STAMP: Spatial-Temporal Adapter with Multi-Head Pooling

Conference: NeurIPS 2025 arXiv: 2511.10848 Code: https://github.com/autonlab/STAMP Area: EEG Signals / Foundation Model Adaptation Keywords: EEG Classification, TSFM Adapter, Spatial-Temporal Encoding, Multi-Head Pooling, Parameter Efficiency

TL;DR

STAMP introduces a lightweight spatial-temporal adapter with only 750K parameters for Time Series Foundation Models (TSFMs). Through three sets of positional encodings (token/spatial/temporal), cross-gated MLP mixing, and multi-head attention pooling, it enables a frozen TSFM (e.g., MOMENT 385M) to compete with or surpass EEG-specific models with 29M parameters (CBraMod) across 8 EEG datasets, achieving 193% higher Kappa than CBraMod on BCIC-IV-2a.

Background & Motivation

Background: TSFMs (e.g., MOMENT, Chronos), pretrained across multiple domains, have demonstrated strong general-purpose representation capabilities. EEG-specific foundation models (CBraMod, LaBraM) perform well on EEG classification but require large parameter counts (29M/5.8M) and EEG-specific pretraining.

Limitations of Prior Work: TSFMs process univariate time series — EEG is spatiotemporal data with 64 channels × 1000+ time steps, making direct application infeasible for the spatial dimension. Feeding each channel independently into a TSFM discards inter-channel spatial relationships.

Key Challenge: TSFMs possess strong temporal representations but lack spatial understanding; EEG models understand spatial structure but require large-scale EEG pretraining. The core challenge is enabling TSFMs to understand EEG's spatial dimension at minimal cost.

Goal: Design a lightweight adapter that enables general-purpose TSFMs to efficiently process spatiotemporal EEG data.

Key Insight: Freeze the TSFM and train only a 750K-parameter adapter — three sets of positional encodings inject spatial information, cross-gated MLP mixes spatial and temporal features, and multi-head attention pooling aggregates representations.

Core Idea: Triple positional encoding (token + spatial + temporal) + cross-gated MLP spatiotemporal mixing + multi-head pooling = 750K parameters enabling a frozen TSFM to process EEG spatiotemporal data.

Method

Overall Architecture

EEG data (\(S\) channels × \(T\) time steps) → frozen TSFM (e.g., MOMENT-L 385M) encodes to \(S \times T' \times D\) (downsampled 1024→128) → Positional Encoding (\(\tilde{e}_{ij} = e'_{ij} + p_{ij} + s_i + t_j\)) → CC-GMLP (decoupled spatial and temporal gated mixing) → MHAP (multi-head attention pooling to fixed-length vector) → Classification Head

Key Designs

1. Three-Set Positional Encoding (PE-NST):

   • Function: Injects spatial-temporal positional information into TSFM output tokens.
   • Mechanism: Token-wise PE \(p_{ij} \in \mathbb{R}^D\) provides an independent embedding for each (channel, time) position; Spatial PE \(s_i\) encodes channel identity (e.g., C3/C4/Oz); Temporal PE \(t_j\) encodes temporal position. All three are summed.
   • Design Motivation: Ablation studies demonstrate that all three PE components are necessary — token PE alone is insufficient (lacking general spatial/temporal structure), and spatial + temporal PE alone is insufficient (lacking position-specific embeddings).

2. Cross-Channel GMLP (CC-GMLP):

   • Function: Performs feature mixing along spatial and temporal dimensions respectively.
   • Mechanism: Spatial gating \(g_S(Z) = Z_1 \odot (W \cdot Z_2)\) (mixing along the spatial dimension); temporal gating operates analogously along the temporal dimension. Both operate independently to preserve spatiotemporal disentanglement.
   • Design Motivation: Transformers are parameter-heavy for spatiotemporal sequences; GMLP is more efficient, and the CC (cross-channel) variant further reduces parameters (0.74M vs. GMLP 0.79M) while achieving better performance.

3. Multi-Head Attention Pooling (MHAP):

   • Function: Aggregates variable-length spatiotemporal tokens into a fixed-length classification vector.
   • Mechanism: Multiple learnable query vectors aggregate token information via attention weights. Final classification: \(\hat{y} = \text{softmax}(W(\lambda z + (1-\lambda)\hat{e}))\)
   • Design Motivation: More flexible than mean pooling — capable of learning to attend to different temporal segments and spatial regions.

Loss & Training

• Standard cross-entropy classification loss
• MOMENT-Large (385M) is frozen; adapter contains 750K trainable parameters
• Compatible with multiple TSFMs (MOMENT S/B/L, Chronos, TSPulse)

Key Experimental Results

Main Results (8 EEG Datasets)

Dataset	STAMP (750K)	CBraMod (29M)	LaBraM (5.8M)	Result
SHU-MI	0.660 AUC	0.657	0.660	Tie
MentalArith	0.811	0.749	0.772	STAMP wins
BCIC-IV-2a	0.409 Kappa	0.212	0.316	+193%
TUEV	0.662	0.618	0.664	Tie
SEED-V	0.208	0.259	0.239	CBraMod wins
FACED	0.278	0.508	0.470	CBraMod wins

STAMP is competitive or superior on 6/8 datasets, with relative weakness on emotion recognition tasks.

Ablation Study

Variant	Description
PE-NST (all three sets)	Optimal
PE-ST (no token PE)	Performance degrades
CC-GMLP vs. Transformer	CC-GMLP outperforms on all 4 datasets with fewer parameters
MHAP vs. Mean Pooling	MHAP substantially better on BCIC-IV-2a; comparable elsewhere
Different TSFM backbones	MOMENT L > B > S; Chronos slightly better on emotion tasks; TSPulse stronger on event tasks

Key Findings

• A 750K-parameter adapter enables a general-purpose TSFM to match a 29M-parameter EEG-specific model on most tasks — 39× improvement in parameter efficiency.
• Emotion recognition (SEED-V, FACED) is a weakness for TSFMs, likely due to the absence of emotion-relevant features in TSFM pretraining.
• CC-GMLP is more efficient than Transformer for spatiotemporal mixing — suggesting EEG spatiotemporal relationships are relatively simple and do not require full attention.
• The choice of TSFM backbone has limited impact — architectural design matters more than pretraining data.

Highlights & Insights

• Exceptional Parameter Efficiency: The combination of a 750K adapter and a frozen 385M TSFM outperforms a 29M EEG-specific model, validating that general temporal representations + lightweight spatiotemporal adaptation constitute a more efficient paradigm.
• Elegant CC-GMLP Design: Decoupled spatial and temporal gated mixing avoids the curse of dimensionality while preserving spatiotemporal interaction.
• Discovery of TSFM Capability Boundaries: Failure on emotion recognition reveals that TSFM pretraining signals lack affective semantics — EEG-specific pretraining may be necessary for such tasks.

Limitations & Future Work

• Poor performance on emotion recognition tasks — TSFM pretraining data does not contain emotion-related neural signal patterns.
• Evaluation is limited to classification; temporal forecasting and generation tasks — the primary use case of TSFMs — are not assessed, as EEG prediction tasks remain scarce.
• The complete design with five positional encoding components introduces additional hyperparameters requiring tuning for different EEG devices.
• Despite being lightweight, the 750K adapter still requires separate training per dataset — zero-shot cross-dataset generalization is not validated.
• CC-GMLP assumes separability of spatial and temporal dimensions, which may be insufficient for tasks requiring joint spatiotemporal modeling (e.g., rapid-response brain-computer interfaces).
• Dependence on patch-level feature extraction from the TSFM may result in information loss for fine-grained events in raw EEG waveforms.

Related Work & Insights

• vs. CBraMod: A 29M-parameter EEG-specific model; STAMP achieves comparable performance with a 750K adapter and a general-purpose TSFM.
• vs. LaBraM: A 5.8M-parameter model; STAMP matches or exceeds it on most tasks.
• Insight: The paradigm of freezing large pretrained models and training lightweight adapters, already successful in NLP and CV, is validated here for the EEG domain.

Rating

• Novelty: ⭐⭐⭐⭐ First work to adapt TSFMs for EEG spatiotemporal data
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ 8 datasets + multiple TSFMs + comprehensive ablations
• Writing Quality: ⭐⭐⭐⭐ Method description is clear and well-organized
• Value: ⭐⭐⭐⭐ Validates the feasibility of general-purpose TSFM + lightweight adaptation for EEG

Related Papers

• [CVPR 2026] T-Gated Adapter: A Lightweight Temporal Adapter for Vision-Language Medical Segmentation
• [NeurIPS 2025] 3D-RAD: A Comprehensive 3D Radiology Med-VQA Dataset with Multi-Temporal Analysis and Diverse Diagnostic Tasks
• [NeurIPS 2025] Learning Relative Gene Expression Trends from Pathology Images in Spatial Transcriptomics
• [NeurIPS 2025] RAD: Towards Trustworthy Retrieval-Augmented Multi-modal Clinical Diagnosis
• [NeurIPS 2025] A Unified Solution to Video Fusion: From Multi-Frame Learning to Benchmarking