A controlled benchmark across eight public MI datasets that holds the feature representation fixed and varies only the decoder is reported, arguing this is a positive, well-controlled result that directly answers the reviewer-style question of whether architectural complexity is warranted.
Abstract
A large and growing body of work applies increasingly complex deep architectures to EEG motor-imagery (MI) decoding, yet rarely tests whether that complexity is justified against a strong, simple geometric baseline under identical conditions. We report a controlled benchmark across eight public MI datasets (3–128 channels, 2–3 classes, single- and multi-session) that holds the feature representation fixed and varies only the decoder. The central method — a compact tangent-space pipeline on the SPD manifold with unsupervised test-time recentering, here called Geometry-Aware — is compared against three classical Riemannian baselines (TS+SVM, FgMDM, MDM) and a family of deep models built from our own prior architecture (a bidirectional Mamba mixture-of-experts, BiMamba+MoE, with two reduced ablation variants, and an SPDNet-style network), all consuming the same single-band covariance features. Across N = 88 subject-level observations cross-session and N = 120 within-session, Geometry-Aware achieves the best average rank cross-session and is statistically tied for the best within-session (second by raw rank but indistinguishable from TS+SVM under the critical-difference test). Its cross-session advantage is large and statistically decisive — it beats every competitor after multiple-comparison correction with large effect sizes (Cohen’s d = 1.06–1.50; all pFDR < 1.1 × 10−12) — yet within session its advantage over its recentering-free twin (TS+SVM) is statistically indistinguishable (d = − 0.00, p = 0.54). This cross/within double dissociation points to recentering as the operative mechanism rather than generic capacity. The deep sequence models (the Mamba variants), despite matched features and a fair, fixed training budget, underperform every classical Riemannian method in both protocols by wide margins; the SPDNet baseline fares better — beating MDM — but still never beats the simple tangent-space pipeline on identical features. We argue this is a positive, well-controlled result that directly answers the reviewer-style question of whether architectural complexity is warranted. We state the limitations — fairness of the deep-model comparison, the absence of a direct mechanistic probe, and dataset scope — and outline how each becomes a concrete next step.
Robust EEG motor imagery decoding remains limited by strong inter-individual variability, making it difficult to identify pipelines that generalize across users. We present a large-scale, standardized within-session benchmark of decoding pipelines across three public datasets: Cho2017 (52 subjects), PhysionetMI (109 subjects), and Zhou2016 (4 subjects). Using a common MOABB LeftRightImagery setting, two frequency bands (8-15 Hz and 8-30 Hz), and a broad combination of feature extraction, preprocessing, and classification steps, we analyzed 216,714 raw evaluation rows, which after structured aggregation yielded 44,928, 109,000, and 4,192 subject-level observations respectively. Covariance tangent-space projection (cov-tgsp) and Common Spatial Patterns (CSP) consistently defined the strongest methodological families, though their relative ordering was dataset-dependent. On Cho2017, the best family-level mean accuracy came from cov-tgsp in 8-30 Hz (0.712 +/- 0.140), whereas Zhou2016 favored CSP (0.832 +/- 0.121 in 8-15 Hz). These aggregate rankings concealed substantial subject-level heterogeneity: 42 distinct winning pipelines across 52 Cho2017 subjects, and 93 across 109 PhysionetMI subjects. We then used the benchmark as an empirical performance landscape for building compact portfolios of pipelines of size K. Several construction procedures were compared, including a ranking-based Top-K Mean heuristic and search-based strategies. Results were broadly consistent, with Top-K Mean giving the best trade-off. A single best global pipeline already retained 94.2% of the oracle in Cho2017 and 81.8% in PhysionetMI; at K = 12, oracle retention rose to 96.5% and 90.0%. The landscape is therefore subject-dependent, and this heterogeneity can be exploited through compact portfolios that make personalization more feasible.
Xavier Vasques, Paul Barbaste, Olivier Oullier· 0 citations
Brain-computer interfaces (BCIs) have been widely used in motor rehabilitation, disease diagnosis, and other neural engineering scenarios. However, conventional neural signal decoding algorithms often suffer from limited generalizability and high adaptation costs, motivating recent interest in BCI foundation models. Existing approaches still struggle to jointly achieve general transferability, accurate decoding, and efficient downstream adaptation. We present STEAM, a hierarchical transfer framework that reconciles general-purpose representation learning with paradigm-specific specialization in EEG foundation models. The framework is instantiated as a dual-branch spatio-temporal encoder in which a shared soft mixture-of-experts (SSMoE) module aligns the spatial and temporal branches, allowing complementary representations to exchange information through a compact set of soft slots. Across seven downstream datasets and fourteen evaluation settings, STEAM attains the best average rank among the compared methods at a competitive inference cost measured in FLOPs. Building upon the Stage-I general initialization, the hierarchical pre-training strategy further specializes the model to a target paradigm without retraining from scratch, yielding consistent gains in paradigm-specific decoding accuracy.
Zhu Chen, Dingkun Liu, Yuheng Chen et al.· 0 citations
In computational neuroscience, the statistical nature of primate visual responses has long served as a benchmark for efficient coding. Specifically, previous works demonstrated that in the primate Anterior Inferotemporal (AIT) cortex, population sparseness $(S_{p})$ significantly exceeds single-neuron sparseness $\left(S_{l}\right)$, demonstrating that while individual neurons respond to relatively simple features, the total available feature space is vast. In this work, we establish a comparative experimental framework to bridge the gap between biological neural responses and the internal representations of Vision Transformers (ViTs) and ResNet-50. We analyze layer-wise kurtosis dynamics across ViT-B/16, ViT-L/16, ViT-B/32, and ResNet-50 using 3000 images of ImageNet-1k validation and Caltech-101 datasets. Our results reveal a “Semantic Snap” that is not a fluke and that it is a robust architectural phenomenon of visual models. Notably, while ResNet-50 inverts the biological signature $(S_{l}>S_{p})$, high-capacity transformer-based models like ViT-L/16 achieve a Lehky ratio, mirroring the distributed manifold coding of the AIT cortex. We further identify a significant magnitude gap in representational bandwidth between attention-based and convolutional architectures. These findings, validated by Pareto tail analysis, robust t-statistics, and Lehky ratio calculations, provide a computational link between transformer scaling laws and the Population Manifold Hypothesis in biological vision.
Ankit Sharma, S. Channappayya· International Conference on...· 0 citations
Brain–computer interface (BCI) systems have advanced with deep learning, but they are still limited by designs tied to specific applications, poor scalability, weak portability, the need for user-specific adaptation, and privacy concerns. We present BELT, a modular Bayesian Edge–Cloud architecture based on three principles: (i) Bayesian priors and posteriors to balance generalization and subject-specific learning, (ii) lightweight classifiers suitable for embedded devices, and (iii) task-aware compression to reduce bandwidth and improve privacy in edge–cloud communication. To show feasibility, we implement BELT-lite as an instantiation of BELT, a lightweight version built only from linear time-invariant operations, making it directly compatible with digital signal processing hardware. Using the BCI Competition IV-2a and IV-2b motor imagery datasets (18 subjects total, ten-fold cross-validation), BELT-lite achieved strong posterior performance after subject-specific fine-tuning: mean accuracy of 87.9%±6.8% on Dataset B and 80.6%±8.6% on Dataset A with data augmentation. After adaptation, four subjects from Dataset B and two from Dataset A exceeded 90% accuracy. On ARM Cortex-A7 hardware, BELT-lite achieved a mean latency of 6.75 ms per sample, significantly faster than EEGNet’s 8.36 ms (p < 10-17)—a 21% speed improvement—at the cost of a modest but statistically significant accuracy reduction of approximately 2.7 percentage points compared to EEGNet. Network Tuning Blocks allowed partial parameter freezing: classifier-only fine-tuning incurred a modest 2–5% accuracy drop while substantially reducing training cost. Compression via the task-unaware autoencoder reduced data size by 3.3× while maintaining high accuracy: prior-model performance stayed within ≈1% of the uncompressed baseline (with slight improvements in some configurations), full posterior fine-tuning showed a ≈1% drop, and classifier-only fine-tuning incurred a ≈3% drop—an acceptable trade-off for privacy-preserving edge–cloud communication, where only a compressed latent representation is transmitted instead of raw EEG. Notably, this task-unaware autoencoder (trained solely to reconstruct the input) consistently outperformed autoencoders that also incorporated classification objectives (task-aware or task-only), providing the best accuracy–compression trade-off across all fine-tuning scenarios. These findings show that BELT provides a principled design for modular and scalable BCIs, while BELT-lite demonstrates that the approach supports accurate, efficient, and portable implementations. Together, they point toward BCI systems that are more practical, mass-producible, and privacy-aware, enabling wider use in real-world settings.
Abolfazl Danayi, H. Soltanian-Zadeh· PLoS ONE· 0 citations
We investigate whether a generative model can supply useful synthetic motor-imagery (MI) electroencephalography (EEG) trials that improve the accuracy of independent downstream classifiers. We train a class-conditional variational autoencoder (CVAE) with an integrated latent classifier on the Zhou motor-imagery dataset, using the learned per-class prior as a generator: sampling the prior for a given label and decoding it into a synthetic, label-consistent signal. A constraint on the covariance matrix of the generated data encourages preservation of covariance structure, and the model is trained with a schedule that alternates ordinary VAE training with a decoder-focused phase that sharpens the generative pathway used for augmentation. We measure the effect of adding synthetic trials to the training set under two evaluation protocols -- within-user (pooled 60/20/20 split across subjects) and cross-user (leave-one-subject-out, LOSO) -- across four representative EEG classification pipelines: Common Spatial Patterns with Linear Discriminant Analysis (CSP+LDA), tangent-space features with a Support Vector Machine (TGSP+SVM), Minimum Distance to Riemannian Mean (MDM), and a neural network based on EEGNetv4 (henceforth EEGNet). Results are aggregated across independent augmentation draws, random seeds (within-user), or leave-one-subject-out folds (cross-user), with uncertainty reported as 95\% confidence intervals (Student's $t$-distribution) computed over per-seed/per-fold averages. We find that synthetic EEG from the CVAE is most credible as a source of class-structured, covariance-like data rather than as a substitute for real raw EEG: it can raise the point estimate for MDM, but the broader augmentation claim remains conservative -- observed gains are small and classifier-dependent.
Matei Moldoveanu, Alain Sirois, Claire Ben Ali et al.· 0 citations
The development of Motor Imagery (MI) Brain–Computer Interfaces (BCIs) is systematically constrained by low signal-to-noise ratios (SNRs), signal non-stationarity, and acute data scarcity. While complex Blind Source Separation (BSS) methods optimize signal clarity, their computational overhead introduces propagation delays that challenge real-time constraints. This study addresses this engineering trade-off by introducing a localized architectural framework to evaluate whether a lightweight pipeline operating without BSS (No-BSS) is sufficiently efficient for real-time control when compared against two BSS-equipped pipelines utilizing Independent Component Analysis (ICA) and Empirical Mode Decomposition (EMD). Validated across the BCI Competition IV Dataset 2A and the PhysioNet MI dataset, all three pipelines share an identical processing chain designed to maximize efficiency. To mitigate low SNRs, an Adaptive Laplacian spatial filter isolates neural intent across target sensorimotor electrodes (C3, C4, and Cz). Data scarcity is countered via a Gaussian noise injection data augmentation strategy, while session-to-session variability is addressed during feature extraction using Wavelet Packet Decomposition (WPD) paired with a Fisher Score criterion to dynamically isolate subject-specific time-frequency nodes. Redundant features are subsequently eliminated using a Genetic Algorithm (GA) before classification. Experimental evaluation reveals a distinct performance stratification: while the ICA (92.80%) and EMD (92.69%) pipelines yield the highest average accuracy for the PhysioNet dataset by isolating non-stationary and physiological noise, the No-BSS baseline (90.28%) remains the superior framework for the BCI Dataset 2A. Across all pipelines across both datasets, a stable classification hierarchy emerges wherein the Support Vector Machine (SVM) leads performance due to its maximum-margin decision boundary, followed by k-Nearest Neighbors (kNN), a modified EEGNet, and Decision Trees. The No-BSS baseline achieves classification accuracies highly competitive with its BSS counterparts while entirely bypassing their algorithmic overhead. Given the strict latency constraints of live BCI control loops, these findings establish the optimized No-BSS pipeline as a highly viable alternative for low-latency, real-time implementations.
Nerita Ramsoonder, Rito Clifford Maswanganyi, Philani Khumalo· Big Data and Cognitive Compu...· 0 citations