acceptodds
Under review as a conference paper at ICLR 2027

AIF-TVR: Active Inference for Time-Varying 3D Flow Field Reconstruction with Sparse High-Resolution Observations

Abstract

Reconstructing time-varying 3D flow fields from sparse high-resolution observations requires selecting informative frames and incorporating their evidence into reconstruction. We propose AIF-TVR, an active-inference-inspired framework that couples sequential observation acquisition with conditional diffusion reconstruction. A degradation-aware temporal belief model propagates acquired evidence across timesteps. An observation-predictive planner evaluates hypothetical belief updates to select frames using an expected-free-energy-inspired objective. The acquired observations then supervise adaptation of the diffusion reconstructor to recover the complete sequence. Experiments on three flow datasets demonstrate lower MSE and MAE than the evaluated baselines, with PSNR gains of 1.99–4.61 dB over the strongest reported baseline on each dataset.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Active Inference, Time-Varying Flow Field Reconstruction, Sequential Observation Acquisition, Conditional Diffusion Models

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_pGBUOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper reconstructs a time-varying 3D flow from the full low-resolution sequence when only B<TB<T high-resolution frames can be acquired. AIF-TVR selects those frames with an active-inference-inspired score over a time-factorized Gaussian belief, fine-tunes a conditional diffusion reconstructor on the acquired pairs, and reconstructs every frame. On three flow datasets at 4×4\times, it reports the better performance over the strongest baseline.

Critical strengths
  1. The acquisition problem is distinct from standard super-resolution.

  2. Selection and adaptation are separated, and the main limitations are stated.

Critical weaknesses

1. Belief updates never affect a later decision

Because the belief factorizes over time and qϕq_\phi depends only on the LR sequence, updating acquired frame aa leaves IkI_k and RkR_k unchanged for every unacquired frame (Section 3.3; Appendix A.4). The procedure is therefore a fixed per-frame ranking with candidate removal and a coverage term, rather than sequential belief-guided acquisition. This limitation is decisive for the headline setting B=1B=1: there is no later decision for the belief update to affect, so Table 1 does not test sequential updating.

2. The score is not connected to reconstruction gain, and its construction can favor uninformative frames

The score measures expected belief revision, not the downstream reduction in full-sequence reconstruction error, and no experiment relates the two. At B=1B=1, this missing link could be tested directly by comparing the score with each candidate's hindsight reconstruction gain. Moreover, the initializer and qϕq_\phi use the same input and target; when they agree, Proposition 2 reduces the information term to a predicted-variance ranking. With the preference set to the current belief and C(a)=0C(a)=0, the active-inference formulation is not shown to provide more than an uncertainty or expected-information-gain heuristic.

3. The experiments do not show that the acquisition rule produces the reported gains

Table 1 compares complete systems and therefore cannot isolate the acquisition rule; matched selection baselines under a fixed reconstructor and adaptation protocol are missing. The ablation does not resolve this because M3†M3^\dagger comes from a separate run. Its 5.17–5.94 dB gains are much larger than the −1.75-1.75 to +0.74+0.74 dB changes in the controlled variants, so they cannot be attributed to a specific component. The reported 43.68 dB one-frame result in Appendix C.2 is also not reconciled with the headline B=1B=1 results. These comparisons use a single seed.

4. The theory does not address reconstruction quality, and the physical results are mixed

The propositions characterize the Gaussian update and estimation error, while Proposition 3 gives a generic descent condition requiring alignment with an unavailable full-sequence gradient. None connects the acquisition score to reconstruction quality. The physical diagnostics are also mixed: vorticity error improves, but normalized energy error increases by 430% on Re6400 and 22% on ShockVortex. Thus the evidence supports selected metric improvements, not a broad physical-fidelity claim.

Questions for authors
  1. If acquiring a frame leaves IkI_k and RkR_k unchanged for every remaining candidate, what makes selection sequential, and what does the belief update contribute to the B=1B=1 results?
  2. At B=1B=1, what is the rank correlation between GG and each candidate’s hindsight reconstruction gain, and does the rule outperform random, uniform, maximum-variance, and LR-entropy selection with the same reconstructor and adaptation?
  3. When do the initializer and qϕq_\phi disagree, and, with C(a)=0C(a)=0 and the preference set to the current belief, what does the score add beyond ranking by latent variance?
  4. Why does M3†M3^\dagger gain 5–6 dB when every controlled variant changes PSNR by at most 0.74 dB, and how is the 9.1 dB drop after one-frame adaptation in Appendix C.2 consistent with the B=1B=1 result?
1 comment

Critical Weaknesses:

1.At B = 1, hypothetical belief updates enter the score used to select the sole HR frame. The observed HR value updates only the selected marginal after that decision, and the acquired LR–HR pair supervises the reconstructor evaluated in Table 1. At larger budgets, the remaining candidates retain their base scores while temporal coverage and eligibility change. With those rules held fixed, precomputed analytic scores and round-by-round scoring yielded identical choices in all 117 decisions.

2.At B = 1, we scored every eligible candidate using LR and independently adapted the same reconstructor on each corresponding HR frame. Across three seeds, Spearman’s ρ between −G and PSNR gain ranged from 0.746 to 0.789 on ShockVortex, from −0.335 to −0.035 on Tangaroa, and from −0.366 to −0.126 on Re6400. When the predictor matches the observation distribution implied by the Gaussian belief, the exact information term reduces to Gaussian mutual information, proportional to the sum of log(1 + vⱼ/r) over latent dimensions. In the general case, differences in predictive means and variances can change both expected belief revision and predictive risk, so candidates with similar average latent variance can receive different scores.

3.Table 1 compares complete systems. To estimate the contribution of frame selection, we fixed the reconstructor and adaptation protocol and compared policies at B = 1 across three paired seeds. Relative to maximum-variance selection, the full policy changed mean PSNR by +0.688 dB on Tangaroa, +0.180 dB on Re6400, and 0.000 dB on ShockVortex. Results against other baselines varied by dataset. The 5.17 and 5.94 dB values in Figure 4 compare M3† from Table 1 with frozen baselines from a different protocol, so they cannot isolate the effect of frame selection or adaptation. Appendix C.2 measures a 9.11 dB within-run decrease under its own one-frame protocol, as detailed in our response to Q4.

4.Proposition 3 gives a sufficient descent condition that depends on alignment with the full-sequence gradient, which is unavailable during acquisition. The link to reconstruction gain is empirical. On ShockVortex, Spearman’s ρ between the acquisition score and hindsight PSNR gain ranged from 0.746 to 0.789 across three seeds. In Figure 6, AIF-TVR reduced median relative vorticity error compared with CD-TVD on all three datasets. Normalized kinetic-energy error decreased on Tangaroa and increased on Re6400 and ShockVortex. We will describe these two diagnostics separately in the revised manuscript.

Questions:

1.The policy builds the acquisition set one frame at a time. Each choice changes the eligible set and temporal coverage, while the remaining candidates retain their exact information and risk scores. At B = 1, hypothetical updates enter the initial acquisition score, and the acquired LR–HR pair is used to adapt the reconstructor. The real posterior update occurs after the sole selection.

2.At B = 1, we independently adapted the same pretrained reconstructor on each eligible HR frame. On ShockVortex, Spearman’s ρ between −G (higher acquisition priority) and hindsight PSNR gain was 0.746, 0.789, and 0.782 across three seeds. The correlation estimates were negative on the other two datasets. On ShockVortex, the full policy improved mean PSNR over random, uniform, and LR-entropy selection by 0.170, 0.163, and 0.061 dB, respectively, and tied maximum-variance selection at 40.366 dB. The policy’s relative performance varied across datasets.

3.The predictor qφ and the belief-implied observation distribution disagree when their means differ (m ≠ μ) or their variances differ (s ≠ v + r). Both discrepancies were largest on ShockVortex in the candidate sweep. Even with C(a) = 0 and the current belief as the preference, expected revision depends on s and (m − μ)², and predictive risk penalizes KL divergence from the current belief (Eqs. 20 and 22). These terms can distinguish candidates with the same latent variance. At B = 8 on Tangaroa, the full policy achieved 38.307 dB, compared with 38.167 dB for variance ranking and 37.773 dB for Gaussian mutual information.

4.The 5.17 and 5.94 dB values in Figure 4 subtract the frozen PSNR of the controlled ablations from M3† results obtained under the Table 1 protocol. These cross-run differences cannot measure an adaptation gain. We will report gains from matched before-and-after comparisons in Figure 4. In Appendix C.2, 1,000 steps of adaptation on Re6400 frame 4 reduced PSNR from 52.79 to 43.68 dB on the same set of 130 unobserved frames. Both Table 1 and Appendix C.2 use B = 1, but the C.2 decrease is measured within a separate run with its own checkpoint, HR frame, and evaluation set.

Reviewer_1i5hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a framework for reconstructing time-varying 3D flow fields from a complete sequence of low-resolution (LR) frames and a limited budget of high-resolution (HR) frame acquisitions. To determine which HR frames to acquire, the method maintains a temporal belief initialized from the full LR sequence and sequentially selects frames based on predicted observations and hypothetical belief updates. The acquired LR–HR pairs are then used to adapt a conditional diffusion model for super-resolving the entire 3D flow sequence. Experiments on three flow datasets demonstrate superior reconstruction performance in terms of PSNR compared with existing methods.

Critical strengths
  1. The problem of actively acquiring informative observations under a limited budget is novel and practical, with potential applications in scientific fields such as deformation measurement where high-resolution data acquisition is expensive.
  2. The framework integrates belief-guided frame acquisition with conditional diffusion reconstruction, providing a unified solution for budget-constrained super-resolution.
  3. Experiments on three flow datasets demonstrate promising reconstruction performance compared with existing methods.
Critical weaknesses
  1. The paper treats 3D flow as an independent data type, but its underlying dynamics are largely neglected in the method formulation. The method resembles an application of conditional super-resolution and active frame acquisition from video processing. If no essential modifications are required for other 3D scalar fields, it may be more appropriate to target general time-varying 3D field reconstruction.
  2. In Section 3.3 (page 5, lines 259-263), the authors state that newly acquired HR observations do not affect the belief-based acquisition scores of other candidates, raising the question of what advantage sequential acquisition offers over collectively selecting a batch of temporally diverse frames.
  3. The related work in the Introduction (pages 1-2) should better highlight the limitations of existing methods and the specific contributions of this paper, rather than simply listing previous studies.
  4. The compared methods are CD-TVD (2025), SRGAN (2017), and SSR-VFD (2020), and I suggest including at least one more recent diffusion-based or neural-operator-based flow super-resolution method.
  5. The ablation studies are insufficient to demonstrate the effectiveness of belief-guided acquisition, particularly without controlled comparisons against random, uniform, uncertainty-based, and diversity-aware frame selection strategies.
Questions for authors
  1. Could the authors clarify which aspects of the method specifically exploit 3D flow dynamics and whether it can be directly generalized to other time-varying 3D fields?
  2. The proposed method measures information value through belief KL divergence, but a larger belief revision does not necessarily imply lower reconstruction error; could the authors provide evidence linking the acquisition score to actual reconstruction improvement?
  3. Could the authors report the computational and memory overhead of belief-guided acquisition compared with simpler selection strategies, given the budget-constrained setting?
  4. Is it possible to extend the proposed framework to a streaming setting where LR frames are also acquired sequentially, without access to the complete LR sequence in advance?
2 comments

Critical Weaknesses: 1.The temporal models learn dependencies from LR flow sequences, and the reconstructor learns spatial structure from historical flow fields. The acquisition score has no explicit fluid governing equations or conservation constraints. We will clarify that the formulation applies to budgeted reconstruction of time-varying 3D fields and that our empirical evaluation focuses on flows.

2.On ShockVortex at 4× spatial super-resolution and (B=8), we held the reconstructor and adaptation protocol fixed. AIF-TVR outperformed the KTV batch selector (Song et al., AAAI 2026), adapted to LR flow features, in all three paired runs (40.235 vs. 40.111 dB on average). Its mean PSNR was within 0.002 dB of the temporal-diversity baseline (40.237 dB). With the full LR sequence available, the analytic selector can plan the entire batch before HR acquisition by combining fixed LR-based scores with temporal coverage updated as each frame is added (see #P1t5 Q2).

3.We will clarify two limitations of prior work in the Introduction. Most flow super-resolution methods focus on reconstruction from given observations, leaving open which HR time steps to acquire under a limited budget. CD-TVD uses LR entropy to select an HR frame. This measures the complexity of the LR input and provides an indirect proxy for the frame’s value to full-sequence reconstruction. We will use these distinctions to explain the scope and contribution of AIF-TVR in the revised Introduction.

4.We compared the conditional diffusion reconstructor in AIF-TVR with an adaptation of the 3D masked diffusion method of Amorós-Trepat et al. (2026, doi:10.1063/5.0304492) on ShockVortex at 4× spatial super-resolution. Both models were trained for 5,000 updates on frames 0–147 and evaluated on frames 201–251 with the same LR inputs and 128 × 64 × 64 crops. Using the same checkpoint and 100 masked-DDIM steps, masked diffusion achieved 16.76 dB from pure noise and 27.89 dB from noised LR, with the latter starting timestep selected on held-out validation frames. The frozen AIF-TVR reconstructor achieved 33.96 dB. This single-seed comparison uses only LR at test time. Table 1 uses one acquired HR frame for test-time adaptation.

5.We compared random, uniform, maximum latent variance, LR entropy, temporal diversity, and feature diversity across all three datasets at HR budgets of 1, 4, and 8. Each comparison used the same reconstructor checkpoint, adaptation loss, 100 optimization steps, and evaluation frames, with three paired seeds. On ShockVortex at B = 1, AIF-TVR achieved 40.366 dB, exceeding random and uniform selection by 0.170 and 0.163 dB, respectively, and tying maximum latent variance. At B = 8, AIF-TVR achieved 40.235 dB. Temporal and feature diversity achieved 40.237 and 40.186 dB, respectively. The relative performance varied across datasets and budgets.

Questions: 1.The initializer and observation predictor learn temporal dependencies from LR flow sequences. The same acquisition formulation can be applied to other time-varying 3D fields by adapting the data representation and degradation operator and training the predictor and reconstructor on suitable historical data. Its performance on other field types remains to be validated.

2.On ShockVortex, we used LR inputs to score all 51 eligible frames. We then adapted the same pretrained reconstructor independently on each candidate’s HR frame at B = 1 and measured the resulting full-sequence PSNR gain over the frozen model. Across three seeds, Spearman’s ρ between the full acquisition score (−G₀) and PSNR gain ranged from 0.746 to 0.789. The rank correlations were the same for MSE reduction. For the belief-revision term alone, ρ ranged from −0.464 to −0.414.

3.We measured selection overhead on 51 ShockVortex frames using an RTX 3090 and five repeats per setting. At B = 1, 4, and 8, median decision times were 0.799, 2.051, and 3.583 s for the eight-sample Monte Carlo selector and 3.91, 3.95, and 4.27 ms for the closed-form selector. Random and uniform selection took at most 0.50 ms, and maximum-variance selection took 2.32–2.48 ms. Including HR reads, encoding, and belief updates, Monte Carlo selection took 1.250, 3.878, and 7.132 s at the three budgets. The peak increase in allocated GPU memory was about 134 MiB for this stage and 1.4 MiB for closed-form selection. At B = 8, the two implementations differed on the final selected frame.

4.AIF-TVR can select HR frames using only the LR frames received so far. We replayed 51 ShockVortex frames in chronological order. At each window boundary, the initializer and observation predictor used the available LR prefix to request one HR frame from that window, assuming the frame could be retrieved from a buffer. We adapted the reconstructor and reconstructed the sequence after the stream ended. With four HR frames, mean PSNR across three paired seeds was 35.744 dB for prefix-based selection and 35.739 dB for selection with access to the full LR sequence.

Reviewer_1i5hOpenReview ↗

Questions partially clarified, but not strong enough

Thank you for the detailed response and the additional experimental evidence. I appreciate the authors' efforts to clarify the problem formulation and address my concerns. However, my primary concern remains with the underlying motivation and methodological design, rather than the experimental performance alone.

W1/W3/Q1: Regarding the claimed research gap, I would characterize the cited prior works as addressing different problem settings, rather than being insufficient for budget-constrained HR reconstruction with controllable frame acquisition. I agree that the latter is a more practical and challenging problem. However, given the scarcity of directly comparable baselines, I would expect a more thorough analysis of the problem itself, including its distinctive challenges and how the proposed methodological design addresses them. I encourage the authors to strengthen this conceptual justification rather than relying primarily on experimental improvements.

W2: My original concern was primarily about the performance benefit of sequential acquisition over a natural batch extension of the proposed strategy. Since the authors acknowledge that the acquisition scores are largely determined by the complete LR sequence, it remains unclear what advantage sequential selection provides. I suggest including a controlled ablation comparing the proposed sequential strategy with its natural batch counterpart, in addition to the KTV batch selector (Song et al., AAAI 2026).

W4: I appreciate the additional comparison with a recent diffusion-based method. This experiment helps clarify the reconstruction capabilities of the proposed framework.

W5: The additional selection baselines are helpful. However, the reported PSNR differences are generally marginal, and the proposed strategy does not consistently outperform simpler alternatives, particularly temporal diversity. This raises concerns about the practical significance of the proposed acquisition mechanism.

Q2: While the reported rank correlation supports the predictive value of the full acquisition score, the negative correlation between the belief-revision term alone and reconstruction improvement raises questions about its individual contribution. Could the authors clarify whether the predictive power of the full score primarily comes from other components? An ablation isolating the contribution of the belief-revision term would be particularly helpful.

Q3: I appreciate the detailed computational measurements. However, given the relatively small reconstruction gains over simpler selection strategies, the additional computational overhead should be justified by a clearer performance benefit. In particular, it would be useful to evaluate the trade-off between selection cost and reconstruction quality.

Q4: The reported streaming result is interesting, as prefix-based selection slightly outperforms selection with access to the complete LR sequence (35.744 versus 35.739 dB). Does this suggest that the streaming setting can perform as well as, or even better than, the offline setting? Given the small difference, I encourage the authors to examine whether it is statistically meaningful and provide a more detailed analysis of the experimental design.

Overall, although the rebuttal addresses several of my experimental concerns, I remain unconvinced that the proposed acquisition mechanism offers a sufficiently clear and substantial advantage over simpler alternatives. My main concern is the lack of a compelling methodological justification for the proposed design, especially given the relatively modest empirical gains.

Comments on the paper

3 comments

Critical Weaknesses:

1.Under Eqs. (6)–(11), the acquired HR value updates only the belief at the selected time, which is then removed from the candidate set. Later decisions use the remaining LR-conditioned information and risk scores together with the updated temporal coverage and eligibility. Under identical rules, precomputed and round-by-round analytic scoring yielded the same choices in all 117 tested rounds. For example, on ShockVortex at B = 8 (seed 2026), both selected frames 244, 239, 232, 251, 226, 207, 210, and 219 in that order.

2.We reran the full M3 and its ablations with the same reconstructor checkpoint, HR budget, and adaptation schedule within each comparison, using three paired seeds. At (B=1), matched M3 achieved 37.891, 33.966, and 40.366 dB on Tangaroa, Re6400, and ShockVortex, respectively. CW3 reports the separate effects of the acquisition rule and the contrastive term. In the revised Table 2 and Figure 4, we will use the matched M3 for the controlled comparisons. Q2 reports the precomputed-score control and the Monte Carlo posterior-update ablation.

3.At B = 1, we evaluated the acquisition policy and the contrastive term separately using the same reconstructor and adaptation schedule across three paired seeds. With L1 plus contrastive loss fixed, the full policy changed mean PSNR relative to LR-entropy selection by −0.241, +0.072, and +0.061 dB on Tangaroa, Re6400, and ShockVortex, respectively. With the selected frames fixed, adding the contrastive term at weight 0.1 changed mean PSNR relative to pure L1 by −1.738, +0.075, and −0.081 dB, respectively. We will report these dataset-dependent effects separately in the revised manuscript.

4.Table 1 uses one HR frame per sequence. The recovered logs record a learning rate of 10⁻⁴ for pretraining and 10⁻⁵ for test-time adaptation. The Tangaroa run uses its best-validation SR checkpoint, frame 182, and 1,000 updates with L1 plus a contrastive term weighted by 0.1. The recovered Re6400 trace records its best-validation SR checkpoint, frame 26, and 1,000 updates. We have yet to verify that this trace produced the exact Table 1 metric. For ShockVortex, the historical comparison reporting 40.05 dB identifies frame 226. The separate 100-step trace on frame 244 leaves the adaptation schedule and checkpoint for the 40.05 dB entry unverified. Appendix C.1 describes a separate 100-step controlled protocol. The recovered belief checkpoints use separate temporal Gaussian encoders for initialization and observation prediction, both with a latent width of 128. The predictor estimates observation variance from the data. A fixed observation-noise variance for the Table 1 runs remains unverified. The fixed r = 0.01 applies to our new Gaussian VFE controls.

Questions:

1.Changing the acquired HR value at a fixed time leaves the next choice unchanged. The update acts on the selected time factor, and no equation passes the observed value to the remaining factors. In nine controlled substitutions of the acquired HR latent, the remaining candidates’ information and risk scores were unchanged, and all nine tests returned the same next frame as their corresponding controls.

2.For the analytic version of M3, precomputing the base scores and recomputing them at each round gave identical scores and frame selections across all 117 tested rounds. We matched checkpoints, HR budgets, adaptation schedules, seeds, coverage, and eligibility rules. For the original eight-sample selector at (B=8), we separately switched off the post-acquisition belief update while keeping coverage, eligibility, and random draws fixed. The selected frames and reconstruction metrics were identical in all nine dataset–seed pairs.

3.Table 1 uses B = 1, with one acquired HR frame per sequence. The recovered logs record learning rates of 10⁻⁴ for pretraining and 10⁻⁵ for test-time adaptation. The 100-step schedule in Appendix C.1 is for the controlled ablation. CW4 gives the dataset-specific settings we could verify and identifies the remaining gaps in the historical record.

Thank you for the candid response and matched experiments. These improve transparency but confirm my central concern: acquired HR values do not influence subsequent selections, and the policy produces the same choices as precomputed scoring with matched coverage and eligibility rules. The new controls also show inconsistent acquisition gains and no consistent benefit from the contrastive term. Furthermore, the exact provenance and adaptation settings of several Table 1 results remain unverified. Consequently, the motivation for sequential belief-guided acquisition and the evidence for its specific contributions remain insufficient, and I am maintaining my score of 2.

Thank you very much for your further feedback, we will address your concerns.

  1. Our framework uses the LR sequence to predict possible HR observations and assess their expected belief revision and predictive risk. Frames are then selected sequentially with updated temporal coverage and eligibility constraints. Under the factorized belief and fixed observation predictor, acquired HR values do not change the remaining candidates’ base scores; consequently, precomputed scores with the same constraints yield identical selections. The acquired HR content affects full-sequence reconstruction through subsequent adaptation of the shared reconstructor. We will clarify that the acquisition contribution lies in observation-predictive scoring and coverage-aware set construction.
  2. We evaluated frame selection while keeping the reconstructor and adaptation settings fixed. On ShockVortex, an independent evaluation of all 51 candidate frames gave Spearman correlations of 0.746, 0.789, and 0.782 between acquisition utility (−G) and the resulting PSNR improvement across three seeds. At B=1, our policy improved mean PSNR over random, uniform, LR-entropy, and feature-diversity selection by 0.170, 0.163, 0.061, and 0.476 dB, respectively. At B=8, the corresponding gains were 0.029, 0.027, 0.012, and 0.049 dB. Each comparison favoured our policy in all three paired seeds. Together, these results support the use of the acquisition score to identify HR observations that benefit reconstruction. The contrastive term provides an auxiliary constraint on the learned representations during adaptation to the target sequence, alongside the voxel-wise L1 loss. Our main contribution concerns the allocation of a limited HR acquisition budget. All controlled selection comparisons use the same reconstructor and adaptation objective, allowing us to assess the contribution of the selected observations. We will clarify the contrastive term’s role as an optional adaptation component and report its dataset-dependent effects separately from the acquisition results.
  3. Table 1 adopts 4x spatial super-resolution, with a pre training learning rate of 10 ⁻⁴. One HR frame (B=1) is collected from each sequence, and the random seed is 2026. Adaptation starts from the pre trained SR checkpoints corresponding to each dataset, with a learning rate of 10 ⁻⁵ and a loss of L1 weighted comparison term of 0.1. The adaptation steps are all 1000.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.