acceptodds
Under review as a conference paper at ICLR 2027

Preserve Broadly, Observe Adaptively: Budget-Decoupled Event-Point Learning for Action Recognition

Abstract

Event cameras capture motion as asynchronous events, making them well suited to action recognition under rapid motion and challenging illumination. However, event recordings vary in cardinality, while existing point-based pipelines commonly form fixed-size network inputs by randomly sampling events from each recording or temporal clip before learning. This one-shot reduction couples event retention to per-pass computation, irreversibly discarding evidence from dense recordings while padding or duplicating events in short ones. We introduce EventVista, a preservation-aware framework that decouples the retention budget from the per-pass computation budget through three advances. First, Authentic Candidate Preservation preserves authentic evidence beyond a single forward pass without increasing per-pass computation by retaining a broader, recording-wide candidate pool while maintaining the natural cardinality of short recordings. Second, Coverage-Aware Runtime Observation constructs bounded, temporally complementary views from this candidate pool and processes them while preserving their actual cardinalities. Adaptive Evidence Integration then sequentially aggregates predictions from the shared network and stops the observation process once the accumulated prediction becomes sufficiently confident and stable.Across THU-EACT-50-CHL, DVSGesture, and DailyDVS-200, EventVista achieves state-of-the-art Top-1 accuracies of 70.40% ( pp), 99.35%, and 50.68% ( pp), respectively.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
event cameras, event-based vision, action recognition, point-based representation learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_h7hJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EventVista, an event-based action recognition framework that decouples the event retention budget from the computation budget of a single forward pass. Authentic Candidate Preservation retains a broader pool of original events, Coverage-Aware Runtime Observation constructs temporally complementary views under a bounded per-view budget, and Adaptive Evidence Integration aggregates predictions with confidence- and stability-based early stopping. Experiments on THU-EACT-50-CHL, DVSGesture, and DailyDVS-200 report Top-1 accuracies of 70.40%, 99.35%, and 50.68%, respectively. The paper includes component ablations, comparisons of observation strategies, and analyses of computational budgets and adaptive inference.

Critical strengths

1.The paper identifies a meaningful limitation of conventional fixed-size event-point processing, where coupling evidence retention with the per-pass computation budget may discard useful motion information prematurely. The proposed retain–observe–stop framework provides a coherent approach to preserving and selectively processing complementary event evidence while accommodating variable-length recordings.

2.The empirical evaluation covers three benchmarks with different event characteristics and reports strong recognition performance. Component ablations, budget sensitivity analyses, comparisons of fixed-view and adaptive inference, and results from three independent runs provide substantial evidence for the potential effectiveness of the framework.

Critical weaknesses

1.Unclear stopping protocols in the single-criterion ablations. Section 3.3 specifies a minimum of two views for the proposed joint stopping rule. However, Table 4 reports mean view counts below two for the confidence-only and stability-only variants on DVSGesture, whose candidate pools all exceed the per-view budget according to Appendix Table 7. These controls may use different stopping rules, but their minimum-view requirements and initialization are not clearly specified. In particular, the stability-only result requires clarification because the stated stability measure compares consecutive predictions. This ambiguity limits the reproducibility and interpretation of the stopping-rule ablations.

2.Incomplete end-to-end efficiency evaluation. Although EventVista maintains a bounded per-pass input budget and reports network GFLOPs, multiple views are generally processed per recording, and the computation estimates exclude candidate construction. The paper does not provide a complete comparison of preprocessing time, end-to-end inference latency, and peak memory consumption under matched resource budgets. Consequently, the practical efficiency advantages of the proposed framework remain insufficiently established.

Questions for authors

1.Section 3.3 defines the proposed joint stopping rule after at least two views, while Appendix Table 7 shows that all DVSGesture test recordings have 48K candidates, exceeding the 24K per-view budget. Why do the confidence-only and stability-only variants in Table 4 report average view counts of 1.86 and 1.72? Were these variants allowed to stop after one view, and if so, how was the stability criterion defined for the first view?

2.Table 2 reports noticeable accuracy differences between packed batching and per-recording execution. Were these variants evaluated using the same trained checkpoint and identical input views, or were they trained separately? Could the authors clarify which implementation or training differences account for these results?

3.Could the authors provide end-to-end inference latency, candidate-construction time, and peak memory consumption, together with comparisons against representative baselines under matched computational budgets?

4 comments
AuthorsOpenReview ↗

Response to Reviewer h7hJ Part (1/2)

We sincerely thank you for your careful review and positive evaluation of our paper, and for fully recognizing its focus on the problem of prematurely discarding discriminative information in fixed-point-count processing, its coherent Retain–Observe–Stop decoupled framework, and its solid empirical performance across three benchmark datasets! Your suggestions on the stopping protocol, batch training, and practical efficiency have helped us further clarify the method's operating conditions and empirical basis. We respond point by point below.

W1 & Q1: Initialization of the stopping policy and minimum number of observations

Thank you for carefully checking the stopping protocol. Both single-criterion ablations in Table 4 allow an exit after the first view; the official Joint policy checks for stopping only after at least two observations for recordings whose candidate counts exceed the single-view budget. The first-view branch of the original Stability-only policy uses a confidence gate and does not compute stability from a single prediction.

Let m(k)m^{(k)} denote the difference between the two highest predicted probabilities after aggregating the first kk views. The stability condition is:

\left\lVert p^{(k)}-p^{(k-1)}\right\rVert_1\le0.10,\qquad \arg\max p^{(k)}=\arg\max p^{(k-1)}.$$ To clarify the initialization and stopping rules of each strategy, we separately list below the exit conditions for the first view and from the second view onward. | Strategy | First-view exit condition | Exit condition from the second view onward | DVSGesture average number of observations | | --------------------------- | ------------------------------------ | -------------------------------------------- | ----------------------------------------: | | Confidence-only | $m^{(1)}\ge0.70$ | $m^{(k)}\ge0.70$ | 1.86 | | Stability-only | $m^{(1)}\ge\tau_{\mathrm{first}}$ | $S^{(k)}$ is satisfied | 1.72 | | Joint (paper configuration) | Not allowed | $m^{(k)}\ge0.70$ and $S^{(k)}$ is satisfied | 2.12 | All strategies continue up to $K_{\max}=5$ if their respective conditions are not satisfied. All DVSGesture candidate pools contain 48K events, exceeding the 24K per-view budget, so the means below two here **indeed arise from the ablation strategies allowing an exit after the first view and are unrelated to the exception for processing short recordings in full once**. Joint retains at least two observations because **high confidence after the first view does not guarantee that the prediction is already reliable**: a single bounded view may miss discriminative events, and subsequent complementary views still have the opportunity to supplement evidence and correct predictions. Therefore, the formal method first obtains subsequent evidence and then jointly assesses confidence and stability. We have added the above initialization and stopping conditions to the revised manuscript to make their correspondence with the official Joint configuration in Section 3.3 clear. ## Q2: Packed Batching vs. Per-recording Execution Thank you for focusing on the specific settings of this control. **The two settings in Table 2 are trained separately and use their own selected checkpoints. The purpose of this control is to examine how batch training can learn more effectively from variable-length event data while preserving the true event cardinality of each recording.** Specifically, Per-recording training computes the loss and updates parameters after processing each recording; Packed training directly packs the variable-length inputs of 8 recordings, averages the batch loss, and then performs a single update. The two therefore have different effective batch sizes, parameter-update frequencies, and optimization trajectories. Packed training uses cumulative offsets to maintain each recording's boundaries during feature extraction and pooling, allowing event recordings of different lengths to participate jointly in batch training while keeping recordings isolated, without adding points or duplicating events to equalize input lengths. These results support our use of Packed training: **it combines genuinely variable-length event inputs with batch learning and achieves consistent recognition benefits on all three datasets under the current training settings.** This implementation allows EventVista's preservation of event cardinality to extend throughout data processing and model training, without forcing all recordings to have the same number of points to form batches.
AuthorsOpenReview ↗

Response to Reviewer h7hJ Part (2/2)

W2 & Q3: End-to-End Efficiency and Resource Consumption

Thank you for requesting a full-pipeline evaluation. Joint reduces end-to-end latency with essentially unchanged peak GPU memory and achieves higher accuracy in budget-matched comparisons.

1. End-to-end latency

Measurements use an RTX 3090, FP32, batch size 1, one CPU thread, TF32 disabled, and 20 warm-up iterations. Cached includes view construction, network inference, and prediction integration from prebuilt candidate pools; Raw E2E additionally includes candidate construction from raw events. Both use the same profiling recordings. Latencies are per-recording means in ms.

DatasetScopeFixed-5 (ms)Joint (ms)Reduction
THUCached557.16515.757.43%
THURaw E2E3,522.483,446.252.16%
DVSGestureCached861.21459.8446.61%
DVSGestureRaw E2E1,010.07605.3440.07%
DailyDVS-200Cached873.74819.866.17%
DailyDVS-200Raw E2E1,149.621,090.365.15%

Joint reduces Raw E2E latency by 2.16%, 40.07%, and 5.15% on THU, DVSGesture, and DailyDVS-200, respectively. THU's smaller end-to-end gain reflects its dominant candidate-construction cost.

2. Preprocessing overhead and peak memory

Raw−Cached is the difference between mean Raw E2E and Cached latencies, estimating additional preprocessing overhead rather than independently timing candidate construction. View construction is already included in Cached. Candidate construction occurs once before observation; its timing variation is not attributed to early stopping. CPU memory is process peak RSS; times are ms and memory is MiB.

DatasetMethodRaw−CachedView constructionPeak GPUPeak CPU RSS
THUFixed-52,965.328.97173.731,052.66
THUJoint2,930.508.25173.741,111.44
DVSGestureFixed-5148.8616.11173.70903.79
DVSGestureJoint145.5010.25173.70962.38
DailyDVS-200Fixed-5275.8815.33173.885,175.98
DailyDVS-200Joint270.5013.84173.895,237.61

Peak GPU memory remains approximately 174 MiB, differing by at most 0.01 MiB. Joint increases peak CPU RSS by approximately 59–62 MiB while reducing view construction time and inference latency.

3. Budget-matched controls

Coverage Static MV and Joint share the encoder, candidate pool, and single-view cap. Static inference uses predetermined integer view-count allocations, independent of test labels and runtime predictions, to match Joint's average encoder calls over the full test set. Short recordings are processed once and included in the mean. Strong Wider increases encoder channel width with a 24K input cap; Strong More Points raises the input cap to 48K.

Each cell reports test Top-1 accuracy (%) / network GFLOPs per recording, excluding candidate construction. Under this estimation convention, both stronger single-view alternatives differ from Joint by at most approximately 1.1%.

DatasetCoverage Static MVStrong WiderStrong More PointsEventVista Joint
THU68.93 / 10.1868.75 / 10.0869.12 / 10.2470.40 / 10.18
DVSGesture98.21 / 14.3798.38 / 14.2398.59 / 14.4299.35 / 14.37
DailyDVS-20049.66 / 20.8849.50 / 20.6549.99 / 20.9450.68 / 20.88

Joint improves over Coverage Static MV by 1.47, 1.14, and 1.02 percentage points, respectively, and outperforms both stronger single-view alternatives. These results support allocating limited computation to complementary evidence according to each recording's prediction state.

Once again, we sincerely thank you for recognizing the research value and potential of our paper and for your important contributions to improving experimental rigor and system-level efficiency evaluation! We have revised the paper according to your suggestions. If the above protocol clarifications and end-to-end efficiency evidence have fully addressed your concerns, we respectfully ask you to consider further raising your score to support acceptance of this paper; your recognition and a higher score would be tremendous encouragement and support for our team!

AuthorsOpenReview ↗

Sincere Thanks and Gentle Follow-up on Our Response

Dear Reviewer h7hJ,

We hope everything is going smoothly with your research. We sincerely appreciate your positive assessment of our framework and your constructive questions about stopping initialization, packed training, and practical efficiency. Your suggestions helped us clarify the execution protocol and strengthen the evaluation with measured end-to-end latency and resource-consumption results.

With the discussion period closing soon on November 18 (AoE), we wanted to gently follow up on our point-by-point response above. If you have a moment, we would be grateful to hear whether the protocol clarifications and additional efficiency measurements address your questions. We remain available to provide any further clarification that would be helpful.

If our responses and additional evidence have addressed your main concerns, we would be very grateful if you could consider raising your score. Your recognition and support would mean a great deal to our team and provide tremendous encouragement for our continued research.

Thank you again for your thoughtful guidance and encouragement. Wishing you continued success in your research!

Warm regards,

The Authors

AuthorsOpenReview ↗

Warm Thank You and Quick Summary of New Experiments & Clarifications

Dear Reviewer h7hJ,

Thank you again for your positive assessment of our framework and your constructive suggestions. Your questions helped us make the execution protocol clearer and strengthen the practical efficiency evaluation. To make our detailed response easier to review, we summarize the three main points below:

  1. Stopping Initialization and Minimum Observations: For candidate pools exceeding the per-view budget, the official Joint policy evaluates at least two views before checking its confidence and stability conditions; shorter pools are processed once in full. The single-criterion ablations explicitly allow first-view exits, with Stability-only using a confidence gate at initialization. This explains their DVSGesture averages below two calls without computing stability from a single prediction.

  2. Packed Training and Variable-Length Inputs: The packed and per-recording settings are trained separately and use their own validation-selected checkpoints. Per-recording training updates after each recording, whereas packed training averages the losses of eight recordings before updating; the comparison therefore reflects different effective batch sizes and optimization schedules. Cumulative offsets preserve recording boundaries throughout feature extraction and pooling, enabling batch learning without padding or event duplication.

  3. Measured End-to-End Efficiency and Resource Costs: Relative to Fixed-5, actual runtime stopping reduces raw end-to-end latency by 2.16%, 40.07%, and 5.15% on THU-EACT-50-CHL, DVSGesture, and DailyDVS-200, respectively. These measurements include raw-event preprocessing, view generation, inference, and prediction integration. Peak GPU memory remains approximately 174 MiB, while peak CPU RSS increases by approximately 59–62 MiB; the additional budget-matched comparisons also favor Joint over static multi-view and stronger single-view alternatives.

We are grateful for your encouragement and for the suggestions that helped us improve the clarity and practical evaluation of EventVista. If you have a moment, we would sincerely value your thoughts on whether these clarifications and measurements address your questions. If they have addressed your main concerns, we would be very grateful if you could consider raising your score. Your support would mean a great deal to our team and encourage our continued research.

Thank you again for your time and thoughtful guidance, and wishing you continued success!

Warm regards,

The Authors

Reviewer_jmG9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes EventVista, an event-based action recognition method that decouples event retention from per-pass computation budget. It preserves a larger candidate pool, constructs complementary runtime views, and adaptively integrates predictions using confidence and stability criteria. Experiments on three benchmarks demonstrate competitive recognition performance.

Critical strengths
  1. Well-motivated problem. Decoupling event retention from computation addresses premature information loss in existing point-based methods. The framework is simple, and potentially generalizable.
  2. Promising results. EventVista achieves competitive performance with approximately only 2.1M parameters. Ablations support the effectiveness of its main components.
Critical weaknesses
  1. Limited novelty. The framework mainly combines established techniques, including voxel sampling, temporal stratification, logit averaging, and confidence-based early stopping. The observation order is fixed rather than prediction-adaptive, further limiting the methodological contribution.
  2. Incomplete comparisons and efficiency evaluation. Recent methods, including SMV-EAR (CVPR 2026), report higher accuracy on DailyDVS-200. Moreover, the multiple forward passes introduce additional computation, but comparisons under matched total computational budgets and end-to-end latency measurements are missing.
Questions for authors
  1. Are the reported DailyDVS-200 results from SMV-EAR and HoloEv-Net directly comparable? If so, how do these methods affect the SOTA claim?
  2. How does EventVista compare with standard multi-view inference or stronger single-view models under matched total FLOPs?
  3. Why does adaptive stopping substantially outperform fixed five-view inference? Could the authors explain this through prediction-transition analysis?
  4. Could the authors clarify the inconsistent fixed-view results across Tables 3–5?
5 comments
AuthorsOpenReview ↗

Response to Reviewer jmG9 Part (1/3)

Thank you for recognizing the budget-decoupling motivation, framework simplicity and generalization potential, and compact-model performance. We provide same-backbone controls, budget-matched experiments, and per-recording analyses to address your concerns.

W1 — Novelty of the budget-decoupled framework

Thank you for your comments. We believe that prior research on the basic operators does not mean that our method has the same evidence-access and inference-decision structure as existing methods; a fixed observation order also does not mean a fixed amount of actual observation or fixed computation allocation. EventVista's novelty should be understood at these two levels.

A one-shot reduction excludes unretained events from subsequent recognition; more computation on that input cannot revisit those source events. EventVista separately controls the candidate retention limit NcN_c, per-pass input limit NtN_t, and actual observations K(x)K(x). Candidate retention keeps real events beyond the first view accessible; temporally stratified views process them progressively under bounded inputs; prediction integration updates the prediction and decides whether to continue or stop. Together, these components implement a recording-level “evidence acquisition—prediction update—continue or stop” mechanism.

For recordings requiring multiple views, V(k)=Ok(C)V^{(k)}=O_k(C) is prediction-independent, but from k≥2k\ge2, the stopping decision a(k)=π(p(k),p(k−1))a^{(k)}=\pi(p^{(k)},p^{(k-1)}) depends on the recording's aggregated prediction; p(k)p^{(k)} is the class probability distribution after aggregating the first kk views. Fixed access order constrains next-view selection, while preserving the ability to decide whether to continue computation from observed evidence. Our contribution is this budget-decoupled evidence-access mechanism for point-based event recognition, rather than an individual sampling, averaging, or thresholding operator.

Keeping the backbone architecture and single-view input limit unchanged, we compare candidate retention and subsequent access. Below are Top-1 accuracies (%) on the three validation sets.

ConfigurationTHUDVSGestureDailyDVS-200
Coupled: Nc=Nt=24KN_c=N_t=24\mathrm{K}65.9195.4144.87
Decoupled: Nc=48K,Nt=24KN_c=48\mathrm{K},N_t=24\mathrm{K}71.2199.0850.79
Gain (pp)5.303.675.92

Retaining more source events and accessing them subsequently improves recognition without enlarging the per-pass input. This tests the joint effect of retention and supplementary observation; Q2 examines cumulative computation through budget-matched controls.

With the same encoder, candidate pool, coverage-aware views, and matched average calls, Joint outperforms static allocation by 1.47, 1.14, and 1.02 percentage points on THU, DVSGesture, and DailyDVS-200, respectively. This verifies the value of recording-dependent observation length even without prediction-dependent view reordering. Details are reported in Q2.

W2 & Q1 — Comparison with recent DailyDVS-200 methods

Thank you for pointing out the two excellent works, SMV-EAR and HoloEv-Net. We appreciate SMV-EAR's sample-level cross-view complementarity modeling and action-speed variation augmentation, and HoloEv-Net's combination of compact spatiotemporal representation and global frequency-domain modeling. These ideas inspire our future evidence-integration and lightweight-modeling work. We will continue following these teams' research. Citations and discussion have been added to the revised main text.

Under the respective published settings, SMV-EAR and HoloEv-Net-Base achieve 54.65% and 54.55% on DailyDVS-200, exceeding EventVista's 50.68%; HoloEv-Net-Small achieves 49.13%, below EventVista. On THU-EACT-50-CHL, EventVista achieves 70.40%, exceeding SMV-EAR's 66.70% and HoloEv-Net-Base's 69.85%.

EventVista's DailyDVS-200 model has 2.168M parameters, versus 23.5M, 12.5M, and 4.4M for SMV-EAR, HoloEv-Net-Base, and HoloEv-Net-Small, respectively: approximately 90.8%, 82.7%, and 50.7% fewer parameters. Against HoloEv-Net-Small, EventVista achieves 1.55 percentage points higher accuracy with approximately half the parameters. All observations share one encoder and classifier, so additional observations do not increase parameter count.

Accordingly, we have narrowed the performance claims in the main text: EventVista achieves SOTA results on THU-EACT-50-CHL and DVSGesture; on DailyDVS-200, it achieves competitive recognition performance with a compact model size.

AuthorsOpenReview ↗

Response to Reviewer jmG9 Part (2/3)

W2 — End-to-end inference efficiency

Compared with Fixed-5, Joint reduces actual end-to-end inference latency on all three datasets, with a 40.07% reduction on DVSGesture. To address the additional computation introduced by multiple forward passes, we have added actual runtime latency measurements. The model executes the encoder view by view, terminates observation once the Joint stopping conditions are satisfied, and performs no further network forward passes for the remaining views. Measurements use an RTX 3090, FP32, batch size 1, a single thread, TF32 disabled, and 20 warm-up iterations. Cached starts from an already constructed candidate pool and includes view generation, network inference, and prediction integration; Raw E2E starts from raw events and additionally includes candidate construction. Both scenarios use the same subset of profiling recordings.

All latencies below are per-recording averages in ms; the arrows indicate Fixed-5 → Joint.

DatasetCachedLatency reductionRaw E2ELatency reduction
THU557.16 → 515.757.43%3,522.48 → 3,446.252.16%
DVSGesture861.21 → 459.8446.61%1,010.07 → 605.3440.07%
DailyDVS-200873.74 → 819.866.17%1,149.62 → 1,090.365.15%

Reducing the number of observations translates into actual latency reductions on all three datasets, with a 40.07% reduction in end-to-end latency on DVSGesture. Candidate construction accounts for most of the runtime on THU, so the end-to-end reduction is smaller than in the Cached scenario. Candidate construction is performed once before observation; early stopping mainly saves the costs of subsequent view generation and network inference.

Q2 — Budget-matched comparisons with standard multi-view inference and stronger single-view models

Thank you for suggesting that the value of additional observations be tested by matching computation budgets. Joint achieves higher accuracy in both the static multi-view controls matched in average number of calls and the stronger single-view controls with approximately the same network computation.

First, static multi-view inference and Joint use the same encoder, candidate pool, and single-view input limit, with the average number of encoder calls matched over the full test set. The static strategies' call allocations are predetermined and do not depend on test labels or runtime predictions; Joint instead determines observation length according to each recording's prediction state. This control tests the role of recording-level adaptive allocation under the same average call budget. The table below reports average test Top-1 accuracy (%).

DatasetRandom Static MVStratified Static MVCoverage Static MVJointMatched average number of calls
THU65.4467.8368.9370.402.64
DVSGesture96.3297.4698.2199.352.12
DailyDVS-20047.4948.7249.6650.683.08

Joint improves over Coverage Static MV by 1.47, 1.14, and 1.02 percentage points, respectively, supporting the value of allocating observation computation according to recording state.

We further use the additional computation to increase the encoder channel width (Strong Wider, with the per-pass input limit kept at 24K) or enlarge the per-pass input (Strong More Points, with the limit increased to 48K), testing whether stronger single-view models can yield the same benefits.

Each cell below reports average test Top-1 accuracy (%) / network GFLOPs per recording. GFLOPs count only network computation and exclude candidate construction.

DatasetStrong WiderStrong More PointsJoint
THU68.75 / 10.0869.12 / 10.2470.40 / 10.18
DVSGesture98.38 / 14.2398.59 / 14.4299.35 / 14.37
DailyDVS-20049.50 / 20.6549.99 / 20.9450.68 / 20.88

The network computation of each of the two stronger single-view models differs from Joint by no more than approximately 1.1%, and Joint achieves higher accuracy on all three datasets. These results support using limited computation for supplementary observations and recording-level adaptive allocation rather than solely enlarging the model or the input processed at once.

AuthorsOpenReview ↗

Response to Reviewer jmG9 Part (3/3)

Q3 — Prediction-transition analysis of adaptive stopping

First, we clarify that after consistently correcting the Fixed-5 results, Joint and Fixed-5 have similar accuracy, and Joint's main benefit is reducing actual calls while maintaining comparable recognition performance. Through two sets of per-recording prediction transitions, we separately test whether supplementary observations before stopping are effective and whether continued aggregation after stopping still provides benefits.

We first compare prediction transitions from the first view → Joint. All counts in the table below are averages per evaluation, and accuracy gains are in percentage points.

DatasetWrong → CorrectCorrect → WrongNet gain in correct recordingsAccuracy gain (pp)
THU26.293.2923.004.23
DVSGesture4.000.004.001.52
DailyDVS-200154.0020.00134.003.27

On all three datasets, more errors are corrected than introduced, showing that supplementary observations before stopping can indeed compensate for discriminative evidence missed by the first view and yield a net recognition benefit.

We further force recordings that Joint has already stopped early based on prediction state to be processed through the fifth view and compare the prediction transitions from Joint → Fixed-5. The accuracies in the table below are Top-1 (%) over the full test set.

DatasetFixed-5 Top-1Joint Top-1Fixed-5 callsJoint callsCorrect → WrongWrong → Correct
THU70.2270.403.282.641.710.71
DVSGesture98.9799.355.002.121.000.00
DailyDVS-20050.5750.685.003.0813.719.00

Recordings that do not stop early have the same outputs as Fixed-5; continuing aggregation for early-stopped recordings can correct some errors but can also change originally correct predictions, reducing the net numbers of correct recordings by 1.00, 1.00, and 4.71, respectively. Therefore, the error-correction value of supplementary observations does not imply that continuously adding views will monotonically improve accuracy. Joint decides whether to continue based on the confidence, stability, and class consistency of the current aggregated prediction, retaining the recognition benefits of multiple views with fewer calls.

Compared with Fixed-5, Joint reduces the average number of calls by approximately 19.5%, 57.6%, and 38.4%, respectively. The 234 short THU recordings are processed in full only once, so the actual average number of calls for Fixed-5 is approximately 3.28; DailyDVS-200 also contains two short recordings, giving a Fixed-5 mean of approximately 4.9980, or 5.00 when rounded to two decimal places.

Q4 — Consistency of fixed-view results across Tables 3–5

Thank you for pointing out the cross-table differences in fixed-view results. Our inspection confirmed that, when preparing the original manuscript, we mixed results from different script versions and view-bank configurations but labeled them as the same setting. We have now selected a consistent checkpoint, ordered view bank, logit-averaging method, and evaluation statistics protocol for each dataset and re-evaluated the results.

The corrected Coverage-aware Fixed-5 Top-1 accuracies are THU 70.22%, DVSGesture 98.97%, and DailyDVS-200 50.57%, respectively. Fixed-5 aggregates the logits of the first five views; short recordings whose candidate counts do not exceed the single-view budget are processed in full only once. These results will be used consistently for the corresponding identical settings in Tables 3–5 of the original paper and will remain consistent with the Fixed-5 control in Q3 of this response.

Thank you again for your specific suggestions on novelty positioning, budget controls, and result consistency; we have revised the paper according to your suggestions. These suggestions have helped us present more clearly EventVista's value in organizing event evidence and allocating computation on demand with a compact model. If the above clarifications and additional evidence have addressed your main concerns, we respectfully ask you to reconsider your assessment and raise your score; your recognition would be tremendous encouragement and support for our team and this work.

Reviewer_J8A4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper points that point-based event action recognizers couple the number of events retained from a recording to the number processed in one forward pass, because they downsample once to a fixed-size input. EventVista separates the two budgets. (i) Authentic Candidate Preservation keeps up to 48K voxel-representative source events per recording, without padding or duplicating short recordings. (ii) Coverage-Aware Runtime Observation draws temporally stratified views of at most 24K events from this pool and encodes them with a PointNeXt-style encoder using offset-based packing. (iii) Adaptive Evidence Integration averages view logits and stops once a top-2 margin criterion and a distribution-change criterion are met . The paper reports 70.40% / 99.35% / 50.68% Top-1 on THU-EACT-50-CHL / DVSGesture / DailyDVS-200.

Critical strengths
  • Clear framing. Separating retention from per-pass computation (Eq. 2) is a useful lens for point-based event recognition.
  • Variable cardinality. Short recordings are handled carefully: there is no padding or duplication, and offset-aware packing is used. A control isolates the effect of batching.
  • Ablations and transparency. The ablations are broad. The authors disclose the unstable-sort issue, caveat their GFLOPs figures, state the offline scope, and release code.
Critical weaknesses

W1. Reported numbers are internally inconsistent

Under single-run micro accuracy on the stated test sets (n = 544 / 264 / 4,093), every reported accuracy must equal round(100k/n, 2) for some integer k.

  • Headline numbers. The DVSGesture headline (99.35) is not attainable; the nearest values are 99.24 and 99.62. The same holds for DailyDVS-200 (50.68), where the nearest values are 50.67 and 50.70.
  • All tables. Across Tables 1–5 and 8–11, 28/31, 26/28, and 19/31 distinct values are unattainable for the three datasets, respectively.
  • Table 12. No three runs on 264 recordings yield 99.29 ± 0.08, and no three runs on 544 recordings yield 70.28 ± 0.24.
  • Evaluation code. The released code computes micro accuracy from integer counts, so macro or batch averaging does not explain these values.

Results for the same nominal condition also disagree:

  • Fixed-5 accuracy. Coverage-aware fixed-5 accuracy on THU is 66.70, 65.56, and 65.40 in Tables 3, 4, and 5. By contrast, the temporal-stratification row matches between Tables 3 and 5.
  • Mean views on THU. Table 4 reports 5.00 and 3.00 mean views for fixed-5 and fixed-3 on THU. However, 234 short recordings are processed once, which caps the means at 3.28 and 2.14.
  • Mean views on DVSGesture. The single-criterion rules average fewer than two views on DVSGesture, despite the stated two-view minimum.

W2. Evaluation protocol and paper–code mismatches

  • Validation data. App. A.1 states that only official splits are used, yet thresholds are "selected on validation data." The released configs hold out 10% of the training data for this.
  • DailyDVS-200 split. The DailyDVS-200 config points to the validation split, not the official test list.
  • Released scripts.
    • Training samples random views rather than stratified ones.
    • Inference uses a random-subset bank instead of coverage-aware views.
    • Stopping is applied offline to saved logits, so the claimed compute savings are not demonstrated.
  • Hyperparameter sensitivity. M and K_max sit at sharp test-accuracy optima. On DVSGesture, M = 4/8/16 gives 95.18 / 99.35 / 93.30, and K_max = 7 drops accuracy to 97.38. No results on the selection split are reported.

W3. Evidence for the core mechanism is weak

  • Later views appear to hurt.
    • Under fixed-5 aggregation, coverage-aware views are worse than temporal stratification on THU (65.40 vs. 67.98).
    • Adaptive inference (2.64 views on average) beats every reported fixed prefix by at least 4.8 pp.
    • Stopping requires a margin of at least 0.70, and recordings that do not stop coincide with fixed-5. The gap therefore implies that about 26 confidently correct recordings (net) become wrong when later views are added.
    • This suggests that later views hurt, which contradicts the complementary-evidence narrative.
  • Views are not as complementary as described.
    • Figs. 4 and 8 show that View 2 adds only ≈9.6K new events and overlaps View 1 by 60%, even though a disjoint 24K view exists. This contradicts the description in Sec. 3.2 and A.3.
    • The observed counts match a sliding window with shift |C_m|/5.
    • In the released code, this shift depends on K_max, so Table 10 does not isolate the effect of the stopping limit.

W4. Positioning and fairness of comparisons

  • Missing related work. The paper does not discuss test-time voting over multiple subsets (e.g., RS-CNN's ten-vote evaluation), multi-window processing in prior event-point methods, or early-exit inference (BranchyNet, MSDNet). Without these, the new elements reduce to a view schedule and a stopping rule.
  • Confounded comparisons. The coupled 24K/24K variant already reaches 65.82 on THU, above every prior method in Table 1. EventMamba uses 1–2K points per window and appears to report window-level accuracy. The Table 1 gains therefore conflate input budget, encoder, and evaluation unit with the decoupling idea.
  • Statistical resolution. Most results are single-run, and several claimed differences are smaller than one test recording (0.379 pp on DVSGesture). Examples include the 99.35 vs. 99.20 margin over EventMamba and the "saturation" in Table 9.
Questions for authors
  1. Accuracy computation. What are the exact denominator and aggregation for each reported accuracy?
    • Please release per-recording outputs (labels, per-view logits, stopping step) for the headline and Table 12 runs.
    • Please state how the headline run was chosen.
    • Please explain which checkpoints and view banks produced the differing fixed-5 results.
  2. Protocol and code. Which split was used to select thresholds and checkpoints on each dataset, and was the final model retrained?
    • Which configs and entry points produce the reported DailyDVS-200 test results, the stratified training views, and the coverage-aware inference views?
    • Please report Tables 9–11 on the selection split.
  3. Mechanism.
    • Please report fixed-prefix accuracy for k = 1…5 and oracle-prefix accuracy.
    • Please report how many early-stopped recordings change their prediction by k = 5.
    • Please add a control in which View 2 is disjoint from View 1.
    • Please confirm whether all K_max settings share one view bank.
  4. Fair comparison. How does EventVista compare with a prior point-based method under a matched point budget and recording-level evaluation? What are the measured latency and memory, including candidate construction?
7 comments
AuthorsOpenReview ↗

Response to Reviewer J8A4 Part (1/5)

We sincerely thank you for recognizing the contribution of retention–computation budget decoupling, variable-length event processing without padding or duplication, and extensive ablations. Your scrutiny has helped us clarify how EventVista retains broader evidence and accesses it on demand under bounded per-pass computation.

W1 & Q1 — Accuracy computation and numerical consistency

Thank you for carefully checking the numerical results and their correspondence across tables. First, we need to clarify that all the questioned results can be directly recomputed from the integer correct counts of individual evaluations. The original manuscript insufficiently distinguished repeated test evaluations from independent training runs, allowing averaged accuracies to be mistaken for single-test results. We clarify this distinction and correct the cross-table inconsistencies below.

1. Accuracy computation and model selection

The main results on all three datasets average repeated evaluations with different evaluation seeds. Each covers the same complete official test set, producing one final prediction per recording and computing recording-level micro accuracy. For nn test recordings, crc_r correct predictions in evaluation rr, and RR evaluations:

\bar A=\frac{1}{R}\sum_{r=1}^{R}A_r=\frac{100\sum_{r=1}^{R}c_r}{Rn}.$$ Each evaluation uses integer correct counts, consistent with the public code. Main results use **the checkpoint from the epoch with the highest validation accuracy**, evaluated directly on the official test set without retraining or test-based run selection. The revised manuscript clarifies this aggregation. Checkpoint identifiers, training seeds, and evaluation view-bank seeds will accompany the anonymous reproduction materials. ### 2. Table 12 and Fixed-5 corrections The original Table 12 contained summary errors. Recomputing mean and sample standard deviation ($ddof=1$) over three independent training runs gives THU **$70.28\pm0.21\\%$**, DVSGesture **$99.37\pm0.22\\%$**, and DailyDVS-200 **$50.55\pm0.20\\%$**. These independent-training statistics are reported separately from the repeated test evaluations of the same checkpoint used for the main results. The original manuscript also mixed Fixed-5 results from different script versions and view-bank seeds. In the unified re-evaluation, strategies share the same checkpoint, ordered view bank, and logit averaging within each round. Corrected Fixed-5 Top-1 accuracies are **THU 70.22%, DVSGesture 98.97%, and DailyDVS-200 50.57%**, used consistently in Tables 3–5 and the corresponding ablations without stopping. The original THU Table 4 reported configured view counts as actual means. Of 544 recordings, 234 short recordings are processed once; thus Fixed-3/5 average-call upper bounds are **2.14/3.28**. Mean views now consistently counts actual encoder forward passes. ### 3. First-view rules for single-criterion ablations **Both Table 4 ablations allow first-view exit; Joint checks stopping only after at least two observations for candidate pools exceeding the single-view budget.** The original description omitted this distinction, especially Stability-only's first-view confidence gate. Stability is evaluated only from the second view. The ablations test earlier-stopping accuracy–computation trade-offs. Joint first acquires supplementary evidence because high first-view confidence does not guarantee sufficient discriminative support, then jointly checks confidence and stability to reduce premature stopping. Let $m^{(k)}$ be the top-two probability margin after aggregating the first $k$ views. Define stability as: $$S^{(k)}:\quad \left\lVert p^{(k)}-p^{(k-1)}\right\rVert_1\le0.10,\qquad \arg\max p^{(k)}=\arg\max p^{(k-1)}.$$ | Strategy | First-view exit | Exit from the second view onward | DVSGesture mean calls | | --------------- | ------------------------------------ | -------------------------------------------- | --------------------: | | Confidence-only | $m^{(1)}\ge0.70$ | $m^{(k)}\ge0.70$ | 1.86 | | Stability-only | $m^{(1)}\ge\tau_{\mathrm{first}}$ | $S^{(k)}$ is satisfied | 1.72 | | Joint | Not allowed | $m^{(k)}\ge0.70$ and $S^{(k)}$ is satisfied | 2.12 | For recordings requiring multiple views, observation continues up to $K_{\max}=5$ if the stopping conditions are not satisfied.
AuthorsOpenReview ↗

Response to Reviewer J8A4 Part (2/5)

W2 & Q2 — Evaluation protocol and paper–code consistency

The original description did not clearly distinguish validation from final testing. The official test sets remain unchanged throughout. Checkpoints and default hyperparameters are selected on the validation sets and then tested directly, without retraining.

1. Validation splits

DailyDVS-200 uses official Train/Validation/Test lists. THU and DVSGesture hold out subject-disjoint validation sets within official training data.

DatasetTrainValidationOfficial testValidation subjects
THU-EACT-50-CHL1,522264544P12, P17
DVSGesture968109264user01, user02

Subjects are indivisible. Selection prioritizes complete training-class coverage and validation-class coverage, then a holdout near 10%, capped at 20%. Actual proportions are 14.78% and 10.12%. The subject combinations are determined solely by data coverage and recording counts, without selection based on model accuracy; once determined, they remain fixed.

We corrected App. A.1's inaccurate statement that no additional random or subject splits were constructed and provided actual split lists. Checkpoints maximize validation accuracy, without merging validation data into training, retraining, or test-based selection.

2. Execution entry points and runtime stopping

The earlier public scripts differed from the main experimental entry points; reproduction instructions were incomplete. Training allocates quotas across temporal intervals, then randomly samples within each interval. The training views are random while also being constrained by temporal stratification; these properties do not conflict. Inference uses a fixed coverage-aware ordered view bank and skips remaining encoder calls once Joint conditions hold. Full-logit offline scripts support fixed-prefix and prediction-transition analyses; runtime-stopping latency measurements establish actual savings.

run.py executes the pipeline; protocols/subject_holdout/run_all.sh orchestrates audit/prepare/train/calibrate/test. DailyDVS-200 main results use 4,093 official test recordings. We will align public entry points and provide dataset configurations, split lists, and standalone commands in anonymous reproduction materials. Runtime measurements appear in W4 & Q4.

3. Full-validation-set results corresponding to Tables 9–11

All accuracies are recording-level Top-1 (%). The latter two tables report accuracy / actual mean encoder calls, including the first view.

Varied setting(Nc,Nt)(N_c,N_t)THUDVSGestureDailyDVS-200
Retention budget(24K, 24K)65.9195.4144.87
Retention budget(36K, 24K)69.3297.2548.84
Default(48K, 24K)71.2199.0850.79
Retention budget(64K, 24K)71.2199.0850.66
Single-view budget(48K, 12K)68.5698.1749.40
Single-view budget(48K, 18K)70.0898.1750.25
Single-view budget(48K, 32K)70.8399.0850.66

With the per-pass input limit unchanged, increasing the candidate retention budget from 24K to 48K improves performance on all three validation sets, providing clear validation support for the 48K/24K choice. Further enlarging the pool or input does not improve validation accuracy.

MMKmax⁡K_{\max}THUDVSGestureDailyDVS-200
4566.29 / 2.6995.41 / 2.1646.45 / 3.13
8571.21 / 2.6499.08 / 2.1250.79 / 3.08
16568.56 / 2.6694.50 / 2.1447.52 / 3.10
8370.08 / 1.9198.17 / 2.0550.09 / 2.75
8770.83 / 2.7298.17 / 2.1650.60 / 3.16

M=8M=8 outperforms both M=4M=4 and M=16M=16 on all three validation sets, showing that this choice is supported by the model-selection data. The same defaults, M=8M=8 and Kmax⁡=5K_{\max}=5, maximize validation accuracy across all three datasets; they are selected on validation data, not test peaks. Raising Kmax⁡K_{\max} to 7 adds calls without improving accuracy. W3 & Q3 isolates prefix-length effects using a shared view bank. THU pools no larger than the single-view budget are processed once, permitting a full-validation-set mean below two; Joint's two-view minimum applies to multi-view recordings.

τm\tau_mτd\tau_dTHUDVSGestureDailyDVS-200
0.600.1070.08 / 2.4298.17 / 2.0049.18 / 2.89
0.700.0570.83 / 2.8699.08 / 2.3050.66 / 3.33
0.700.1071.21 / 2.6499.08 / 2.1250.79 / 3.08
0.700.1570.45 / 2.5198.17 / 2.0250.47 / 2.92
0.800.1069.70 / 2.8498.17 / 2.3749.56 / 3.39

Default thresholds achieve the highest or tied-highest validation accuracy. Compared with τd=0.05\tau_d=0.05, they maintain equal or higher accuracy with fewer calls, reflecting the trade-off between validation accuracy and observation cost.

AuthorsOpenReview ↗

Response to Reviewer J8A4 Part (3/5)

W3 & Q3 — Complementary evidence, prediction transitions, and stopping behavior

Thank you for carefully examining the core mechanism. First, we clarify that our claim is not “more views lead to better recognition,” but “acquire supplementary evidence according to each recording's needs and stop once the aggregated prediction satisfies the confidence and stability conditions.” Also, overlap between views does not imply a lack of complementary evidence, nor does strict non-overlap guarantee higher recognition accuracy.

1. Prediction changes from the first view to Adaptive

We compare per-recording predictions, first view → Adaptive. Counts are per-evaluation means; gains are percentage points.

DatasetCorrect → WrongWrong → CorrectNet correct gainAccuracy gain (pp)
THU3.2926.2923.004.23
DVSGesture0.004.004.001.52
DailyDVS-20020.00154.00134.003.27

On all three datasets, more errors are corrected than introduced, yielding a net gain in correct predictions. This supports access beyond the first view.

2. Fixed prefixes and oracle-prefix

Fixed prefixes aggregate the first k=1,…,5k=1,\ldots,5 views; pools with ∣C∣≤Nt|C|\le N_t are processed once. Oracle-prefix marks a recording correct if any prefix is correct, using ground truth only as a post-hoc upper bound, never for inference. Values are recording-level Top-1 (%).

DatasetPrefix 1Prefix 2Prefix 3Prefix 4Prefix 5AdaptiveOracle-prefix
THU66.1868.3869.6770.0470.2270.4071.69
DVSGesture97.8498.5998.9799.2498.9799.35100.00
DailyDVS-20047.4149.3950.1350.6150.5750.6853.29

All multi-view prefixes outperform the first view, but accuracy does not increase monotonically with prefix length. Adaptive maintains comparable accuracy with 2.64, 2.12, and 3.08 mean calls, versus about 3.28, 5.00, and 5.00 for Fixed-5.

3. What happens after early stopping?

We force prediction-based early stops to continue to view five, excluding 234 THU and two DailyDVS-200 short recordings already processed in full once. Counts are per-evaluation means.

DatasetJoint early stopsFinal class changedCorrect → WrongWrong → CorrectWrong → Different Wrong
THU202.004.711.710.712.29
DVSGesture260.001.001.000.000.00
DailyDVS-2003,650.0048.4313.719.0025.71

Continuation corrects errors or reverses correct predictions, yielding net correct-count losses of 1.00, 1.00, and 4.71. Different class support across views means mean-logit aggregation need not improve each recording monotonically. Thus, the error-correction value of supplementary observations and the limited benefit of continued aggregation after stopping can both hold. Joint uses confidence, stability, and class consistency to decide whether to continue.

4. Strictly non-overlapping second-view control

Unseen-first prohibits reuse of previously seen events when sufficient unobserved events remain, encompassing the strictly non-overlapping second-view setting you requested.

Let Uk−1U_{k-1} be the observed-event union. Unseen-first selects from C∖Uk−1C\setminus U_{k-1}, using all unseen events before refilling with seen events if fewer than NtN_t remain. Within-view repetition is forbidden; full coverage permits cross-view reuse. When ∣C∣≥2Nt|C|\ge2N_t, the first two NtN_t-event views are disjoint. Every DVSGesture recording has 48K candidates and 24K events per view, so the second view covers exactly all events outside the first view. Smaller pools reuse events only as needed; pools of at most NtN_t are processed once. Test Top-1 (%):

DatasetCoverage-awareIndependent RandomIndependent StratifiedUnseen-first
THU70.4067.6568.7569.85
DVSGesture99.3597.4698.2198.76
DailyDVS-20050.6847.8549.1750.38

The illustrated ~9.6K new events and 60% overlap describe the main schedule, distinct from Unseen-first. These results show that maximizing additional coverage of event indices does not automatically yield higher accuracy; the organization and aggregation of discriminative evidence are equally important.

5. Shared view bank across Kmax⁡K_{\max}

All Kmax⁡K_{\max} values in the experiments share the same ordered view bank, with only the maximum prefix length changing; the event content and order of common prefixes remain 100% identical. Cells: test Top-1 (%) / actual mean encoder calls.

DatasetKmax⁡=3K_{\max}=3Kmax⁡=5K_{\max}=5Kmax⁡=7K_{\max}=7
THU69.49 / 2.1270.40 / 2.6470.22 / 2.71
DVSGesture98.97 / 2.0599.35 / 2.1298.97 / 2.15
DailyDVS-20050.13 / 2.7550.68 / 3.0850.54 / 3.17

This removes the original Kmax⁡K_{\max}-dependent step-size confound. A limit of five outperforms three; seven adds calls without accuracy gains, consistent with validation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Comment

Dear Area Chair and Reviewers,

We sincerely thank you for your careful reviews and constructive discussion. We especially appreciate the shared recognition of the motivation for decoupling event retention from per-pass computation, alongside the positive assessments of our coherent Retain–Observe–Stop framework, authentic variable-length processing, extensive experiments, and competitive recognition with compact models.

During the rebuttal, we focused on strengthening the evidence for EventVista’s central contribution: preserving broader event evidence and allocating observations according to each recording’s needs under bounded per-pass inputs. Below, we summarize the additional evidence and corresponding manuscript revisions, with changes highlighted in blue.

  • Controlled Evidence for the Core Contribution (for Reviewers jmG9, J8A4, h7hJ):

    • Additional rebuttal experiments include same-backbone retention controls, static multi-view inference matched in average encoder calls, and stronger single-view models with approximately matched network FLOPs.
    • Joint improves over Coverage Static MV by 1.47, 1.14, and 1.02 percentage points across the three datasets and also outperforms both stronger single-view alternatives, supporting the value of how EventVista organizes evidence and allocates computation.
  • Direct Validation of Supplementary Observations (for Reviewers J8A4, jmG9):

    • App. F.1–F.3 and Tables 16–18 provide prediction-transition analyses, fixed-prefix and oracle-prefix comparisons, and a strictly disjoint second-view control.
    • Supplementary observations yield net accuracy gains of 4.23, 1.52, and 3.27 percentage points over the first view, while Joint preserves accuracy comparable to Fixed-5 with fewer actual encoder calls.
  • Practical Benefits Verified by Runtime Measurements (for Reviewers h7hJ, J8A4, jmG9):

    • App. D.2 and Tables 14–15 report measured latency, preprocessing overhead, view-generation time, and peak memory.
    • Actual runtime stopping reduces raw end-to-end latency relative to Fixed-5 by 2.16%, 40.07%, and 5.15% across the three datasets, demonstrating that reduced observation counts translate into measured savings in the complete processing pipeline.
  • Consistent Evaluation and Transparent Reporting (for Reviewers J8A4, jmG9, h7hJ):

    • Sec. 4.1, App. A.1–A.2 and A.6–A.7, and Tables 3–5 and 12 clarify validation splits, checkpoint selection, repeated-evaluation aggregation, actual encoder-call counts, stopping initialization, and packed training.
    • We corrected inconsistent cross-table entries and independent-training summaries, and provided additional validation-set sweeps in the rebuttal; the official test sets remain unchanged.
  • Updated Comparisons and Precise Contribution Scope (for Reviewers jmG9, J8A4):

    • Secs. 2 and 4.2, Table 1, Fig. 3, and App. D.3 incorporate recent methods and explain EventVista’s relationship to test-time voting and network early exit.
    • The updated comparisons also establish its competitive accuracy–model-size trade-off on DailyDVS-200, where it achieves 50.68% accuracy with 2.168M parameters.

We hope this summary provides the AC with a clear basis for assessing EventVista’s contribution, supported by controlled comparisons, direct mechanism analyses, and measured computational benefits across three benchmarks. Thank you again for your time and careful consideration. We welcome further questions during the discussion!

Warm regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.