acceptodds
Withdrawn from ICLR 2027

Functional Autoencoders for Amplitude-Phase Representation Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Functional data are intrinsically infinite-dimensional, and often exhibit phase variation, where corresponding events occur at different times across observations. Existing linear dimension reduction methods struggle with nonlinear amplitude variation, while functional autoencoders without an explicit warp entangle temporal misalignment with shape. We propose the Amplitude–Phase Functional Autoencoders (AP-FAE), an unsupervised framework for functional data that spans both univariate and multivariate cases, with emphasis on the multivariate setting, and factorizes the latent space into separate amplitude and phase embeddings derived from all channels. A smooth functional decoder reconstructs channel-specific amplitude functions in canonical time, and a shared monotone, endpoint-preserving warp captures phase variation. We prove a bound linking amplitude recovery to registration, reconstruction, and noise errors, and validate it numerically. Across synthetic data and six real-world benchmarks, AP-FAE outperforms state-of-the-art baselines on most clustering and alignment metrics and on all reconstruction metrics. Clustering with amplitude embeddings alone consistently surpasses joint amplitude–phase clustering, confirming the benefit of explicit disentanglement. Code is available at https://anonymous.4open.science/r/APFAE-418C/.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Functional Data Analysis, Autoencoders, Registration, Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VKE7OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

To deal with functional data with phase variation, the paper proposes a network architecture with the following features: (1) a functional layer to map the multi-dimensional function into a multi-dimensional vector ss; (2) use one MLP branch to learn an amplitutde-related vecor aa, and another MLP branch to learn a phase-related vecor pp; (3) the vector aa after transformation is used basis-expansion coefficients to construct the multi-dimensional function ff, and the vecor pp after MLP transformation is used to construct the warping function w. The network is trained by minimizing the reconstruction error between the original x and the reconstruction f(w).

Critical strengths

None

Critical weaknesses

Overall, the paper's technical contribution appears incremental, primarily combining an existing functional autoencoder with a monotone warping decoder. The central claim of amplitude-phase disentanglement is not adequately established.

(1) The architecture implicitly assumes that the finite-dimensional projection vector ss preserves sufficient information to recover both amplitude and phase. However, the authors do not establish under what assumptions this is possible, particularly for complex warping functions. Information discarded by the projection cannot be recovered by subsequent network layers.

(2) There is no guarantee that the two branches will learn amplitude and phase separately. Why the amplitude branch will lean the amplitude information only, and the phase branch lean the phase information only? The training objective (i.e., reconstruction loss) can not help disentangle the phase from the amplitude. Even a global minimum of the reconstruction loss need not identify the intended decomposition. Moreover, learning the warping function involves non-convex optimization, and the authors provide no guarantee that the estimated warp is close to the true one.

(3) Although the implementatio subtracts the maximum network output before exponentiation to prevent overflow, the normalized exponential in Eq. (8) may still produce highly concentrated warp densities. These can yield nearly flat regions, steep local transitions, and poorly conditioned gradients.

(4) In Theorem 1, the amplitude recovery bound depends on both reconstruction error ϵx\epsilon_x and ϵg\epsilon_g. Training directly minimizes the former, but the theorem provides no guarantee that the latter is small. Moreover, ϵg\epsilon_g generally cannot be evaluated without knowing the true warp. The result therefore describes amplitude recovery conditional on accurate registration, rather than establishing that the proposed method achieves it. Simply verifying that the inequality holds does not establish accurate recovery.

(5) The functional layer does not explicitly control redundancy among the learned functional weights. Multiple filters could capture similar dominat modes of variation, reducing the effective dimension of the projection vector ss.

Questions for authors

The concerns above may be challenging to resolve fully, the following issues in the numerical evaluation are more straightforward and must be addressed.

(1) Some benchmark methods do not explicitly account for phase variation, so improvements over them may reflect the benefit of accommodating temporal deformation rather than successful amplitude-phase separation. Although alignment-aware baselines are included, a registration-then-autoencoding pipeline would help establish the value of the proposed joint architecture. For example, a clear baseline is elastic registration + a functional autoencoder. The Tucker, Wu, and Srivastava paper (reviewed in the introduction) is also a strong baseline.

(2) The simulation provides a limited assessment of amplitude-phase separation. The reported three metrics do not directly establish accurate amplitude and phase recovery. The warping functions in Fig. 3 belong to a one-parameter exponential family and are globally convex/concave. Increasing the parameter range changes deformation severity but not the complexity of the family. Moreover, Fig 3 compares leading PCAs, rather than subject-level amplitude recoverty. Comparing clustering performance using aa, pp, and (a,p)(a, p) does not establish that these codes selectively represent amplitude and phase.

Reviewer_B6CLOpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The authors propose an unsupervised framework that jointly learns nonlinear functional representations and aims to disentangle amplitude and phase variation through separate embeddings. They establish a theoretical bound quantifying how registration, reconstruction, and noise errors jointly control amplitude recovery. Their architecture consists of a functional encoder that embeds the input functional data using a B-spline basis set for both amplitude and phase. The amplitude and the phase decoders then separately reconstructs both the amplitude and phase (that relies on warps). The authors show results on both synthetic and real datasets.

Critical strengths

The paper tackles an important problem in functional data analysis.

The idea of the decoder separeing the channel-specific amplitude functions and a shared inverse warp is interesting. This has several applications, especially for data from synchronized channels.

The method is simple to implement. Owing to the construction, both the memory and time complexity is linear in the number of samples, the number of functional channels, and the number of grid points and their combinations.

In the synthetic data, the method is used to demonstrate the amplitude and the phase separation. The method also performs well on real data.

While there is a problem with identifiability of warps, the authors do state that the identifiability of amplitude and phase is up to a common aligned coordinate and that amplitude recovery is contingenet on registration error, reconstruction error, and the noise level.

The authors have also released their code.

Critical weaknesses

There are several concerns about novelty and related work is not adequately discussed. There were some early ideas that attempted to use auto-encoders and/transformers to extract the warping function (although their goal was not to perform amplitude and phase separateion). For e.g. the works by Koneripalli et al., Rate-Invariant Autoencoding of Time-Series (ICASSP 2020), Shapira Weber et al., Diffeomorphic Temporal Alignment Nets (DTAN, NeurIPS 2019), Nunez et al., SrvfNet: A Generative Network for Unsupervised Multiple Diffeomorphic Functional Alignment (CVPR Workshops 2021). This is distinct from the SrvfRegNet baseline, which the authors have cited and used in the paper. Especially the idea in Koneripalli uses an unsupervised autoencoder that constructs a latent space of warps and then decodes a canonical signal represetnation. DTAN learns input-dependent diffeomorphic warps for unsupervised joint alignment, whereas SrvfNet learns warps and a template with a variational autoencoder for SRVFs. Thus the claim of "first unsupervised work to learn functional representations..." couldbe appropriately revised and restated.

The idea of identifiability of the amplitude and phase, especially as it pertains to disentanglement is not adequately demonstrated. The recovery of the amplitude and phase is shown mainly at the population level (comparing functional principal components) or qualitatively (Figure 3b). No individual signal/channel level measures/comparisons are shown. For e.g. one ca compute relative amplitude or phase errors (angles) to determine how the method performs amplitude and phase recovery. Similary for the clustering results, the metrics such as ACC measure clustering accuracy, adjusted total variance (ATV) measures registration, and mean squared error (MSE) measures the reconstruction fit. They don't exactly measure how well you actually recover the amplitude and the phase.

There are no constraints on the warping functions and thus it is not known if extreme dilation or expansion of the time warps can result. For e.g. the earlier works DTAN, SRVfnet etc. regularized the warps including those by first and second derivatives. It is possible that hte low dimensional phase representation may thus provide implicit regularization. But the authors should clarify whether explicit warp penalities are needed.

One of the criteria for performance of the recovery (disentanglement) is registration accuracy. However, this is not explicitly achieved in the architecture, but rather through the reconstruction (training loss) perspective. Thus it is not clear how the estimated warps are aligned across subjects, which is an important step for phase recovery. The synthetic examples do provide qualitative exmaples of warp recovery, but no quantitiative evaluation of the registration mechanism is included.

The experimental performance for AP-FAE is not decisively better. For e.g. SRC is still a good contender across datasets. The improvements in clustering accuracy on ERing, Plane, and Symbols are modest, while SRC was able to outperform AP-FAE on UWaveGestureLibrary. As stated above, neither the reconstruction MSE nor adjusted total variance could be used to establishe accurate recovery of the underlying amplitude and phase components.

Questions for authors

IS there an attempt to define the common aligned coordinate that was mentioned in Remark 1? This question has implications for registration and reconstruction accuracy.

Can you quantiativeely demonstrate performance of actual amplitude and phase recovery. I understand it may not be possible on real data, especialy if those measurements are not already known. However, what about synthetic data. In addition to the experiments performed in the paper, also show specific metrics for recovery performance.

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

This paper introduces AP-FAE an unsupervised functional autoencoder model for univariate and multivariate curves. It splits these signals into two embeddings: an amplitude embedding, decoded as B-spline curves on a common time axis and the phase embedding into a monotone time warp shared across channels. The model is trained with a reconstruction loss only. The authors derive a bound relating amplitude-recovery error to registration, reconstruction and noise errors, and evaluate the method on synthetic data and six UCR/UEA datasets, reporting clustering, alignment and reconstruction results against functional clustering, registration and autoencoder baselines.

Critical strengths

Main strengths of the paper:

  1. The paper introduces a simple but sensible idea: The model splits each curve into two separate embeddings: one for its shape and one for its phase. The phase part always produces a valid time warp and it is shared by channels recorded together. The pre-training set-up is simple but effective with only a reconstruction loss, such that the pre-trained embeddings can be re-used for various downstream tasks.
  2. Multiple experiments on synthetic data validates the method: On synthetic data, where the true shapes and warps are known, the paper shows that both are recovered. The method is also tested on six real datasets from three angles (clustering, alignment, reconstruction), and under missing data, irregular sampling, extreme or absent timing differences, and large datasets.
Critical weaknesses

Main weaknesses of the paper:

  1. The amplitude–phase disentanglement is claimed but not directly shown: The main evidence is that clustering on the amplitude embedding works better than clustering on both embeddings. This is indirect: in the synthetic data, phase is unrelated to the classes by design, so adding it can only hurt, and on one real dataset (BasicMotions) the phase embedding alone clusters almost as well as the amplitude embedding (0.800 vs 0.807). Nothing in the training objective prevents phase variation from leaking into the amplitude embedding, and the theorem assumes the warps are estimated well rather than showing it. A simple check on the synthetic data, e.g., whether the phase embedding predicts the true warps while the amplitude embedding does not, would make the claim convincing.
  2. Some basic/baseline comparisons and error bars are missing: The paper argues that learning amplitude and phase jointly is better than the classic two-stage approach (register the curves first, then apply dimension reduction such as FPCA), but never tests that approach, nor standard clustering methods that already handle phase variation (e.g., based on dynamic time warping, DTW). Results are averaged over 50 runs without standard deviations, so small differences (e.g., 0.875 vs 0.868 on Plane) cannot be judged.
  3. The usefulness of the embeddings for downstream tasks is not demonstrated. The embeddings are presented as task-agnostic and reusable (the introduction mentions regression with both embeddings), but the only downstream use evaluated is K-means clustering with the true number of classes. A supervised evaluation on the real datasets, e.g., a linear probe or k-NN classifier on the amplitude, phase and combined embeddings using the standard UCR/UEA train/test splits, compared with FAE embeddings and the raw curves, would show whether the learned embeddings actually improve downstream performance.
Questions for authors

Some questions for clarification:

  1. Could you report the three terms of the bound separately in Table A1?
  2. How were the hyperparameters chosen, and were labels used (Figure 5)? What varies across the 50 repetitions on the real datasets?
  3. On the synthetic data, can you show directly that the two embeddings separate amplitude and phase, e.g., by reporting how well the true warp parameter can be predicted from the phase embedding versus the amplitude embedding?
  4. Could you evaluate the embeddings on a supervised task, e.g., a linear probe or k-NN classifier on the amplitude, phase and combined embeddings using the standard UCR/UEA train/test splits, compared with FAE embeddings and the raw curves?

Suggestion for future work:

  • Phase variability is well documented in neural signals, e.g., trial-to-trial latency jitter of event-related potentials in EEG or of high-gamma responses in iEEG. Applying AP-FAE to such data, and evaluating the embeddings on standard downstream decoding tasks against task-specific deep learning models and EEG foundation models (e.g., Banville et al 2025 - 2605.08495), would show whether explicit amplitude–phase disentanglement helps. It would also be worth discussing whether the shared-warp assumption holds for such signals, as it may suit event-related responses better than oscillatory activity.

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

Related papers

Discussion (0)

Sign in to comment.