acceptodds
Withdrawn from ICLR 2027

Self-Supervised Learning with Recursive Quadrature Filters for Neural Time Series

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

12 views

Abstract

We develop self-supervised methods for training Recursive Quadrature Filter (RQF) encoders on unlabeled time series. RQFs are biologically motivated, complex-valued temporal filters that form a special class of diagonal state-space models (SSMs). Each RQF acts as a band-pass filter whose selectivity is determined by two learnable parameters: its tuning frequency and bandwidth. To train these encoders, we introduce a self-supervised learning objective based on prediction error; the encoder and a prediction head are trained to predict targets derived from the encoder’s future states. In the windowed mode, the head uses the summary of the current window to predict the summary of a later window; in the dense mode, it uses the state at the current timestep to predict future log-power. Each RQF encoder is pretrained on a dataset and then a readout (e.g., a task-specific classifier) is trained on the output of the RQF encoder, with the encoder either frozen or fully fine-tuned. For multichannel EEG and intracranial recordings, a shared encoder processes each channel separately, and a spatial readout combines the channel representations using electrode locations. Across four neural benchmarks using scalp or intracranial EEG, the pretrained RQF system performs close to the state of the art while using a small fraction of the parameters of the competing models. For example, a 110k-parameter RQF system reaches 0.9130 ± 0.0036 test AUROC on TUAB, within the range achieved by state-of-the-art EEG foundation models with 1.0M–369M parameters.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
self-supervised learning, state space models, eeg, ecog, neural time series

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_Rff8OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a new model base on Recursive Quadrature Filtering. This model gets very good performances over different EEG tasks, almost reaching the performances of FM models for EEG with way less parameters. To go a bit further the paper proposes different study for the pre-training setup showing in which setup the model fails.

Critical strengths
  • The idea to use RQF to embed the signal allows to give more meaningful representation.
  • The RQF method allows to get almost same performances with way less parameters, showing that smart model is sometimes better.
  • The paper is pretty easy to read.
Critical weaknesses
  • The paper lacks of motivation for the new model they propose. For example the introduction is unbalanced. Most of the introduction is focus on contribution of the paper (5 paragraphs out of 6). It would be interesting to have a better introduction to the field rather than one page on the contribution. This introduction could point out what are the missing part of the field that RQF try to fill.

  • This comes to the second point. If RQF is an interesting method, in practice model like Reve or CBraMod have better results. I understand that RQF has way less parameters but a study of the computational time to fine-tune the backbones could be interesting to see that in practice using RQF reach almost the same results with less computational time.

  • My last point that is the most critical, I'm wondering how trustful are these results. The majority of the values on the tables are from initial paper. But is the split and pre-processing the same ? Looking at CBraMod paper, for TUAB they used specific pre-processing and a split of 80/20%. It seems that is not what the paper is doing. Nothing in the paper is telling us that the tables are fully comparable because the splits are the same.

Questions for authors
  • In plot top right in Figure 2, USleep is said train only on ISRUC, but in the orignal paper it is trained on separate corpus no? Do you re-train it?

  • Do you think that having this kind of representation could help to have more interpretable features for our signals ? In this case it could be interesting to try to interpret the features.

  • I'm wondering how much your work can be related to the work of [1], where they propose filtering in the frequency space EEG for better generalisation? I'm wondering if this step could be applied to the different filter you get.

[1] Gnassounou et. al., PSDNorm: Test-Time Temporal Normalization for Deep Learning in Sleep Staging

Reviewer_KL2jOpenReview ↗
Rating
2
Confidence
1
Informativeness
2
Presentation
1
Soundness
3
Read review
Paper summary

I have to apologize in advance: I don't know why I was assigned to review this paper as I have a very different background, both in terms of application and algorithms involved in this work. My understanding of this work is that they propose to use RQF, which resembles SSM such as MAMBA, though with domain-specific architectural choices in combination with self-supervised learning choices to pre-train models. They ablate downstream performance as a function of the number of available labelled data-points. They ablate freezing or tuning the encoder during supervised fine-tuning. The experimental evaluation seems extensive, the appendix revealing many training details.

Critical strengths

The main strength of this paper lies in the RQF formulation and the evidence the authors provide on the effectiveness in reducing parameters for the given task performance (Figure 2 as the headline figure and multiple further tables). The pre-training vs. no-pretraining control in combination with the fine-tuning study is well-done.

Critical weaknesses

The main weakness as far as I can tell is in their disambiguation on what exactly contributes: The architecture or the self-supervision algorithm? Results show that the improvement of using SSL can be small: TUAB 0.9092 -> 0.9130. BrainTreebank: 0.8709 -> 0.8997; ISRUC-SG1: 0.7953 ->0.7990 Ideally, the authors would compare their architectural choice with several SSL approaches, such as those requiring teachers and/or masking and compare their SSL approach to different architectural choices, ideally parameter-count matched.

The authors are straight-forward about some of the limitations on their experimental evaluation and that setups might not be comparable. How valuable do direct comparisons remain given those limitations?

Given that performance with&without pre-training is often quite close, it might be interesting to seed the pre-training runs in addition to seeding the fine-tuning runs.

Questions for authors

See above

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

The paper pretrains Recursive Quadrature Filter (RQF) encoders with a self-supervised prediction objective. RQFs are complex-valued band-pass filters, hence diagonal state-space models, with learnable frequency and damping. Two objectives are compared: a windowed one that predicts a later window’s summary from an earlier one with InfoNCE and a variance hinge, and a dense one that predicts future log-power from a single state; neither uses masking, a quantizer, or an EMA teacher. A frozen or fine-tuned encoder feeds a task readout. Experiments cover TUAB, BrainTreebank, ISRUC-SG1 and BCI-IV-2a, with a paired-initialization control and a label-efficiency sweep.

Critical strengths
  1. The control design measures what pretraining buys. The paired-initialization control starts from byte-identical weights without pretraining, turning “pretraining helps” into a number: +3.8 AUROC on TUAB with a frozen encoder, +13.2 balanced accuracy at 1% labels, and almost nothing at 100% labels.
  2. The parameter-efficiency result is strong. 110k parameters reach 0.9130 AUROC on TUAB, between LaBraM-Base (5.8M) and LaBraM-Huge (369M); 293k reach 0.9102 on BrainTreebank, 5.7 points above BaRISTA at 3.4× the parameters.
Critical weaknesses
  1. Most of the headline performance comes from the architecture, not the objective. Under full fine-tuning SSL adds only +0.38 AUROC on TUAB, +0.37 on ISRUC and +0.87 on BCI-IV-2a, all within seed noise, while the title and abstract still credit the objective.
  2. The strongest evidence rests on one small subset and one checkpoint. The 1% label gains come from a single 16-patient, 279-segment subset, and the variance covers only three downstream seeds.
Questions for authors

How much does the 1% label gain vary across label subsets and pretraining checkpoints? What is the BrainTreebank result under participant-disjoint splits, and does the margin over BaRISTA survive? Does the gain over paired initialization come from the learned (ω, γ) spectrum or from the readout?

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

Related papers

Discussion (0)

Sign in to comment.