acceptodds
Under review as a conference paper at ICLR 2027

An Analytical Theory of Auxiliary Learning

Abstract

Auxiliary learning is an optimization paradigm in which a neural network's performance on a target task is improved by jointly training it on additional tasks. However, the mechanisms behind this improvement remain poorly understood. We study this problem using a teacher-student framework and derive a closed system of differential equations describing the dynamics of online stochastic gradient descent in the large-input limit. For linear networks, we obtain a closed-form expression for the generalization error to leading order in the learning rate, quantifying how task correlations and label noise determine the benefit of auxiliary learning. For non-linear activation functions, we develop a fluctuation-dissipation analytical theory that establishes a general relation linking the main and auxiliary errors to the corresponding single-task error. Numerical experiments support the theoretical predictions and show how auxiliary tasks improve generalization by balancing the forcing dynamics towards the optimal solution with gradient noise.

Primary area:
learning theory
Keywords:
auxiliary learning, multi-task learning, teacher-student, fluctuation dissipation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_Fy32OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies auxiliary learning in a teacher–student framework, deriving online SGD dynamics and a leading-order expression for the steady-state generalization error of linear networks. It extends the analysis to nonlinear activations through a reduced Lyapunov equation, with numerical simulations supporting its predictions of how loss weights, readout correlations, and label noise affect generalization error under the fixed-readout setting.

Critical strengths
  • The paper develops a clean, tractable teacher–student framework for auxiliary learning, with an explicit ODE description of online SGD and steady-state analyses covering both linear and nonlinear activations. The theoretical analysis yields concrete results, including a closed-form leading-order error for linear networks and, for nonlinear activations, a reduction of the stationary fluctuation analysis to a K×KK\times K Lyapunov equation.
  • The presentation is clear, and the analytical predictions are supported by comparisons with ODE integration and SGD simulations in the regime studied.
Critical weaknesses
  1. The connection between the theoretical findings and the auxiliary-learning phenomena motivating the paper remains insufficiently established. The framework provides a tractable account of stationary prediction error, but it is less clear what these findings tell us about when auxiliary supervision helps or harms the main task through learning a shared representation. Both targets are generated from the same teacher features and admit a common realizable optimum, while the main analysis fixes the readouts to their teacher values. The experiments largely validate the theory within these assumptions, leaving unclear which weighting laws and error relations survive when the readouts are learned jointly, task demands compete for shared capacity, or training repeatedly uses finite data. A modest controlled check relaxing one of these assumptions, or a substantive analysis explaining which findings should persist and why, would help bridge the model to the broader auxiliary-learning questions raised in the introduction. Successful extrapolation or a complete theory is not required; identifying where the regularities change would itself clarify the contribution’s scope and significance.

  2. The introduction could state the main findings and their significance more concretely. The abstract and introduction explain what is analysed and say that task correlations and label noise determine the auxiliary benefit, but do not make the resulting regularities explicit. For example, highlighting the correlation-independent optimal auxiliary weight in the linear setting or the weighted-error identity, with their conditions, would give readers a clearer understanding of what has been discovered. The revision should foreground a few principal findings and explain what they add to the understanding of auxiliary learning, rather than primarily describing the analytical framework and technical outputs. This is primarily a presentation concern.

Questions for authors

The main requests are in the Weaknesses section. The following clarifications would also be helpful; a discussion would suffice, with additional experiments optional.

  1. Role of the learning rate. Appendix C identifies an effective learning rate and label noise for ρ=1\rho=1, while Appendix E gives ϵmain/ϵ0=1−β\epsilon_{\rm main}/\epsilon_0=1-\beta at ρ=0\rho=0, where the improvement comes from reducing the main-task learning rate. How should the auxiliary benefit be interpreted relative to single-task training with a reasonably tuned learning rate, particularly in terms of the error–convergence-speed trade-off or performance under a fixed training budget? This does not question the validity of the fixed-learning-rate comparisons, but would help clarify the significance of the reported gains.

  2. Positioning relative to closely related work. For a single output, at a realizable optimum with independent homoscedastic label noise, the gradient-noise covariance is proportional to the Hessian, which gives a simple solution to the leading-order stationary covariance equation. Could the authors clarify the specific advance of Eq. (22), ϵ0=μσ2E[g′(h)2]∥v∥2/2\epsilon_0=\mu\sigma^2 E[g'(h)^2]\lVert v\rVert^2/2, relative to classical LMS misadjustment results and the stationary analysis of Mandt et al. (2017)? In particular, is the advance its explicit evaluation for general activations in the teacher–student setting? It would also help to discuss Mori et al. (ICLR 2025), which studies task-selection protocols in closely related teacher–student models for multi-task and continual learning.

References:

  • B. Widrow, J. M. McCool, M. G. Larimore, and C. R. Johnson, “Stationary and nonstationary learning characteristics of the LMS adaptive filter,” Proceedings of the IEEE, 64(8):1151–1162, 1976.
  • F. Mori, S. Sarao Mannelli, and F. Mignacco, “Optimal Protocols for Continual Learning via Statistical Physics and Control Theory,” ICLR 2025 (arXiv:2409.18061).
3 comments

We thank the Reviewer for their constructive comments on our manuscript. We addressed the two critical weaknesses by adding simulations with a learnable second layer (in the manuscript, Sec. 3.4 and Fig. 5) and by expanding the introduction for clarity. We also added App. E and Fig. S4 showing how the additional head speeds up convergence across β\betas, as well as performance. We also added Mori et al. (2025) to the related work.

Critical Weaknesses 1

To address these concerns and expand the scope of our results, we relaxed the assumptions of frozen readouts, and we added Sec. 3.4 and Fig. 5 to show what happens when training the second layer. Networks with non-homogeneous activations such as erf converge to the frozen 2nd-layer Lyapunov theory, while networks with ReLU and linear activations keep improving without plateauing. This is because if an activation function is 1-homogeneous (f(cx)=cf(x)f(cx)=cf(x)), there is a degeneracy of solutions. SGD then moves the weight norm from the second to the first layer, improving generalization by effectively decaying its learning rate. But rescaling this error by the second layer weights once again recovers our theoretical prediction.

As for the weighting laws, we updated Fig. 4 to show results with different KK. With larger KKs, optimal β∗\beta^*s are closer to the optimal 2−1\sqrt 2-1 of the linear model. We were not able to derive the analytical limit for K→∞K\to\infty, if such a limit exists, but it could be a future direction to explore. Training with a finite dataset is another interesting direction, which we believe to be outside the scope of this analysis. We note DMFT as a promising route (for example, Bordelon&Pehlevan, 2023). In general, “competing demands for shared capacity” is not covered and is a stated limitation of our work.

Critical Weaknesses 2

We agree with the reviewer about the introduction, and we have expanded it, hopefully making our paper's findings clearer.

Questions For Authors:

Q1. To answer this question, we added App. E and Fig. S4. They show how convergence time changes with β\beta (with other hyperparameters fixed, and α=1\alpha=1). We consistently observe that having an additional head speeds up convergence, and the gain is larger with smaller ρ\rhos and larger β\betas. With smaller learning rates, models require more steps to converge, and thus training a single-head model to match the generalization error of a multi-head model with a well-tuned β\beta also requires significantly more steps. This is important because in online learning, convergence time also indicates how much data is needed to converge.

Adding to this, the results on the trainable second layer we discussed above show that the effective learning rate of the network can change dynamically directly from SGD, without any scheduling. In a way, SGD keeps the same learning rate until convergence, and then starts diminishing it by moving the weight norm from the second to the first layer. Possibly, a learning rate schedule might perform even better, but deriving it analytically is left to future research.

Q2. We agree with the observation on the Hessian. While there are similarities between our results and Mandt et al. (2017), which partially inspired our approach, there are important differences. From a technical angle, their results rely on using mini-batches and invoking the central limit theorem, and the noise comes from mini-batches resampled from a fixed, finite dataset. In our setting, we specialize in a solvable model trained with online SGD and noise coming directly from the teacher. Thus, the Lyapunov equation is solvable a priori because the thermodynamic limit reduces our problem to O(K2)\mathcal O (K^2) variables. But our goals are also different: they focus on tuning SGD parameters from a Bayesian point of view and using SGD as a variational EM algorithm, while we want to solve our models analytically to better understand their behaviours.

We also thank the reviewer for pointing out that we missed Mori et al. (2025) among the previous literature. We have corrected this. The main difference in our work is that we train tasks in parallel, while they use continual learning, where tasks alternate. They choose a teacher with only one neuron for each task. They then introduce a task-similarity parameter γ\gamma between the 1st-layer weights, while our ρ\rho is between the 2nd-layer weights. But the biggest difference is that they focus on convergence and forgetting, two dynamical aspects of training, while we study the behaviour of the network already at the steady-state solution.

Reviewer_Fy32OpenReview ↗

Reviewer Fy32: response to rebuttal — score raised to weak accept

Thank you for the response and the revision.

W1. Section 3.4 and Figure 5 are the controlled check I had in mind. For erf\mathrm{erf}, the fixed-readout prediction still agrees with the experiment when the second layer is trained. For 1-homogeneous activations, the error continues to fall as norm transfers between layers, and the results agree with the theory only after rescaling by ∣v∣2|v|^2. The larger-KK results in Figure 4 are also useful.

Two details from the discussion should be included in the paper:

  1. The response identifies finite training data and competing demands for shared capacity as limitations. I could not find these in the revised PDF. Please state them explicitly, along with the explanation given in the reply to Reviewer Grnh: in the realizable online setting studied here, the auxiliary task improves generalization through noise averaging and an effective change in learning rate.
  2. Figure 5 shows a single run at one hyperparameter setting. Please make that clear in the caption, or add a few seeds or a second value of β\beta.

W2. The revised introduction now motivates the work in terms that match the setting analyzed. This resolves my concern. One typo remains on p. 1: “the optimal weigh and that the optimal auxiliary weight.”

Q1. Appendix E answers my question. The auxiliary head converges faster and reaches a lower steady-state error. These results are sufficient for my assessment, and I do not need an additional single-task curve with a tuned learning rate.

Q2. The positioning relative to Mandt et al. (2017) and Mori et al. (2025) is now clear. Please add one sentence relating Eq. (22) to the classical LMS misadjustment result, citing Widrow et al. (1976).

The revision addresses my main concerns, and I am raising my score to weak accept.

We thank the reviewer for the follow-up and for supporting acceptance. We updated the PDF with the requested changes and renamed Section 4 to "Future Work and Limitations" to better reflect its content.

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

The paper studies auxiliary learning in a two-output teacher-student model: a student with a trainable first layer WW and two readouts fixed at the teacher values (vv for the main task, uu for the auxiliary task) is trained by online SGD on the loss αLmain+βLaux\alpha \mathcal{L}_{main} + \beta \mathcal{L}_{aux}. Labels are generated by a teacher of the same architecture and corrupted by Gaussian label noise. Since both tasks share the teacher's first layer, the auxiliary term does not change the global minimizer; its effect is purely dynamical.

Contributions and validation: (1) Linear networks. The authors derive a closed-form expression for the main-task error at convergence, to leading order in the learning rate, and the optimal weights. Theory agrees well with SGD simulations. (2) Nonlinear networks: Using simulations and the mean-field ODEs, the authors show how the main-task error depends on the auxiliary weight β\beta for erf and ReLU networks and explain their results with a quantitative theory using a Lyapunov equation for the weight fluctuations. In the non linear case the optimal weight depends on the activation and on the specific readouts.

Critical strengths
  1. A closed-form result for linear networks For linear networks, the authors derive an explicit expression for the main-task error (to leading order in learning rate), which shows directly how the task correlation ρ\rho and the weight β\beta affect the error, and gives a clean optimal weight β∗=2−1\beta^*=\sqrt2-1. The theory agrees well with the ODEs and the SGD simulations.

  2. A general approximation for nonlinear networks. Expanding SGD around the optimum gives a small (K×KK \times K) Lyapunov equation that applies to any activation function and reproduces the simulations for erf and ReLU.

Critical weaknesses

W1. The setup captures only one of the usual mechanisms of auxiliary learning. Auxiliary tasks are usually expected to help because the main task alone is not enough to learn a good shared representation (Ruder, 2017, Sec. 4). The paper assumes a shared representation, which is reasonable, but with unlimited data the main task alone already contains everything needed to learn this representation. The only mechanism left is averaging out the label noise across the two tasks.

Freezing the readouts at the teacher values restricts the setup further. Fixed readouts are common in the teacher-student literature, but here they mean that the auxiliary loss is minimized by the same WW as the main loss throughout training, so the auxiliary task can never pull the shared layer in the wrong direction. In practice the auxiliary head is learned, and when auxiliary tasks help and when they hurt is a central question; here, the possibility that they hurt is ruled out by construction. The setup is clean and tractable, but it is not clear to me that it explains why auxiliary learning helps in practice. I believe this is the main weakness of the paper.

Minor / presentation:

  • The notation I2(i,j)I_2(i,j) (Eq. 7) is confusing, since the same symbol and indices are used for student-student, student-teacher and teacher-teacher averages.
  • Sec. 3.1 would benefit from a short sketch of how Eq. 8 is derived, rather than only a pointer to App. B.
  • Fig. 3: the vertical line marking the linear-model optimum is labelled only in panel B, and is easily confused with the dashed theory curves.
Questions for authors

Q1. Learnable readouts. Do the main results (the benefit of the auxiliary task and the value of β∗\beta^*) still hold when vv and uu are trained? The ODEs in Eqs. 29a,b seem to allow checking this numerically. In particular, can the auxiliary task then hurt the main task?

Q2. Spread in Fig. 4. For a fixed ρ\rho, Fig. 4 shows a large spread in both β∗\beta^* and the gain. What causes this spread, and do the values concentrate around a specific value for large KK?

2 comments

We really thank the Reviewer for their precise comments, which reflected a clear understanding of our work. We agree with the minor presentation details they pointed out, which we fixed, and we included two additional results to answer their broader questions: we simulated our networks with a learnable second layer, added in Sec. 3.4 and Fig. 5, which we found matches our theory when homogeneity degeneracy is accounted for, and expanded Fig. 4 to show larger values of K.

Critical Weaknesses: W1 and W2

To start addressing this limitation, we relaxed the constraint of frozen readouts, and we added Sec. 3.4 and Fig. 5 to show what happens when training the second layer. Here, we summarize two observed behaviours. Networks with non-homogeneous activations such as erf converge to the frozen 2nd-layer Lyapunov theory, while networks with ReLU and linear activations keep improving without plateauing. This is because if an activation function is 1-homogeneous (f(cx)=cf(x)), there is a degeneracy of solutions. SGD then moves the weight norm from the second to the first layer, improving generalization by effectively decaying its learning rate. However, rescaling the error by the norm of the second-layer weights recovers our theoretical prediction even for homogeneous activation functions, thus supporting the generality of our theory and conclusions.

Minor / presentation:

We agree with all points raised.

  • We improved the notation for clarity.
  • We added a paragraph sketching the derivation in Sec. 3.1.
  • We removed the vertical line.

Questions For Authors:

  1. In our setup, the auxiliary task can hurt generalization even with a fixed 2nd layer. When the two tasks have the same noise variance, this happens for β>1\beta>1. But, as shown in App. B, when the auxiliary task has different noise, generalization can be worse than the corresponding single-output model even for smaller β\betas. In the linear model, for example, the critical value above which generalization worsens is m2ain/a2ux. We already discussed our addition of a trainable second layer: either it does not change results (non-homogeneous functions like erf) or makes the network learn arbitrarily well (1-homogeneous functions).

  2. We find this observation to be quite interesting. We updated Fig. 4 to show results with different K. Indeed, the spread of optimal β∗\beta^* is smaller, and they are closer to the optimal 2-1 of the linear model, but the spread of corresponding errors remains similar. We were not able to derive the analytical limit for K, if such a limit exists, but it could be a future direction to explore.

We thank the authors for the careful response and for the revisions. The new results with a trainable second layer and the revised presentation indeed improve this work. However, having considered the response and the revised manuscript, our overall assessment is unchanged. The work is sound and clearly presented and should be accepted, hence the score of weak accept. But in our view the scope of the setting and of the conclusions does not place it above the bar for a clear accept.

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

The authors consider the problem of predicting generalization error in an auxiliary learning context. In this context, a neural network is trained on a ``main'' task as well as one or more auxiliary tasks. It is thought that training on auxiliary tasks improves the generalization error on the main task. The authors derive an analytical expression for the generalization errors of the main and auxiliary tasks under certain learning rate regimes in the linear activation as well as the nonlinear activation settings. They use fluctuation-dissipation-type arguments for their derivation. The resulting predictions (in closed form for the linear settings) agree with those from ODEs derived in the population (mean field limits) in previous work, but these are simpler to compute. From their expression, the authors are able to derive the weighting in the loss function for the auxiliary task in order for the generalization error to be minimized.

Critical strengths
  1. The teacher student set up is elementary and applies aptly to the learning problem at hand.
  2. The similarities and differences between the predictions of generalization error (up to leading order in learning rate) are well-explained.
  3. The somewhat surprising result that the correlation between the output weights of the main and auxiliary tasks does not affect the optimal weighting (β∗\beta^*).
Critical weaknesses
  1. Reading the related work section and one of the papers (Goldt et al 2019), I am not able to see the novelty in the paper. Goldt et al do a substantial analysis of the dynamics used in the paper and the fluctuation-dissipation results from the statistical physics literature appear to have already been applied to this problem. The Lyapunov function type approach to look at behavior around W∗W^* are also present in the cited literature. It is possible that some expressions for the generalization error are more computable using the authors' derivation but I fail to see how this provides any additional insight into why auxiliary learning fails or succeeds.

  2. I read a couple of appendices (B and D.1) where various approximations are made to derive the leading order terms of the generalization error and a simplified equation for the Lyapunov equation (reducing it to KxK). Are these new? If yes, these seem to be the only new contributions of the paper but they are all in the appendices!

  3. While keeping a simplified setup (freezing u, v and considering only one layer) is appealing for mathematical analysis, since each analysis technique (see 1.) has appeared before, at least some connection to how we can use these results to provide insight into auxiliary learning in practice can be useful. Also, an argument for why the effect of higher order terms in the expansion do not change the generalization error (which is also a small quantity) could be insightful as well as extensions (at least empirically) to deep networks.

Questions for authors
  1. The ODEs are discussed multiple times (and comparison with them are plotted in the figures) but never mentioned in the paper.
  2. What makes it possible to derive a Lyapunov equation of smaller size?
  3. Why does the auxiliary loss not change the minimizer when β\beta is not zero?
  4. Some statements really need clarity: e.g., "It is well understood that a linear model trained with online SGD on an MSE loss converges to a generalization error governed by the dynamical effects of training" --> generalization error in any setting depends on the training algorithm although the effects may vary. Do the authors mean that in this setting the effect of the training dynamics is substantial? Why?
  5. What is the main intuition behind your theory? Does the expression for generalization error just follow from x being Gaussian?
  6. Why does ρ\rho not affect β∗\beta^*?
  7. Why do you get the same predictions for ϵmain\epsilon_\mathrm{main} using different expressions for the order parameters compared to the ODE solutions (fixed points of the ODE) in Goldt et al? Or are they the same expressions?
1 comment

We thank the Reviewer for the time spent studying our work and for their insightful comments. To address the broader issues, we added Sec. 3.4 and Fig. 5, showing what happens when training the second layer (which still matches our theory when homogeneity degeneracy is accounted for), and expanded the Introduction section. We also expanded the discussion of the Lyapunov equation in the text to improve clarity, and moved ODEs to the main text and added a derivation sketch in Sec. 3.1 as suggested.

Critical Weaknesses 1 and 2

Our novel contributions can be summarized as follows:

a) Obtaining closed-form expressions of the generalization error and the optimal beta for the linear model as a function of ρ\rho, as well as for different noise scales between heads and for multiple auxiliary labels.

b) Applying fluctuation–dissipation theory to the large-input teacher–student framework. To our knowledge, papers in our reference list did not apply it to this setting, and particularly to the auxiliary learning problem. This application allows us to obtain both the generalization error and the optimal β\beta analytically.

c) Computing the generalization error of the one-output model for any activation function for which J2 is known analytically (while Goldt et al. required computing I2, I3, I4, and J2), and any choice of second layer weights (Goldt et al. required vi=1∀iv_i=1 \forall i). Crucially, I4 is not known analytically for ReLU (or ReLU-like) activations. Our approach is therefore the first to allow an analytical computation of this generalization error for these activations.

d) Providing novel mathematical derivations. Although our derivation of the Lyapunov equation shares elements with (and draws inspiration from) Mandt et al. (2017), there are important differences. We point to our answer to reviewer Fy32 for additional information.

We also better understand auxiliary learning in online training. In this setup, where the solution is perfectly learnable, auxiliary learning improves generalization through noise averaging and effective learning-rate reduction. For the linear model, we showed that the critical β\beta above which the auxiliary task hurts generalization is σmain2/σaux2\sigma^2_\text{main} /\sigma^2_\text{aux}, and β∗\beta^* is independent of ρ\rho. For nonlinear activations, we updated Fig. 4 to show that larger KKs move the optimal β∗\beta^* closer to the linear model one. Finally, we showed a trade-off relationship between tasks, and added App. E to show that, empirically, auxiliary learning speeds up convergence.

Critical Weakness 3

In Appendix D.6 we show the difference between our theory and ODE convergence for different learning rates. For smaller μ\mus, they match almost exactly, with a difference around 1%, which increases to 10% with μ=0.1\mu=0.1 and 40% with μ=1\mu=1. Since matching ODEs and SGD itself requires a small learning rate, the regime where the ODE description holds is the one where our first-order theory is accurate.

Questions For Authors:

  1. We agree, and we expanded Sec. 2.2 to improve clarity and moved the ODEs to the main text.
  2. We expanded the explanation in Appendix D.2-3 to improve clarity. The short reason is the isotropy of Gaussian data. Almost all input directions (N−KN-K out of NN) are perpendicular to the direction of the teacher weights and equivalent to each other, while other directions give up to O(K/N)\mathcal O(K/N) correction. Since we always choose N≫KN\gg K, we can discard them.
  3. Even when β\beta is not zero, the realizable optimum is still W∗W^*, since the first layer is shared between the two heads.
  4. We agree on the lack of clarity of this sentence; it specifically refers to a teacher-student linear model trained on Gaussian inputs with online SGD on mean-squared error. These are what are usually referred to as adaptive filters (Widrow et al., 1985). We rephrased it.
  5. We hopefully addressed the first part of the question when answering the critical weaknesses. The derivation of an analytical expression for this generalization loss does indeed depend on the Gaussian inputs. It may be possible to generalize our results to non-Gaussian inputs using Gaussian equivalence theory (Goldt et al., 2022). Online learning is ever more relevant because it mimics training large models on large datasets, each sample seen once or a few times.
  6. We find this β∗\beta^* both from the adaptive filter-like derivation and the Lyapunov equation, both matching our simulations. But this does not give any intuition for why β∗\beta^* does not depend on ρ\rho, which is true only for α=1\alpha=1 (App. F). We don’t yet have a simple intuition for this, and it remains open.
  7. Our setup differs from Goldt et al. because of the additional output weighted by β\beta and in how we derive the generalization error at the fixed point. When α=1\alpha=1 and β=0\beta=0, we expect to obtain the same generalization as Goldt et al. (at least at first order in μ\mu), which we do.

Comments on the paper

2 comments

Several observations and questions were shared by the reviewers. To address them, we made the following changes (shown in red in the updated PDF):

Clarity

-expanded the Introduction section to make our contributions clearer

-expanded the discussion of the Lyapunov equation in the main text, moved ODEs to the main text, and added a derivation sketch in Sec. 3.1 as suggested

Scope

-added Sec. 3.4 and Fig. 5 to show what happens when training the second layer (which still matches our theory when homogeneity degeneracy is accounted for)

-expanded Fig. 4, showing generalization for larger values of KK

-added App. E, an analysis of how the additional head speeds up convergence across β\betas

Related works

-added Mori et al. (2025) to the previous literature

Area_Chair_pdQxOpenReview ↗

Discussion Period

Dear Reviewers,

Thank you for your reviews!

The authors have submitted their responses. Please read through their rebuttal, acknowledge their reply, and follow up with them if any of your concerns haven't been fully addressed or if you have additional questions.

Please make sure to engage in the discussion before the deadline on November 18 (AoE).

Best,

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.