acceptodds
Under review as a conference paper at ICLR 2027

LBTAR: Learnable Brain-Topology-Guided Autoregressive Channel Recovery for EEG Spatial Super-Resolution

Abstract

High-density electroencephalography (EEG) captures rich spatial information, but its acquisition and deployment costs are relatively high. Low-density EEG is easier to acquire, but its limited spatial coverage makes it difficult to characterize regional brain activity at a fine spatial scale. Therefore, reconstructing high-density EEG from low-density observations has become an important problem in EEG spatial super-resolution. Existing methods mostly adopt one-shot reconstruction or fixed-order autoregressive recovery, which limits their ability to adapt to sample-specific channel dependencies and dynamically evolving recovery states. To address this limitation, we propose LBTAR, which replaces predefined recovery paths with a state-dependent sequential recovery process. LBTAR treats EEG channels as nodes and integrates channel-to-region assignments, inter-regional adjacency relationships, and electrode spatial information to construct a brain-region-structure-guided channel topology. It further jointly considers the currently visible context, the remaining candidate set, the dynamic channel topology, and channel-level prediction reliability to adaptively select the next group of channels for recovery. At each step, the mask-conditioned reconstruction network produces full-channel predictions, while only the selected channel signals are written back to progressively update the recovery state. Channel-level uncertainty is further used to adjust subsequent topology and recovery decisions, thereby mitigating error propagation. Experiments on three public datasets, namely SEED, SEED-IV, and Localize-MI, show that LBTAR achieves favorable time-domain reconstruction performance and frequency-domain consistency across multiple spatial super-resolution ratios, while maintaining relatively low computational overhead and stable performance at high missing-channel ratios. Ablation experiments further validate the effectiveness of dynamic recovery, brain-region structural priors, and uncertainty-aware modeling.

Primary area:
applications to neuroscience & cognitive science
Keywords:
EEG spatial super-resolution, autoregressive recovery, brain topology, channel reconstruction, uncertainty-aware modeling, learnable recovery policy

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_kA9MOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the problem of reconstructing high-density EEG from low-density observations by proposing LBTAR, which addresses the difficulty existing methods face in adapting to sample-specific channel dependencies and changing recovery states. Its main contribution is incorporating the recovery order into the learning process. The method combines brain-region structural priors, the current recovery state, and prediction reliability to adaptively select the channels written back at each step. Uncertainty feedback adjusts channel connectivity and subsequent selections, with the aim of mitigating error propagation during stepwise recovery. Comparative experiments show that LBTAR outperforms the evaluated baselines while maintaining relatively low computational overhead. Ablation experiments further support the roles of dynamic channel selection, brain-region structural priors, and uncertainty modeling.

Critical strengths
  1. LBTAR allows the model to determine the channel recovery order based on the current state, adapting to channel dependencies across samples and recovery stages. Brain-region topology provides structural information beyond electrode distances; stepwise write-back enables subsequent predictions to use recovered channels; and uncertainty modeling assesses prediction reliability to guide subsequent recovery. Together, these designs address the limitations of a fixed recovery order.
  2. Compared with the evaluated methods, LBTAR achieves better reconstruction performance across datasets and spatial super-resolution ratios. Ablation experiments support the contributions of dynamic channel selection, brain-region topology, and uncertainty modeling. The model also has a relatively low parameter count, computational cost, and inference time, suggesting potential for deployment.
Critical weaknesses
  1. The paper randomly divides preprocessed samples into training, validation, and test sets without subject-level separation. Related samples from the same subject may therefore appear in different sets, so the results do not establish generalization to new subjects.
  2. The appendix provides training hyperparameters, some stage-wise recovery settings, and descriptions of ablation variants, but does not clarify how brain-region labels and adjacency matrices are constructed or specify the reconstruction backbone architecture. It also reports results using only a fixed random seed of 42, without variation across runs or statistical tests. These gaps make the method difficult to reproduce and the stability of its reported advantages difficult to assess.
  3. The non-differentiable Top-K operation is handled by a soft auxiliary loss with stop-gradient, which is a standard heuristic rather than a new methodological contribution. The uncertainty head is also a common log-variance weighting mechanism. Overall, the technical novelty is incremental and closer to an engineering combination than a fundamentally new formulation.
  4. Only three datasets are used, and two of them (SEED and SEED-IV) are same-lab 62-channel emotion datasets, which limits domain diversity. The paper does not report cross-dataset or cross-subject generalization, does not evaluate on additional high-density EEG datasets or different electrode layouts/montages, and does not test robustness to realistic channel corruption or subject variability.
Questions for authors
  1. Why did the authors use a random sample split instead of a subject-level split? Under a subject-level split, would LBTAR still outperform the evaluated baselines?
  2. How are the brain-region topology and reconstruction backbone constructed, and which training write-back strategy was used in the reported experiments?
Reviewer_9tVfOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes LBTAR, a learnable brain-topology-guided autoregressive framework for EEG spatial super-resolution. Given a subset of observed EEG channels, the method progressively reconstructs the missing channels over several recovery stages. At each stage, a shared mask-conditioned reconstruction network predicts all channels, while a learned policy selects a subset of missing channels whose predictions are written back into the current state and used as context in later stages.

The recovery policy uses the current channel representations, the visible and unrecovered channel sets, a topology combining electrode distance and brain-region relations, and a learned estimate of channel-level prediction reliability. An uncertainty head predicts channel-level log-variance, which is used both as a training objective and to attenuate subsequent graph connections. Because hard Top-K selection is non-differentiable, the policy is trained with a soft auxiliary objective based on current channel reconstruction errors. Scheduled sampling is used to reduce the mismatch between training-time write-back and inference-time self-rollout.

The method is evaluated on SEED, SEED-IV, and Localize-MI at several channel-subsampling ratios. The reported results show substantial improvements over spatial interpolation, GAN-based, Transformer-based, diffusion-based, and fixed-order autoregressive baselines in time-domain and spectral reconstruction metrics. The paper also reports ablations of sequential recovery, learned channel ordering, topology components, and uncertainty feedback, as well as parameter count and inference cost.

Critical strengths
  1. The paper addresses a practically relevant EEG reconstruction problem and proposes a coherent extension of fixed-order autoregressive recovery. Treating the recovery order as state-dependent is well motivated because the available context changes after every write-back step. Incorporating estimated prediction reliability into subsequent topology construction is also a reasonable mechanism for reducing error propagation.

  2. The empirical study covers multiple datasets, channel-subsampling ratios, reconstruction metrics, and method families. The ablations separately examine one-shot versus sequential recovery, random and geometric recovery orders, spatial and region-level topology, shuffled region assignments, and uncertainty feedback. The model is also compact and has low reported inference cost relative to the selected neural baselines.

Critical weaknesses
  1. The random sample-level train-test split creates a serious risk of leakage and limits the generalization claim. For SEED and SEED-IV, continuous recordings are first divided into non-overlapping four-second segments and the resulting samples are then randomly split. This can place segments from the same subject, session, trial, and nearby time periods in both training and test sets. Localize-MI trials are also randomly split without a patient- or recording-level separation. EEG spatial covariance is strongly subject- and session-dependent, so performance under this protocol may largely reflect interpolation within already observed subjects and recording conditions. The paper does not provide subject-independent, session-independent, or trial-grouped evaluation, and therefore the reported results do not establish generalization to new users or independent recordings.

  2. The experiments do not adequately validate the paper's central claim of sample-specific, state-dependent recovery. The ablations compare the proposed policy mainly with random order and a hand-designed geometric order. They do not include a learned but sample-independent global ordering, a fixed ordering obtained from average training-set prediction errors, or a policy without access to the evolving channel representations. The paper also does not report how often the selected order actually differs across samples or recovery states. Consequently, the gains may arise from sequential reconstruction, additional supervision, topology, or a generally favorable easy-channel-first ordering rather than from sample-specific adaptive decision making. The hard Top-K decisions are trained only through a soft auxiliary loss, creating a further mismatch between the optimized policy distribution and the actual discrete rollout.

Questions for authors
  1. Can the authors provide subject-independent and session- or trial-grouped evaluations? For SEED and SEED-IV, were four-second segments from the same original trial randomly distributed across training, validation, and test sets? For Localize-MI, can trials from the same patient or stimulation condition occur in multiple splits? Please also report results using a split in which no subject, or at minimum no original trial or recording block, is shared between training and testing.

  2. How sample-specific and state-dependent is the learned recovery order in practice? Please report the diversity or agreement of selected channel groups across samples and stages, and compare against a learned but sample-independent global order under the same reconstruction backbone. A fixed order obtained from average training-set channel errors would also help determine whether the proposed policy learns dynamic decisions or mainly an easy-channel-first curriculum.

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

The paper proposes LBTAR, an autoregressive approach for EEG spatial super-resolution, focusing on reconstructing high-density EEG signals from a limited number of observed channels. The proposed LBTAR builds on previous approaches to sequential channel recovery and introduces a learnable recovery policy that is used to determine the order in which the channels are reconstructed. Specifically, LBTAR leverages a channel topology, a channel graph built from electrode positions corresponding to brain regions. Using such a structural prior together with evidence from previous recovery steps, the proposed learnable recovery policy ranks the remaining channels to be recovered and gets updated. LBTAR also includes an uncertainty estimation mechanism that adjusts the topology and subsequent selection decisions to reduce error propagation. The experimental analysis includes experiments on 3 datasets (SEED, SEED-IV, and Localize-MI) at different scales, ranging from 2-16x, comparing against 5 baseline methods (SI, EEGSR-GAN, ESTformer, SRGDiff, CAFE). Reported results show that LBTAR can lead to improved performance while being more efficient at inference. The authors also provide an ablation study, evaluating the effects of different LBTAR components.

Critical strengths
  • The motivation of the work is clear and relevant, particularly for applications where EEG acquisition is infeasible.
  • The ablation study is informative, considering alternative recovery strategies, different structural priors, and uncertainty modeling.
  • The proposed model is both smaller and faster compared to other approaches.
Critical weaknesses
  • W1 I am not entirely convinced by the formulation of the learnable recovery policy. As stated in Eq.21, the policy prioritizes channels with lower reconstruction error. While this is one way to minimize error propagation in subsequent steps, it does not necessarily mean that the ones that are the easiest to reconstruct are the most informative/beneficial for the remaining steps. Therefore, it's unclear to what extent the policy captures these effects. Maybe a comparison against an oracle (based on the final reconstruction) or with a predetermined order can be beneficial.

  • W2 The experimental setup is rather narrow, consisting of only 3 datasets, two of which have rather related acquisition protocols (SEED and SEED-IV). Although the authors report improvements across the 3 datasets, it is unclear how well these findings may generalize to signals with different spatial and temporal characteristics, such as motor imagery, ERP, or sleep recordings etc. Even in the reported experiments, there is a discrepancy between Localize-MI and the two SEED datasets.

  • W3 The evaluation protocol is also concerning when it comes to generalization to unseen subjects. As stated in the manuscript (App. E), the samples are randomly split into train/test, meaning that recordings from the same subjects and sessions can appear in both splits. As the EEG spatial characteristics can be very subject-specific, it is unclear whether the model learns generalizable properties or overfits to the subjects. Therefore, in this case, a subject-independent evaluation, like LOSO (leave-one-subject-out) would provide stronger evidence of performance, especially wrt to new subjects.

  • W4 The work, as it is, is hardly reproducible. There are neither sufficient details about the model parameters and design choices given, nor is the code provided.

  • W5 The main methodological contributions of the work are rather narrow. Some of the key elements have already been published (like the sequential group recovery).

Questions for authors
  • See weaknesses

  • How were the baselines trained? Why are the SRGDiff scores about two times worse than in the original paper (NMSE of 0.31 vs. 0.16 on SEED at 2×), under what is reported to be an identical protocol to the one in the SRGDiff paper?

-Could the authors report the mean and standard deviation over several runs with different seeds? Do the differences/gains in the ablation experiments remain?

  • It would be interesting to see whether the reconstruction order is similar or substantially different across subjects.

  • Could the authors clarify how the brain region assignments and adjacency matrix are constructed (Eq. 4 and Eq. 5)? In particular, which regional partition is used for the different electrode configurations, and what criteria are used to determine whether two regions are considered adjacent?

  • In terms of efficiency, are all baselines run under identical hardware settings and training conditions? Have the authors analyzed how the individual components, particularly the learnable recovery policy, affect the overall computational efficiency?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.