acceptodds
Withdrawn from ICLR 2027

Online Newton Learning for Multi-Output Nonlinear Models under Non-i.i.d. Data

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

We study online learning for a class of multi-output nonlinear stochastic regression models under non-i.i.d. data and general convex losses. This problem arises in several representative settings, including multi-output linear regression, learning from component-wise saturated observations, and multiclass softmax classification under the KL-divergence loss. We develop an adaptive Newton-type estimation algorithm that combines observation-driven gradients, recursively accumulated second-order information, and local curvature of the nonlinear function, allowing its updates to adapt to the evolving geometry of sequential data. We then establish parameter convergence without imposing i.i.d. data assumptions, requiring only an information-growth condition that matches the minimal requirement for least- squares estimation in linear stochastic regression. Moreover, we show that the average prediction regret decays at a logarithmic-over-linear rate without requiring additional excitation on the data. Experiments on synthetic and real-world datasets corroborate the theoretical results and demonstrate the effectiveness of the proposed method.

Primary area:
learning on time series and dynamical systems
Keywords:
Online learning, Non-i.i.d. data, Nonlinear stochastic model, Newton-type algorithm, Convergence

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_sLSEOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The manuscript considers estimation for the model y_{k+1} = g(Θ*ᵀφ_k, e_{k+1}) by using Newton-type recursion. As the main results, authors derive average error of size log n/ n and prove strong consistency (Theorem 1-2 and Corollary 1-2).

Critical strengths

The unified framework provides clear error analyses on model class covering various interesting examples. Examples provided in the manuscript are clear and important, with sufficient level of details presented.

Critical weaknesses

The error rates are classical and exactly what are expected, and the novelty of the manuscript remains partially unclear.

Questions for authors

All the results follow from Lemma 1 in the appendix. Could one position clearly the novelty of Lemma 1 with respect to the literature?

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

This paper proposes a projected online Newton-type algorithm for multi-output nonlinear stochastic regression under non iid data and a class of prediction losses. Under suitable assumptions, it establishes almost-sure logarithmic cumulative prediction error without persistent excitation, together with strong parameter consistency under an additional condition.

Critical strengths

The paper tries to extend the results on multi-output with non iid data online learning to nonlinear models with log-over-linear rate under structural conditions. Comparison between the proposed method with standard methods shows the improvements on selected tasks.

Critical weaknesses
  1. The analysis of this paper is based on a model whose geometric landscape is restricted, especially shown in a series of assumptions from assumption 1-6. Some of the assumptions is not common. For example, assumption 5 requires a symmetric positive-defined Jacobian for the prediction. It is quite stronger than the common L-smooth condition because it needs additional lower bound with the smooth constant mfm_f. The similar situation happens in assumption 6. Combined with assumption 3, the model can be roughly regarded as a loose quadratic. That means the analysis is quite simplified and the claimed non iid data seems not to be a stronger strength compared with the model restrictions, and same for the contributions of this paper.

  2. The synthetic experiments are only included the example 1, without the claimed nonlinear and non-quadratic cases. The comparisons lack the direct comparison with Newton method. Ablations are also required.

Questions for authors
  1. I wonder if the assumption is standard aligned with the traditional SGD analysis, how does the result change?
  2. Why use the PAMAP2? Does else dataset not work?
Reviewer_SSqgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes an online Newton method for multi-output nonlinear models under non-i.i.d. data. It provides convergence guarantees for prediction and parameter estimation, supported by synthetic and real-world experiments.The distinction between prediction convergence and parameter consistency is interesting, although the underlying approach is not entirely new.The theoretical analysis is reasonably developed, but I have concerns about whether the experimental settings satisfy the required convergence conditions.

Critical strengths

The convergence guarantees under weak excitation conditions are theoretically interesting. The experiments show competitive results against existing online optimizers.

Critical weaknesses

1.The PAMAP2 hyperparameters appear inconsistent with the condition (C>0) in Eq. (25). This raises questions about how the theoretical guarantees relate to the experiments. 2.The empirical evaluation relies on only one real-world dataset. The lack of released code and unclear hyperparameter selection procedure also limit reproducibility.

Questions for authors

1.Do the experimental hyperparameters satisfy Eq. (25)? If not, how do the theoretical guarantees apply? 2.How were the hyperparameters selected? Was an independent validation set used? 3.Assumption 1 does not seem to ensure that the optimal predictor f is time-invariant. Is an additional assumption needed when the noise distribution changes over time? 4.What is the specific improvement over Zhang and Zhang (2025) that the authors regard as the central contribution, and can they support it with a controlled comparison or a simple ablation of the local-curvature update? This would have the greatest impact on my assessment.

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

Related papers

Discussion (0)

Sign in to comment.