SutraNets: Sub-series Autoregressive Networks for Long-Sequence, Probabilistic Forecasting

We propose SutraNets, a novel method for neural probabilistic forecasting of long-sequence time series. SutraNets use an autoregressive generative model to factorize the likelihood of long sequences into products of conditional probabilities. When generating long sequences, most autoregressive approaches suffer from harmful error accumulation, as well as challenges in modeling long-distance dependencies. SutraNets treat long, univariate prediction as multivariate prediction over lower-frequency sub-series. Autoregression proceeds across time and across sub-series in order to ensure coherent multivariate (and, hence, high-frequency univariate) outputs. Since sub-series can be generated using fewer steps, SutraNets effectively reduce error accumulation and signal path distances. We find SutraNets to significantly improve forecasting accuracy over competitive alternatives on six real-world datasets, including when we vary the number of sub-series and scale up the depth and width of the underlying sequence models.

Paper

References (94)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer BQ6D6/10 · confidence 4/52023-07-03

Summary

The paper "SutraNets: Sub-series Autoregressive Networks for Long-Sequence, Probabilistic Forecasting" propose to model uni-variate time-series with two interleaved networks, that model 'fine-grained' and 'coarse-grained' time-steps. This has the advantage of reducing signal paths, reducing inference error accumulation, enable parallel training, and sharing parameters potentially more efficiently. The paper investigates 5 additional architectures on 6 standard datasets (e.g. electricity, but also MNIST), and demonstrates superior performance over the baselines.

Strengths

+ the paper is well and very clearly written (including the literature review, methods, Figures etc.) + the core idea of the paper makes intuitive sense, and the paper proposes several good ideas how to extend the standard autoregressive architecture, and investigates their impact on the final performance. + the method is shown to improve performance over baselines

Weaknesses

mainly minor things: - Figure 4: it's a bit unclear what is meant with 'more confident' - I think it is probably meant that it is less blurry? I'd be great to mention this explicitly if this is the case. Also, it'd be great to somehow quantify these findings: what are the likelihoods on the hold-out test data in the different settings? Maybe in this case (MNIST), also image-based evaluations could be performed (e.g. FID). - Figure 6: "per 100" in the figure itself. I think it'd be better to re-do the Figure with "inference time per forecast" and then put it there directly - otherwise it lacks a unit. - There is some prior work on multi-variate forecasting that has some similarity to this paper, but hasn't been mentioned. E.g. [1] (please also check reference & citations of this work), though focussing on the full multi-variate time-series problem, do also use an 'inner' architecture to generate data, and then have an outer autoregressive model. These should be mentioned in the related literature. [1] Rasul et al 2021. Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows.

Questions

- in eq.2 the conditioning is on x^k_{1:T+N}, i.e. y_{\tile t} is conditioned on co-variates from the future. I assume this should be x^k_{1:\tile t} instead?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

+ The paper has a good limitations section, which covers the limitations. E.g. the method being more complex, and requiring additional tuning for better performance. Also the societal impacts are considered in this section.

Reviewer 7ywN6/10 · confidence 5/52023-07-06

Summary

The authors propose new ways execute the recurrence in auto-regressive forecasting models. The approach is practical and shouldn't be too complicated to implement for a forecasting practitioner. Experiments show performance uplifts on common real and toy datasets.

Strengths

The approach is interesting and practical, although the motivation appear a bit heuristically. One important advantage that is mentioned but not emphasized much, is that due to the lagged recurrence, roughly $K$-fold parallel processing can be achieved during training, which can be rather significant for RNN models on long sequences. The paper is well organized, clearly written and easy to follow. The background section is nice and informative and provides a good overview over related work.

Weaknesses

The authors explore different interleaved recurrence orders, but there is more to the topic that could strengthen the paper. For example, how can the performance be attributed to different recurrence order on the one hand, and different hidden state dynamics on the other hand? As an example Fig 2(a) and (c) have the same recurrence order, but different hidden state dynamics. Accordingly Backfill-alt Fig 2(f) could perhaps perform even better with the straight hidden state dynamics of a standard RNN. Another analysis I'd love to see is how the conditional distributions $p(y_t| ...)$ look like for the different methods. Is there a recurrence order where these are mostly Gaussian or have another preferable simple shape? Take for example $y_1\sim\mathcal N(0,\sigma_1)$ and $y_2\sim\mathcal N(y_1,\sigma_2)$, then both $p(y_1)$ and $p(y_2|y_1)$ are Gaussian, but for $p(y_2)$ and $p(y_1|y_2)$ the later has a complicated bimodal form. Similar effects may impact the performance of auto regressive models in general, and it would be interesting to see that analyzed for the proposed models.

Questions

* The equation (2) is incorrect: the right hand side terms must condition on $\mathbf x_{1:T+N}$ not $\mathbf x^k_{1:\tilde T+\tilde N}$ in order for both sides to be equal. That has implications for the RNN formulation in l131 where in the current form the RNN state $h^k_\tilde t$ does not depend on any $\mathbf x_t$ with $t>K\tilde t$ effectively, while in equation (1) it does. * Equation (2) and the inline formula in l131 describe the Regular-alt case. It would be helpful to have these explicitly formulated also for the other 3 cases, so please add them. That could save space between l135 and l146 * l143-145: *"When not conditioning on > k values, SutraNets can instead generate the complete prediction range of the kth sub-series before generating any predictions for > k ones"*. I think that's not correct. When sampling according to xxx-non schema, the RNN has to be conditioned on $h_{\tilde T+\tilde N}^{<k}$ in order to get (wrapped up) information about the whole past sub-series $y_{\tilde T:\tilde T+\tilde N}^{<k}$ and not just $y_{\tilde t}^{<k}$. This may have implications to the experiments and currently impose an unfair disadvantage to the xxx-non methods.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

limitations are properly addressed

Reviewer MDp26/10 · confidence 4/52023-07-06

Summary

The author proposed SutraNets, a novel method for neural probabilistic forecasting of long-sequence time series. It addresses challenges faced by previous autoregressive approaches, such as error accumulation and modeling long-distance dependencies. SutraNets treat long predictions as multivariate predictions over lower-frequency sub-series, reducing errors and signal path distances. The authors conduct extensive experimental results on real-world datasets, and they show significant improvements.

Strengths

1. The sub-series in long-sequence analysis have seldom been investigated, especially for RNN-type networks. 2. The paper is well-organized, and the discussion is very clear.

Weaknesses

1. What is the deviation point of high and low frequency? What is the connection between traditional signal analysis? 2. The performance improvement seems moderate. Could the author include std for complete? 3. Finding 4 could include more recent SOTA methods.

Questions

1. Could the author discuss more about this work compared with pathTST (2023)? 2. See above.

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

4 excellent

Contribution

3 good

Limitations

Yes

Reviewer 6VCF4/10 · confidence 5/52023-07-07

Summary

This manuscript proposes SutraNets for long-range probabilistic forecasting on time series data and pixel sequences. SutraNets is a type of recurrent neural network that transforms a long series into a collection of shorter sub-series. Sub-series forecasts are generated autoregressively, sequentially conditional on each other, enabling coherent outputs.

Strengths

Figure 2 presents the generation ordering in SutraNets well, making it easy to follow.

Weaknesses

Since this work focuses on probabilistic forecasting, it should include related work and baselines on time series probabilistic forecasting models. However, most related baselines are missing. For example: - NeurIPS’21 Probabilistic Transformer for Time Series Analysis - ICLR’21 Multivariate Probabilistic Time Series Forecasting via Conditioned Normalizing Flows - ICML’21 Autoregressive Denoising Diffusion Models for Multivariate Probabilistic Time Series Forecasting “long-sequence” in the title means to predict a long sequence, it would be better to compare recent long-range time series prediction models. This is because the main work in this paper is the designed recurrent structure, rather than a new way to estimate the uncertainty. In other words, the connection between the proposed model and the probabilistic modeling is weak. Moreover, I believe that the structure of SutraNets is quite similar to recurrent neural networks with skip or residual connections in the temporal direction. And it does not depend on the K. It is unclear what advantages do SutraNets have over RNNs with skip connections. The main differences between SutraNets and the following works should be discussed. - NIPS’17 Dilated Recurrent Neural Networks - NIPS’16 Architectural Complexity Measures of Recurrent Neural Networks

Questions

see weaknesses

Rating

4: Borderline reject: Technically solid paper where reasons to reject, e.g., limited evaluation, outweigh reasons to accept, e.g., good evaluation. Please use sparingly.

Confidence

5: You are absolutely certain about your assessment. You are very familiar with the related work and checked the math/other details carefully.

Soundness

2 fair

Presentation

3 good

Contribution

2 fair

Limitations

Limitations and broader impacts have been discussed in Section 3.

Reviewer iCD85/10 · confidence 4/52023-07-07

Summary

This paper proposes a novel method for probabilistic forecasting of long-sequence time series. It uses an autoregressive generative model to factorize the likelihood of long sequences into products of conditional probabilities. The proposed model SutraNets treat long and univariate prediction as multivariate prediction over lower-frequency sub-series to effectively reduce error accumulation. Experimental results show improved forecasting accuracy on a variety of datasets.

Strengths

1. It's novel to convert a univariate series into a multivariate series, each dimension comprising sub-series of the original sequence. 2. It's interesting that SutraNet model generates each sub-series conditional on both its own prior values and on other sub-series.

Weaknesses

The proposed model is only applied to LSTM based model, while many of the state-of-the-art time series forecasting model is Transformer based. Not sure whether it can be integrated into Transformer based models.

Questions

Could this approach be applied to Transformer based models?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

4: You are confident in your assessment, but not absolutely certain. It is unlikely, but not impossible, that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work.

Soundness

3 good

Presentation

2 fair

Contribution

2 fair

Limitations

The limitations are mentioned in Section 3. However, it doesn't show the limitations to extend it to Transformer based models which are very popular in time-series forecasting recently.

Reviewer BQ6D2023-08-14

Reply

I thank the authors for their diligent efforts in addressing the concerns raised in my initial review. The minor concerns I had initially identified have been adequately discussed in the authors' rebuttal. At this point, I do not feel to have further points to discuss with the authors.

Reviewer iCD82023-08-16

Thanks a lot for authors on addressing my initial concerns. However, my main concern is in recent 2-3 years, latest top-tier conference, other backbone models have far more superior performance than RNN based model. Even for probabilistic forecasting, it's very easy to adapt current transformer based model in a probabilistic fashion. Please refer to [1] section 4.4 (appendix table 9) for the results with various Transformer based backbone model. Could the results on this paper beat the results shown the appendix table 9? [1]Shabani, Amin, et al. "Scaleformer: iterative multi-scale refining transformers for time series forecasting." ICLR2023

Authorsrebuttal2023-08-17

Thank you very much for engaging in this discussion and for suggesting the insightful comparison with Scaleformer. We absolutely do understand your concern. While our new paragraph above clarifies that Sutranets can be applied to Transformers, it is definitely not certain that the improvements that we demonstrated will generalize to Transformers. That being said, I am sure you would agree that NeurIPS may still include papers demonstrating advances in RNNs. As you know, RNNs are more biologically-plausible than Transformers (maintaining and updating a memory-like state), and, unlike Transformers, which must use a fixed-size context window, RNNs can theoretically condition on context of any length. This latter point is especially relevant to comparisons with Scaleformer ([Shabani et al, 2023](https://arxiv.org/abs/2206.04038)). In Table 9 of Shabani et al. (and in all their results), note the underlying Transformer is restricted to 96 historical values, i.e., only **4 days** of hourly data (Section 4.1: "the look-back window size is fixed to 96, and the horizon is varied from 96 to 720"). Yet also note the `traffic` dataset (and to a lesser extent, `electricity`) has extremely strong **weekly** seasonality - i.e., a future value is accurately predicted by the value 168 steps (hours) in the past. Since the implemented Scaleformer cannot "see" these values from one week ago, it makes sense that it performs much worse than both SutraNets, *and* the seasonal-naive-1week baseline (see below). (One may also qualitatively compare the results in their Figure 9 to our Figure 3, to see how Scaleformer may miss weekly "spikes" that SutraNets can predict). Furthermore, note that probablistic Scaleformer uses the non-SOTA Gaussian output distribution of DeepAR (Salinas et al., 2020), while SutraNets are evaluated with the SOTA approach of C2FAR (Bergsma et al., 2023). Although Scaleformer with Transformers may be limited to short conditioning windows, the core *idea* of Scaleformer can be applied with RNNs. Indeed, we implemented and evaluated a similar method to Scaleformer, but for RNNs (and using C2FAR as the output distribution), which we called Low2HighFreq. We noted the connection to Scaleformer in our related work. The full comparison (repeating results from our Table 1) is summarized in the following table: | System | Length of conditioning | Length of prediction | CRPS/wQL on `electricity` | CRPS/wQL on `traffic` | |--------------------------|------------------------|-----------------------|---------------|-----------| | Informer (Table 9, Shabani et al., 2023) | 96 | 96 | 0.330 | 0.372 | | Scaleformer (Table 9, Shabani et al., 2023) | 96 | 96 | 0.238 | 0.288 | | seasonal-naive-1week | 168 | 168 | 0.111 | 0.175 | | Low2HighFreq (Scaleformer for RNNs) | 168 | 168 | 0.082 | 0.166 | | SutraNets (Backfill-alt) | 168 | 168 | 0.074 | 0.128 | Here we *predict* 168 steps ahead vs. 96 for Scaleformer (which should disadvantage *our* results). In advance of the camera deadline, we will repeat these evaluations to use length-96 prediction windows, and we will also train versions of our system using length-96 conditioning. Also note Scaleformer evaluates using CRPS, while we evaluate using wQL, a 10-point approximation to CRPS (for point predictors like seasonal-naive-1week, note CRPS reduces to normalized deviation). While in some ways it is unfair to compare systems using different amounts of context, it is nevertheless quite notable that Scaleformer, FEDformer, Autoformer, TimesNet only condition on 96 historical values for the hourly predictions in all their experiments. We evaluated with up to length-2016 contexts, or 21X the length used in Scaleformer. The above proposed table should provide helpful information - for both practitioners and researchers - on the potential cost of restricting the look-back context.

Reviewer iCD82023-08-17

Thanks a lot for the author's latest experiments. This addressed my main concern and I increased the rating to 5.

Reviewer 6VCF2023-08-17

Thank you for your response. After carefully reading your response and re-evaluating the manuscript, I still have the following questions: I believe the main contribution of this work lies in the long-sequence decoding strategy (Backfill-alt is suggested) for autoregressive models, which aims to reduce error accumulation (mentioned in the abstract). However, it is important to note that this strategy is still limited. For instance, in Figure 1, the approach uses (y0, y3, y6, y9) to generate (y12, y15, y18) initially, and then employs (y12, y15, y18) along with the history (y1, y4, y7, y10) to generate (y13, y16, y19). This raises some questions: - I noticed that y12, y15, and y18 are directly included in the final output, even though they were generated using only (y0, y3, y6, and y9) without considering the recent values y10 and y11. **If (y12, y15, y18) turn out to be incorrect predictions, would these errors accumulate and affect the subsequent predictions of (y13, y16, y19)?** - Why wasn't (y10, y11) used to generate y12? It seems that predicting y12 using (y10, y11) would be much easier (short-term prediction). - Moreover, since y12 is decoded using y0, y3, y6, and y9, **does it imply that the model assumes the time series to be periodic**? (The time series depicted in Figure 3 appears to be relatively straightforward to predict using existing long-term prediction methods.) There is a related work TimesNet [1], which shares a similar idea to SutraNets. Both approaches transform long time series into multiple subseries. Considering this, I have another concern regarding **why the authors did not compare their proposed model with recent long-time series prediction methods such as [1]**. This question was not addressed in their response. [1] [Timesnet: Temporal 2d-variation modeling for general time series analysis](https://arxiv.org/abs/2210.02186) I believe it would be relatively straightforward to adapt long-time prediction methods for probabilistic forecasting or to evaluate SutraNets in the context of non-probabilistic forecasting. This would be valuable in illustrating the prowess of SutraNet in long-sequence prediction. Given the limited evaluation presented and the absence of baselines for long-time series prediction methods, I remain uncertain about the true effectiveness of the proposed model in **long sequence** prediction.

Authorsrebuttal2023-08-18

Reply to Official Comment by Reviewer 6VCF (1/2)

We really appreciate that you have both helped improve the paper, and are now re-evaluating the manuscript in light of our work together. > the main contribution of this work lies in the long-sequence decoding strategy ... which aims to reduce error accumulation Based on reviewer feedback, we will definitely revise the abstract to emphasize the novel benefit of SutraNets is the ability to simultaneously: (1) capture long-term dependencies by reducing signal path, (2) reduce error accumulation in autoregression, (3) enable a K-fold improvement in training parallelism for RNNs, and (4) generate coherent outputs. > in Figure 1, the approach uses (y0, y3, y6, y9) to generate (y12, y15, y18) initially, and then employs (y12, y15, y18) along with the history (y1, y4, y7, y10) to generate (y13, y16, y19) We understand you to be extrapolating the process in Fig. 1(d), for **Regular-non**. Indeed, it is a *non*-alternating model: the first sub-series prediction is generated end-to-end while only using historical values for that subseries. > If (y12, y15, y18) turn out to be incorrect predictions, would these errors accumulate and affect the subsequent predictions of (y13, y16, y19)? Errors do still accumulate, but to a lesser extent. Consider generating y19. In a standard RNN (Fig. 1a), we must generate y11, y12, y13, y14, y15, y16, y17, y18 - 8 steps - in order to generate y18, i.e., the value that precedes y19. However, in Regular-non, we can generate y13, 15, y18 - 3 steps - in order to generate y18. So the prediction of y19 can be based on proximal values (y18) themselves generated with $K$X less error accumulation. > Why wasn't (y10, y11) used to generate y12? It seems predicting y12 using (y10, y11) would be much easier Regular-non by definition generates y12 using y9 and earlier, while standard RNNs (Fig. 1a) and Regular-alt (Fig. 1c) generate sequentially as you are suggesting. Yet interestingly, Regular-non sometimes works better than Regular-alt in experiments. Why? Well, During *training*, sequential RNNs sometimes find it effective to just repeat the previous gold/true value, i.e., to predict y12 based on gold y11. But in *inference*, if y11 is generated after a long chain of error accumulation, repeating y11 to predict y12 is problematic. By forcing the network to attend to values $K$ steps in the past, we reduce train/test "discrepancy" and improve forecasting accuracy. > since y12 is decoded using y0, y3, y6, and y9, does it imply that the model assumes the time series to be periodic? SutraNets do not require data to be periodic to be effective, e.g., dataset `mnist`$^{\pi}$ comprises images after a fixed random permutation, and `azure` and `wiki` comprise many trending but non-seasonal series. However, for time series that *are* periodic, we asked a similar question to yours: does it matter whether y0, y3, y6, etc. captures seasonality, i.e., does it matter whether $K$ divides into the seasonal period? The results are in Table 3 and Finding 5: non-alternating models are comparatively weaker than alternating versions. > it is important to note that this strategy is still limited While *non-alternating* RNNs generate future values without "seeing" all the history, keep in mind that **all the alternating RNNs do see all the history in their generative process**. That is, the orange RNN in Fig. 1(f) will have seen y0, y1, y2, ..., y7, y8 when it goes to generate y11. True, it does not see y9 or y10, but that's the point: all values from y9 onward must be generated, and Backfill-alt generates y11, then y10, then y9. Since this seems to work quite well, we would like to share these findings with the community. > There is a related work TimesNet [1], which shares a similar idea to SutraNets. Both approaches transform long time series into multiple subseries We can certainly include discussion of TimesNet. It is interesting that TimesNet transforms univariate time series into 2D tensors of *multiple* periods, which are then processed via Inception-style 2D kernels (i.e., now using a 2D computer vision backbone, and outputting point predictions). SutraNets on the other hand transform time series into $K$ sub-series of the *same* period, and proposes an autoregressive factorization (with different generative orderings) for generating these sub-series using their own and previous values (with a 1D sequential generative model that produces probabilistic outputs).

Authorsrebuttal2023-08-18

Reply to Official Comment by Reviewer 6VCF (2/2)

> why the authors did not compare their proposed model with recent long-time series prediction methods such as [1] Experimentally, we compare SutraNets to other methods that reduce signal path and mitigate error accumulation, while generating coherent probabilistic outputs. For experimental rigour, we only selected methods that we could fairly re-implement in our codebase, so that all methods share the same normalization (based on the conditioning window), optimizer, sampling of training data, decaying learning rates, criteria for early stopping, fixed and tuned hyperparameters, etc. - We implemented C2FAR (baseline), C2FAR+lags (to improve signal path), C2FAR+dropout (to mitigate error accumulation), Low2HighFreq (to leverage low-frequency predictions to guide high-frequency generation, similar to Scaleformer), plus four SutraNet variations. Ideally we could also compare to the performance of other implementations on standard datasets. Unfortunately, **comparing to the evaluation in the TimesNet paper (and other recent work) is problematic**. - As explained to Reviewer iCD8 (regarding Scaleformer), TimesNet (like Scaleformer, Autoformer, and FEDformer) is restricted to only using at most 96 historical values, i.e., only **4 days** of hourly data (Table 2: "the past sequence length is set as 36 for ILI and 96 for the others"). Note we use up to 2016 historical values -- **21X the maximum lookback context used in TimesNet** -- so TimesNet and SutraNets are not really comparable in terms of the scale of historical data considered. - More importantly, the `traffic` dataset (and to a lesser extent, `electricity`) has extremely strong **weekly** seasonality - i.e., a future value is accurately predicted by the value 168 steps (hours) in the past. Since the implemented TimesNet cannot "see" values from one week ago, it is likely to perform much worse than even the seasonal-naive-1week baseline (which itself performs worse than SutraNets). We strongly advocate including the seasonal-naive baseline in all forecasting results, as it facilitates comparison. In the words of [Hewamalage et al., 2023](https://arxiv.org/abs/2203.10716): "on series that have clear seasonal patterns, models should accordingly be benchmarked against the seasonal naive model." - That being said, please see our recent response to Reviewer iCD8, where we now compare our proposed model to the recent long-time series prediction method of Scaleformer, based on published Scaleformer results on `traffic` and `electricity`. Unlike with TimesNet, this comparison is facilitated by Scaleformer generating a probabilistic forecast, and evaluating using comparable metrics to us (albeit still with restricted context). > it would be relatively straightforward to adapt long-time prediction methods for probabilistic forecasting or to evaluate SutraNets in the context of non-probabilistic forecasting. This would be valuable in illustrating the prowess of SutraNet in long-sequence prediction Indeed, it would be quite impressive if a system designed for probabilistic forecasting was competitive with the accuracy of all the non-probabilistic forecasting systems on that task, and yet we have reason to believe that this is the case: most recent "long sequence" systems will be negatively impacted by only using limited context windows, as described above. However, while this would be nice to show definitively, it's still an apples-to-oranges comparison. That is, and I am sure you would agree, it need not be a *requirement* for publication that systems designed for probabilistic forecasting also establish competitive results on non-probabilistic forecasting. > Given the limited evaluation presented and the absence of baselines for long-time series prediction methods, I remain uncertain about the true effectiveness of the proposed model in long sequence prediction While your point is well taken regarding SutraNet evaluation on the related task of *non-probabilistic* forecasting, we do provide "extensive experimental results on real-world datasets" (as noted by Reviewer MDp2) in our evaluation of *probabilistic* systems. We evaluated: - 8 different neural models (4 comparison, 4 SutraNet) on 6 different datasets (Table 1) - 3 neural models with 4 different variations in neural depth on 2 datasets (Table 2) - 2 neural models with 3 different variations in $K$ on 2 datasets (Table 3) - 2 neural models with 5 different variations in random seed on 2 datasets (supplemental Fig. 3) Each of these 100+ results required running a hyperparameter tuning grid search, so, as you can imagine, the total GPU compute time was rather significant for this project, but necessary for the depth and breadth of analysis that we provide in the paper.

Reviewer 7ywN2023-08-18

Thanks to the authors for their rebuttal. I think my concerns where addressed and suggestions followed, but no major step changes were made to the paper. My initial rating seems to resonate with the other reviewers, so I'm gonna stick with it.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC