NeurLZ paper reports compression gains, and I’m curious how those gains compare with LATEC at the same error bound. Could you compare the two on the same data, reporting the final bitrate including model weights, total encoding time including training, and decoding time?

The checkpointing example addressed my concern about LATEC’s lower throughput relative to SZ3. When checkpoints are rarely read back, smaller files can matter more than faster decoding. The other experiments addressed my remaining concerns. Thank you again.

We thank the reviewer for their feedback. To answer your question, we also compared LATEC with NeurLZ. On average, LATEC saves 40.4% vs 35.9% for NeurLZ. We did not find any open source implementation of the algorithm, so we follow NeurLZ's own protocol on the same data to be able to reuse the reported results in the paper (table 2): the same NYX, Miranda, and Hurricane fields, and the same error bounds (1E-2, 5E-3, 1E-3, 5E-4, 1E-4). We compute the bit-rate saving over SZ3 v3.1.8 (while our benchmark used v3.4) at equal PSNR, averaged over the 5 bounds. Here are the detailed results:

MethodNyx TemperatureNyx Dark Matter DensityNyx Velocity YNyx Baryon DensityMiranda DiffusivityMiranda Velocity ZMiranda ViscocityHurricane CloudHurricane PrecipHurricane WMean
NeurLZ41.8467.9229.5375.6631.7226.8831.1211.7118.6923.3835.85
LATEC43.8470.7933.1043.9715.1326.1016.5250.0057.4447.0540.39

NeurLZ encoding times are derived from the ratios in its Table 3 (Nyx 512^3), applied to our SZ3 times. Training is extrapolated linearly from 10 to 100 epochs. Miranda and Hurricane are not in Table 3, and the ratios grow for smaller fields, so those estimates are likely low.

DatasetSZ3NeurLZ (est.)LATEC
Encode
Nyx3.8~3606.7
Miranda1.18~1102.2
Hurricane0.64~601.9
Decode
Nyx1.6~3.04.8
Miranda0.53~1.01.4
Hurricane0.31~0.61.0
Hardware
DeviceCPU8x RTX 6000 Ada1x RTX 5070

Note that NeurLZ timings are reported from their setup, using 8 GPUs while LATEC only uses 1.

Thank you for the additional experiments. The higher encoding cost of online training makes sense and matches my intuition. LATEC’s bitrate savings are impressive, and I’ve raised my rating to 4.

Reviewer_bKyjOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces LATEC, a compression method designed for scientific data. Based on a coarse-to-fine predictive coding framework, LATEC incorporates a dimension-agnostic axial graph neural network. By utilizing shared neural network modules to aggregate neighborhood information across different axes of a tensor, it predicts both the mean value and the residual distribution scale for each data point, thereby enabling generalization across dimensions. Compared to traditional compression methods, LATEC exhibits significantly slower encoding and decoding speeds, while achieving an improvement in bitrate efficiency of approximately 10%.

Critical strengths
  1. The paper is generally well-written.
  2. The dimension-agnostic GNN design is sound; furthermore, this design enables the model to learn effective scientific data representations through pre-training on synthetic data, allowing for the transfer of pre-trained models from 2D/4D data to 3D/5D data.
  3. The experimental evaluation is comprehensive, and the method demonstrates superior performance.
Critical weaknesses
  1. The scope of dimensional generalization is limited; for instance, testing has only been conducted for 5D, and the gains observed in these 5D experiments are not particularly significant. Given the emphasis on cross-dimensional generalization, why were 6D or higher dimensions not tested? Furthermore, why is there no systematic analysis of how model performance evolves as the number of dimensions increases?
  2. The inference speed is significantly slower than that of SZ3. While the overall bitrate savings amount to only 10%, the inference time is 10 to 20 times longer.
  3. The contributions of individual modules appear incremental. Many of these modules are adapted from existing methods; the authors should more clearly articulate what their core, fundamental innovation actually is.
Questions for authors

see the Weaknesses

1 comment

We thank the reviewer for the positive assessment of the writing, the dimension-agnostic design and the evaluation. The questions on dimensionality led directly to a new appendix and the scale shift improvement.

W1. 6-D and beyond, and the trend with D.

  • The 5-D gains were understated in the submission. As explained in the global response (W2), the near-parity on GyroSwin came from a loader error. Corrected, LATEC saves 27.4% against SZ3 (30.8% fine-tuned), and 48.6% on Synthetic 5D. Gysela is genuinely out of distribution: pretrained LATEC saves 3.0% there, and fine-tuned LATEC saves 16.8% against SZ3 (Table 2).
  • New App. E.3 (Fig. E.7) studies the trend with D. The same pretrained weights code synthetic fields from D = 2 to 6, and the gain against SZ3 grows with D:
D23456
bpp saving vs SZ3 (%, higher is better), RME 1e-312.416.322.426.029.9
bpp saving vs SZ3 (%, higher is better), RME 1e-49.011.816.416.220.2
  • Real 6-D data (App. E.3, Table E.4). On stacked GyroSwin timesteps, native 6-D LATEC uses 32.8% fewer bits than SZ3 at RME 1e-3, and 16.1% fewer at 1e-4. On this dataset, coding the time axis natively saves only a further 0.4-3.5% over coding each timestep separately in 5-D.
  • Beyond 6-D. Nothing in the architecture depends on D. We did not go further because we know of no real dataset above 6-D, but we are happy to evaluate LATEC on higher-dimensional data if the reviewer can point us to a dataset.

W2. Speed versus savings. The trade-off is better than in the submission on both sides:

  • Savings. The mean saving against SZ3 is now 18.1% (median 17.9%), up from 10.8%. As explained in the global response (W2), this includes the GyroSwin correction and the per-stage scale shift.
  • Speed. With the Triton and Numba implementation, LATEC encodes at 220-270 MiB/s from 2-D to 4-D and 153 MiB/s in 5-D (Fig. 9). SZ3 on 8 CPU threads encodes 1.4-1.5× faster than LATEC in 3-D and 4-D and decodes 4-5× faster, instead of 12-24× and 13-27× in the submission, while in 2-D LATEC encodes 1.5× faster than SZ3. TTHRESH, the only baseline that runs natively in 5-D and 6-D, encodes 5-7× slower than LATEC there.

LATEC remains overall slower than SZ3, but still practical. It best suits data that are compressed once and then stored for long periods or read rarely, where rate matters more than coding time.

W3. What is new. We see the main contribution as conceptual: to our knowledge, LATEC is among the first to show that a dimension-agnostic, prior-fitted network (PFN), trained once on synthetic fields only, is useful for error-bounded compression (see also global response, W3). The revised Sec. 2 now states explicitly which parts are inherited: the coarse-to-fine residual coding of the SZ family (dyadic schedule, quantizer, escapes), and the learned per-point scale of Ballé et al. (2018). The three components we add are each supported by new evidence:

  1. Axial GNN predictor: with the same learned scale, the GNN mean saves a further 7-33% over classical predictor selection on the real fields other than the very smooth S3D, and the savings also grow with D = 2…6 (App. E.3).
  2. Scale-guided gating: it controls, point by point, which predictor is used, at a cost of 16 bits per chunk-stage (Figs. 6 and 8).
  3. Synthetic pretraining: without any target data, LATEC is the best pointwise error-bounded codec on 12 of the 13 fields, and fine-tuning from it lowers the rate on all adapted datasets (Table 2, Fig. 7).

We also release an open, modular Benchopt benchmark (13 fields, five baselines), where new fields or codecs plug in without touching the rest.

Reviewer_3eAqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces Learned Axial TEnsor Compression (LATEC), an error-bounded codec for high-dimensional scientific tensors. It combines a dimension-agnostic axial GNN predictor with coarse-to-fine residual coding and scale-guided selection between learned and classical predictions. The GNN predicts reconstruction means and residual scales, while quantization and escape coding control pointwise reconstruction error in normalized coordinates. A single model is pretrained on synthetic 2D and 4D fields and evaluated on 13 scientific fields spanning 3D–5D, with additional fine-tuning experiments. At matched relative maximum error (RME) targets, pretrained LATEC reports an average 10.8% reduction in bits per point (BPP) relative to SZ3.

Critical strengths
  1. The paper tackles an important problem: building an error-bounded compressor that can transfer across tensor dimensionalities. Using a GNN to exploit structured neighborhood information is an intuitive design choice, and the overall approach is easy to follow.

  2. The reported compression results are promising. Compared with SZ3, LATEC achieves an average BPP reduction of 10.8% and a median reduction of 9.8% across 13 fields, with improvements on 10 of them (Table 2). The chunked design also supports memory scalability. For a 4 GiB input, peak host RAM is 10.5 GiB for LATEC versus 43.5 GiB for TTHRESH (Section 3).

Critical weaknesses
  1. I would appreciate a clearer explanation of why an axial GNN is needed and how it differs from existing prediction-based compressors. The approach includes a GNN with 102,710 trainable parameters (embedding width 64) and considers 288 gating configurations per chunk-stage. Although the design is carefully described, it remains unclear whether a simpler predictor using comparable local context could achieve similar BPP. A controlled comparison would make the benefit of the GNN architecture easier to assess.

  2. The compression gains come with a substantial throughput penalty. At RME = 10^-4, Table 3 reports encoding approximately 12–24 times slower and decoding 13–27 times slower than SZ3. Although the authors acknowledge this trade-off and prioritize compression rate, it is still unclear which practical workloads justify the additional computation. The maximum 47.3% saving on Synthetic 5D is encouraging, but it does not by itself establish the same benefit for typical archival workloads.

Questions for authors
  1. Could the authors clarify the main technical differences between LATEC and earlier prediction-based compressors such as SZ3 and HPEZ, particularly the contribution of the axial GNN?

  2. Given the throughput gap in Table 3, which scientific workloads do the authors expect to benefit most from LATEC, especially when compressed data must be accessed frequently?

  3. Could the authors compare the axial GNN with a simpler MLP-based predictor under comparable local-context, model-capacity, quantization, and entropy-coding settings? This would help clarify whether the GNN design is necessary.

  4. Could the authors provide a break-even analysis for a representative archival workload, quantifying whether the storage savings from the improved compression ratio can offset the substantially higher encoding and decoding costs? In particular, under what assumptions about data volume, retention period, and access frequency would LATEC become more cost-effective than existing compressors such as SZ3?

If the authors can address my concerns, I would be willing to increase my score.

1 comment

We thank the reviewer for their constructive review and for acknowledging our promising results. As the reviewer proposed, we added a controlled comparison to the paper (Fig. 8) and give a break-even analysis below, which clarify the paper's positioning.

W1, Q1. How LATEC differs from SZ3 and HPEZ. SZ3 and HPEZ predict each point using simple multidimensional interpolators, then code the quantized residuals with Huffman coding and Zstd. LATEC keeps their error-bounded loop, but our axial GNN predicts a mean and a scale for each point. The scale drives the entropy coder and gates, point by point, between the GNN and classical interpolation, at a cost of 16 bits per chunk-stage: the GNN is chosen in nonlinear regions and interpolation in smooth ones (Fig. 6).

Beyond these components, the main contribution is conceptual. To our knowledge, LATEC is among the first to show that a dimension-agnostic, prior-fitted network, trained once on synthetic fields only, is useful for error-bounded compression. The same weights, pretrained on 2-D and 4-D, compress real 3-D to 5-D data without fine-tuning, and they are a good starting point when target data exist (Fig. 7).

W1, Q3. Is the GNN needed? Would a simpler predictor do? The GNN is needed for two reasons:

  1. Dimension agnosticism. An MLP over a local context of width $k$ takes $k^D$ inputs, so each D needs its own architecture and training, and it cannot run on unseen dimensions. For this reason we did not train a dense MLP predictor. The axial GNN shares its weights across axes, so one model runs for any D (Sec. 1).
  2. Rate. Instead, the new ablation (Fig. 8) isolates the learned mean: it replaces only the GNN mean with the best classical interpolation, keeping the same learned scale, gate and entropy coder. Values are % bpp vs SZ3 on crops at RME 1e-3; lower is better:
HurricaneMirandaS3DMHD-64RT-Inst.GyroSwinSynth. 5D
LATEC−26.7−18.8−13.3−22.6−43.0−26.2−55.6
w/o GNN mean−6.7−12.8−13.5−4.3−15.2−10.5−17.5

The GNN mean saves a further 7-33% of the bits on five of the six real fields, and 46% on Synthetic 5D. The only exception is S3D, which is smooth enough for interpolation; there the classical mean is 0.3% cheaper.

W2, Q2. Which workloads benefit. With the Triton and Numba implementation, the gap to SZ3 is much smaller than the 12-24× of the submission (Sec. 3, Fig. 9, App. E.4). We now time LATEC on one H100 and the baselines on 8 CPU threads, at RME 1e-4. LATEC encodes at 220-270 MiB/s from 2-D to 4-D, 153 MiB/s in 5-D and 108 MiB/s in 6-D, and decodes at 187 MiB/s in 2-D down to 78 MiB/s in 6-D. Compared with the baselines:

  • SZ3 encodes 1.4-1.5× faster than LATEC in 3-D and 4-D, and decodes 4-5× faster. In 2-D, LATEC encodes 1.5× faster than SZ3.
  • SPERR encodes and decodes faster than LATEC in 3-D, but in 2-D it encodes 16× slower and decodes 6× slower.
  • TTHRESH, the only baseline that runs natively in 5-D and 6-D, encodes 5-7× slower than LATEC there and decodes 2.2-2.8× slower.

LATEC reconstructs a 1 GiB tensor in 7-10 s from 3-D to 5-D. It suits data that are compressed once and rarely read, such as simulation checkpoints, which are decompressed only if the simulation must be restarted, and long-term archives. For data that are read frequently, SZ3 remains preferable.

W2, Q4. Break-even analysis. We compare the storage saved per month with the one-off extra GPU encoding cost, plus the extra decoding cost of each full read. The assumptions are:

  • LATEC timed on a consumer RTX 5070 at about USD 0.10 per GPU-hour (owned or rented), and SZ3 on the same hardware
  • parallel file system (PFS) storage at USD 0.03 per GiB-month
  • tape storage at USD 0.0015 per GiB-month

Under these assumptions:

  • 4-D and 5-D fields (MHD-64, RT-Instability, GyroSwin) repay their extra encoding cost within 1-3 months on a PFS, and within about 2-5 years on tape. Over 5 years, they stay cheaper than SZ3 as long as the data are fully read fewer than 30-64 times.
  • 3-D fields in our benchmark repay it more slowly: Hurricane in 7-14 months, and Miranda in 14 months at RME 1e-3. As these fields are smoother, their low bpp leaves little to save.
  • For any storage price: LATEC pays off within 5 years whenever storage costs more than USD 0.0006-0.0016 per GiB-month.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Global Response

We sincerely thank the reviewers for their thoughtful reviews. To our knowledge, LATEC is one of the first learned, pointwise error-bounded compressors in which a single model runs on scientific tensors of any dimensionality. Overall, all reviewers pointed out that this is an important problem, with a sound design and promising results on a broad evaluation. The main weaknesses they note are the throughput of the method and its gain on high dimensional data, as well as better justification for the design choices. The experiments they prompted substantially strengthened the paper. Below, we answer these shared concerns, in particular with improved throughput and expanded ablation studies.

W1. Throughput (3eAq, bKyj, Rbts). We agree that this point was the main weakness of LATEC, but we were able to improve it significantly during rebuttal. We rewrote the GPU kernels in Triton and parallelized the entropy coder with Numba. LATEC encodes 1.5× faster than SZ3 in 2-D, and SZ3 encodes only 1.4-1.5× faster in 3-D and 4-D, and TTHRESH, the only baseline that runs natively in 5-D and 6-D, encodes 5-7× slower than LATEC there. We updated the throughput comparison and reporting for all methods in the revised manuscript (see updated Fig 9, app E.4). LATEC now suits many scientific workloads, and most particularly data compressed once and rarely read such as checkpointing.

W2. Gains on real high-dimensional data (bKyj, Rbts). We run extra experiments to better show performance evolution with the field dimension. On synthetic fields, the gain grows with D, from 12.4 % (2-D) to 29.9 % (6-D), and LATEC saves 32.8 % on an additional real 6-D GyroSwin (App. E.3). Moreover, the reviewers' questions on the gap between synthetic and real data led us to add a per-stage scale shift to the entropy coder (Eq. 7), which corrects a scale that is systematically too large or too small on a stage. It improves LATEC on most datasets. We also realized that the on-par performances on GyroSwin actually came from a loader error (wrong real/imaginary axis, timestep t=0). Once corrected, LATEC saves 27.4 % vs SZ3. Together, these changes raise the mean saving from 10.8 % to 18.1 %, and LATEC is now the best bounded codec on 12 of 13 fields.

W3. Contributions and their necessity (3eAq, bKyj, Rbts). We agree that LATEC builds on top of existing blocs: LATEC inherits the error-bounded loop of the SZ family and the learned-scale entropy model of learned image compression. Our contributions are what we add to this loop:

  1. A dimension-agnostic GNN predictor. It is, to our knowledge, one of the first learned predictors for pointwise-bounded compression that runs, with one set of weights, on tensors of any number of axes in their native shape, ablated in Fig. 8.
  2. Scale-guided gating. The predicted scale selects, point by point, between the learned and the classical prediction, for 16 bits per chunk-stage (Fig. 6).
  3. Synthetic pretraining. A generator with no governing equation is the only training signal. It transfers to real 3-D to 6-D data (App. E.3), and is a better initialization than random for fine-tuning (Fig. 7).

We hope that these clarifications and performance improvement convince the reviewers of the relevance of LATEC as a novel dimension agnostic and pretrained learned compressor for scientific data, and we are ready to run further analyses during the discussion if deemed necessary.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.