PaperlayerView exact revision on arXiv

Source-linked AI summary

EEG-AS: Instance-Level Foundation Model Selection for EEG Foundation Models via Behavior Reconstruction

Yunzhen Zhang, Ruoxi Piao, Hasan Onur Keles, Mustafa Misir

arXiv:2609.00653v1cs.LGcs.AI

TL;DR

Because foundation-model performance varies across EEG instances and no single model is consistently best, efficient instance-level selection is needed. EEG-AS reconstructs unavailable model behaviors from inference-available EEG representations and an anchor model, achieving stronger selection across seven benchmarks while narrowing the SBS-to-Oracle gap.

  • Problem

    Foundation-model performance varies across individual EEG instances, while instance-level selection among EEG foundation models remains largely unexplored.

  • Method

    EEG-AS formulates foundation-model deployment as per-instance algorithm selection and reconstructs unavailable candidate behaviors from EEG representations and a single inference-available anchor model.

  • Results

    EEG-AS achieves the highest average performance among deployable methods, obtains the best result on five datasets, and outperforms SBS on all seven datasets.

  • Takeaways & Limitations

    Instance-level algorithm selection supports adaptive deployment of EEG foundation models without exhaustively executing the entire portfolio.

  • Takeaways & Limitations

    Selection gains vary across datasets, and remaining performance depends on identifying decision-relevant model behaviors; future work will incorporate accuracy, inference time, and computational resources.

Abstract

from arXiv · show

Electroencephalography (EEG) is a non-invasive technique for measuring neural activity and has been widely used in neuroscience applications. Recent advances in EEG foundation models have enabled strong performance across diverse neural decoding tasks. However, no single foundation model consistently performs best across datasets or individual EEG instances, while instance-level model selection remains largely unexplored. To address this limitation, we formulate EEG foundation model selection as an instance-level Algorithm Selection (AS) problem. We propose \textbf{EEG-AS}, an instance-level algorithm selection framework that characterizes each EEG instance using inference-available latent EEG embeddings, handcrafted neurophysiological features, and an anchor foundation model. During training, EEG-AS learns to reconstruct unavailable foundation-model behaviors from privileged prediction tokens conditioned on an anchor foundation model, while during inference it estimates these behaviors without executing the entire model portfolio, enabling efficient selection from seven EEG foundation models. Experiments on seven public EEG benchmarks demonstrate that EEG-AS substantially narrows the gap between the Single Best Solver (SBS) and the oracle upper bound for each instance. These results highlight the effectiveness of instance-level AS for adaptive deployment of EEG foundation models.

Introduction

EEG-AS reframes foundation-model deployment as instance-level algorithm selection because model performance varies across individual EEG instances. It reconstructs unavailable model behaviors using an inference-available representation and anchor model, improving selection across seven benchmarks.

  • EEG foundation models learn transferable representations and perform strongly across diverse downstream EEG applications.
  • The SBS assumption that one model is uniformly best within a dataset does not hold because performance varies across EEG instances.
  • Different foundation models correctly classify different instances, creating instance-level complementarity beyond dataset-level model competition.
  • EEG-AS reconstructs unexecuted candidate-model behaviors from an EEG representation and one inference-available anchor model.
  • EEG-AS selects models without exhaustively evaluating the portfolio and outperforms SBS on all seven benchmarks.
  • 53.7% SBS accuracy rises to 66.9%, closing 44.0% of the SBS-to-Oracle gap toward the 83.7% Oracle upper bound.

Background

EEG foundation models use self-supervised pretraining and diverse architectures to learn representations for downstream applications. Algorithm selection research has progressed from handcrafted features toward learned representations, including CNNs, GNNs, Transformers, and large-scale embeddings.

  • EEG foundation models use self-supervised pretraining on large-scale unlabeled data to learn transferable downstream representations.
  • EEG foundation models differ in architecture and pretraining strategy, reflecting distinct approaches to spatial, temporal, and semantic EEG modeling.
  • Automated algorithm selection has been applied to Boolean satisfiability, combinatorial optimization, and machine learning.
  • Algorithm selection traditionally uses performance prediction models with handcrafted instance features, while representation learning reduces reliance on domain-specific feature engineering.
  • Learned instance representations have progressed from CNNs and GNNs to Transformer architectures and large-scale embedding models.

Method

EEG-AS formulates foundation-model deployment as per-instance algorithm selection, using inference-available EEG representations and one anchor model to reconstruct other models’ behaviors before selecting among them. Training uses privileged prediction tokens and joint reconstruction-selection objectives, while inference executes only the anchor model.

  • Problem formulation: EEG-AS selects one pretrained EEG foundation model per instance from a portfolio whose models can perform differently across instances.The selector aims to approximate the Oracle assignment while outperforming the Single Best Solver without evaluating the entire portfolio for each instance.
  • Framework: The framework combines an inference-available EEG representation, a conditional token predictor, and a token-based model selector.Representations may use pretrained encoders, handcrafted EEG features, or both; this implementation concatenates a frozen BIOT embedding with handcrafted features.
  • Privileged token representation: During training, prediction tokens from all foundation models provide privileged information, preserving complete instance-specific prediction distributions rather than scalar performance scores.Each token is a predicted class-probability vector, and the full token matrix captures model behavior for the current EEG instance.
  • Conditional token prediction: The conditional predictor reconstructs non-anchor tokens from the EEG representation and observed anchor token, producing sample-specific approximations of unavailable model behaviors.The reconstructed matrix combines the observed anchor token with predicted non-anchor tokens before selection.
  • Model selection and optimization: The selector uses cross-attention over reconstructed token embeddings to score each foundation model and choose the highest-scoring candidate.Training first pretrains token reconstruction, then jointly optimizes reconstruction and model selection using pairwise ranking constraints and a balancing coefficient.
  • Inference: At inference, EEG-AS extracts the EEG representation, executes only the anchor model, reconstructs missing tokens, and selects a model without exhaustively running the portfolio.This inference procedure uses information available before executing the candidate foundation models.

Experimental Setup

EEG-AS is evaluated across seven diverse EEG benchmarks using a portfolio of seven foundation models and standardized selection metrics. The setup compares practical and privileged selection scenarios under a lightweight anchor-based deployment protocol.

  • Datasets and Evaluation Protocol: Evaluation uses seven public EEG benchmarks spanning diverse downstream tasks and recording conditions under official train/validation/test splits.The datasets are ADFTD, BCIC-2a, MIMUL-11, SEED-V, SEED-VII, THINGS-EEG-2, and Workload.
  • Foundation Model Portfolio: The candidate portfolio contains seven frozen EEG foundation models: BIOT, EEGPT, BENDR, CBraMod, CSBrain, LaBraM, and REVE.The models differ in architecture, pretraining strategy, and representation-learning paradigm.
  • Prediction Tokens: Each foundation model produces a class probability vector used as its prediction token for EEG-AS training.These tokens encode the foundation models' prediction behavior for each EEG instance.
  • Inference Protocol: During inference, BIOT supplies the anchor token and EEG representation, while EEG-AS reconstructs the remaining foundation-model behaviors without executing the full portfolio.BIOT is selected as the anchor because it is lightweight and has 3.4M parameters.
  • Evaluation Metrics: Selection effectiveness is measured primarily by accuracy, bounded by the Virtual Best Solver that chooses the best foundation model for each instance.The VBS-SBS gap quantifies the available performance potential for instance-level algorithm selection.
  • Baselines and Reference Methods: Baselines include SBS, VBS or Oracle, MLP and Random Forest selectors using z_x, plus global-token and privileged-teacher reference scenarios.The privileged teacher accesses the complete prediction-token matrix and therefore serves as an upper bound rather than a practical deployment method.

Results

Across seven EEG benchmarks, EEG-AS improves instance-level foundation-model selection over global and conventional selectors, while ablations show complementary representations, anchor conditioning, and token supervision matter. Reconstruction quality is meaningful but does not always translate into optimal selection because headroom, decision alignment, and foundation-model diversity constrain gains.

  • Overall Performance: 66.9% average accuracy exceeds the Single Best Solver’s 53.7% across all seven EEG benchmarks.EEG-AS consistently outperforms SBS on every dataset.
  • Overall Performance: EEG-AS surpasses MLP (64.8%), Random Forest (62.7%), and Global Token Predictor (65.4%) in average accuracy.It achieves the best deployable performance on five of seven datasets and improves BCIC-2a by 4.9 percentage points over the strongest deployable baseline.
  • Input Representation: 66.9% is achieved by combining FM embeddings with handcrafted EEG features, outperforming encoder-only inputs at 64.4%.Removing the FM embedding reduces average accuracy to 66.3%, while removing handcrafted features produces larger task-specific drops, including SEED-V from 43.0% to 39.6% and Workload from 91.8% to 89.6%.
  • Anchor Conditioning: 66.9% with instance-specific anchor conditioning exceeds 66.3% without the anchor token and 65.9% with a global token.The advantage is particularly evident on BCIC-2a and SEED-V.
  • Supervision Target: 66.9% with Token Supervision exceeds 65.4% with Correctness Supervision for dataset-level average accuracy.Token supervision preserves the complete prediction distribution, including confidence and class preference information, rather than binary correctness labels.

Conclusion

Across seven EEG benchmarks, EEG-AS outperforms conventional selection strategies and standalone foundation models, while reconstructed tokens capture meaningful model behaviors. Remaining performance depends on the difficulty of identifying decision-relevant behaviors, and future work will incorporate deployment costs.

  • EEG-AS consistently outperforms conventional selection strategies and standalone FMs across seven EEG benchmarks.
  • Reconstructed tokens capture meaningful foundation-model behaviors, but the remaining performance gap depends on identifying decision-relevant behaviors.
  • Future work will investigate cost-aware foundation-model selection using accuracy, inference time, and computational resources jointly.
Loading 2609.00653v1…