🤖 AI

STAMP: Spatial-Temporal Adapter with Multi-Head Pooling

This contribution introduces STAMP, a lightweight and flexible Spatial-Temporal Adapter with Multi-Head Pooling that enables general time-series foundation models to achieve performance on clinical EEG tasks comparable to that of specialized EEG foundation models by leveraging univariate embeddings to implicitly capture spatiotemporal features.

Original authors: Brad Shook, Abby Turner, Jieshi Chen, Michał Wiliński, Mononito Goswami, Jonathan Elmer, Artur Dubrawski

Published 2026-05-07

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Brad Shook, Abby Turner, Jieshi Chen, Michał Wiliński, Mononito Goswami, Jonathan Elmer, Artur Dubrawski

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: The Problem of the "Universal Translator"

Imagine you have a brilliant, world-renowned translator (a Time Series Foundation Model or TSFM) who has read millions of books about weather patterns, stock markets, and power grids. This translator is incredibly skilled at recognizing patterns in data that change over time.

Now, you want this translator to read EEG data – the electrical signals from a human brain. The problem is that brain data is chaotic and unique. It has two special dimensions:

Time: How the signal changes from second to second.
Space: How different parts of the brain (the electrodes on the scalp) communicate with each other.

If you simply give the translator the brain data without help, it gets confused. It's like giving a chef who only knows how to make soup a raw steak and expecting them to grill it perfectly without any instructions. The paper found that if you simply feed the translator the brain data and ask for a simple average (like "what is the average taste?"), the result is basically random guessing.

The Solution: STAMP (The Specialized Adapter)

The authors developed a new tool called STAMP (Spatial-Temporal Adapter with Multi-Head Pooling). Imagine STAMP as a specialized translator headset or a custom-made adapter plug.

Instead of retraining the entire brilliant translator from scratch (which is expensive and slow), STAMP builds on top of the existing translator. It takes the translator's raw "understanding" of the data and adds a layer of specific instructions to help it understand the unique layout of the brain.

STAMP does three main things to make the translator work:

1. The "Name Tags" (Positional Encodings)

The translator naturally doesn't know where a signal comes from. Does it come from the left hemisphere? Was it 5 seconds ago?

The Analogy: Imagine a room full of people shouting. The translator hears the noise but doesn't know who is speaking or where they are standing. STAMP puts name tags on everyone. It marks the data with "Left Hemisphere," "Right Hemisphere," "Time 1," "Time 2." This helps the translator understand the geography and timeline of brain activity.

2. The "Cross-Channel Mixer" (Criss-Cross GMLP)

Brain signals are complex. The left side might influence the right side, and a signal from 2 seconds ago might affect what is happening now.

The Analogy: Imagine a group of musicians playing different instruments. A standard translator might listen to each musician individually. STAMP acts like a conductor who hears how the violinist talks to the drummer and how the trumpeter from 10 seconds ago influenced the saxophonist right now. It specifically considers how different brain locations (space) and different moments (time) interact with each other, rather than just looking at them in isolation.

3. The "Smart Highlighter" (Multi-Head Attention Pooling)

After the translator has processed all the data, you need to make a final decision (e.g., "Is this person sleeping?" or "Did they imagine moving their hand?").

The Analogy: Imagine you have a 100-page report. A naive approach would be to read every single word and average the meaning. STAMP is like a smart highlighter. It scans the report and says, "Hey, these three sentences on page 42 are most important for this decision. Ignore the rest." It learns to weight the most critical parts of the brain signal to make the final prediction.

Why This Is a Big Deal

The paper tested this new adapter on 8 different brain datasets, covering tasks like seizure detection, emotion identification, and controlling brain-computer interfaces.

The Result: STAMP performed just as well as or better than the most advanced brain-specific models (EEG Foundation Models) that were built from scratch specifically for brains.
The Efficiency: The "brain-specific" models are like massive supercomputers with millions of parameters (gears and cogs). STAMP is a lightweight, pocket-sized tool. It uses about 750,000 parameters, while the specialized brain models use millions (sometimes 29 million!).
The Flexibility: STAMP works with various "universal translators" (like MOMENT, Chronos, etc.). Even if the translator was never previously trained on brain data, STAMP can help it effectively understand brain signals.

What the Paper Does Not Claim

It is important to stick to what the authors actually said:

They did not claim this is a cure for Alzheimer's or a way to diagnose depression in a hospital tomorrow. They only tested it on existing, public datasets for research purposes.
They did not claim it works perfectly for every task. They noted that it had some difficulty with certain "emotion recognition" tasks (identifying complex feelings like "disgust" vs. "fear"), suggesting that the underlying translator might need more training with emotional data.
They did not say you need to retrain the entire system. The "brain" of the system (the TSFM) remains frozen; only the small STAMP adapter is trained.

Summary

STAMP is a clever, lightweight add-on that takes a general AI expert for time-series data and teaches it how to read brainwaves. This is done by giving the AI a map of the brain (spatial), a clock (temporal), and a way to focus on the most important signals, all without the need to build a massive new computer from scratch.

Technical Summary: STAMP: Spatial-Temporal Adapter with Multi-Head Pooling

Problem Statement

Time Series Foundation Models (TSFMs), pre-trained on diverse domains, have demonstrated strong performance across various modeling tasks. Concurrently, specific Foundation Models for Electroencephalography (EEG) data (EEGFMs) have been developed to capture the unique spatiotemporal characteristics of electrical brain activity. However, a comparative analysis between general TSFMs and EEG-specific Foundation Models for EEG tasks has not yet been conducted. Furthermore, EEGFMs often incorporate complex architectures to model spatial and temporal dependencies, whereas general TSFMs typically treat inputs as univariate time series, inherently rendering them ineffective for EEG tasks where spatial relationships between channels and temporal evolution are critical. A naive application of TSFMs to EEG data (e.g., using simple Mean Pooling) results in performance near random chance.

Methodology

The authors introduce STAMP (Spatial-Temporal Adapter with Multi-Head Pooling), a lightweight, flexible adapter built upon general TSFMs to enable effective EEG modeling. STAMP utilizes univariate embeddings generated by a frozen TSFM and implicitly models the spatiotemporal features of EEG data through three core components:

Positional Encodings (PEs): To address the lack of spatial and temporal context in standard TSFM embeddings, STAMP adds learnable positional encodings. These include token-wise positional encodings ( $p_{ij}$ ), spatial encodings ( $s_i$ ), and temporal encodings ( $t_j$ ). The modified embedding is calculated as $\tilde{e}_{ij} = e'_{ij} + p_{ij} + s_i + t_j$ , where $e'_{ij}$ is the dimensionality-reduced TSFM embedding. Ablation studies show that using all three encoding types yields the best performance.
Criss-Cross Gated MLP (CC-GMLP): To capture relationships across the spatiotemporal grid, STAMP employs a novel CC-GMLP architecture inspired by the Criss-Cross Transformer (CC-TF) and Gated MLPs (GMLP). This module separates the modeling of spatial and temporal relationships. It processes the input grid $\tilde{E}$ through a series of blocks where temporal gating units ( $g_T$ ) and spatial gating units ( $g_S$ ) operate independently along their respective axes before being concatenated. This enables parameter-efficient learning of cross-channel relationships, independent of the specific temporal evolution of the signal.
Multi-Head Attention Pooling (MHAP): To aggregate the mixed token embeddings for final prediction, STAMP utilizes MHAP. This mechanism learns to weight the importance of each token relative to the final task. It projects tokens into low-dimensional spaces, computes attention scores against query vectors, and aggregates the results. A final aggregation step over the queries determines the overall significance of each head and generates a summary vector. This is combined with a residual connection to the mean of the mixed token embeddings before a final linear projection.

The adapter is trained on a frozen TSFM (primarily MOMENT Large, but also tested with MOMENT Small/Base, Chronos, and TSPulse). The total number of trainable parameters in STAMP is approximately 750,000, representing only a fraction of the millions of parameters found in typical EEGFMs.

Main Contributions

First Comparative Analysis: The work provides the first comparative analysis of EEG-specific Foundation Models versus general TSFMs for EEG-specific tasks.
Novel Adapter Architecture: The introduction of STAMP, which successfully adapts univariate TSFM embeddings for multivariate spatiotemporal EEG data without requiring the TSFM to be retrained on EEG data.
Efficiency: STAMP achieves performance comparable to the state-of-the-art in EEGFMs (such as CBraMod and LaBraM) but uses significantly fewer trainable parameters (approx. 0.74M compared to 29M for CBraMod and 5.8M for LaBraM).
Comprehensive Evaluation: The method was evaluated on 8 benchmark EEG datasets covering diverse clinical tasks, including Motor Imagery, mental arithmetic, emotion recognition, and event classification.

Results

Performance vs. Baselines: STAMP achieves performance similar to or surpassing that of EEGFMs (CBraMod, LaBraM) and strictly outperforms non-Foundation EEG models (EEG Conformer, ST-Transformer) across most datasets.
Ablation Studies:
- Positional Encoding: The inclusion of all three types of Positional Encodings (token, spatial, temporal) consistently improved performance, with the exclusion of PEs leading to the worst results.
- Token Mixing: The CC-GMLP architecture outperformed both simple GMLP and Transformer-based mixing strategies (B-TF, CC-TF) across the datasets.
- Token Aggregation: While MHAP and Mean Pooling performed similarly on most datasets, MHAP provided a significant performance boost on the BCIC-IV-2a dataset, indicating its utility in capturing complex token relationships.
TSFM Compatibility: STAMP performed well with embeddings from various TSFMs (MOMENT, Chronos, TSPulse), even with those not pre-trained on EEG data (Chronos, TSPulse), demonstrating the adapter's generalizability.
Limitations: STAMP showed inferior performance compared to other methods on emotion recognition datasets (SEED-V and FACED). The authors suspect that current TSFMs may not extract features relevant to fine-grained emotion recognition, as performance did not significantly improve even when switching to Chronos embeddings.

Significance and Claims

The work claims that STAMP demonstrates that general Time Series Foundation Models can be effectively adapted for EEG tasks with minimal additional trainable parameters. The authors argue that while TSFMs may underperform "out of the box" compared to EEGFMs, relatively few adaptations—specifically the addition of Positional Encodings and a token mixing strategy—are sufficient to achieve competitive results.

The significance lies in the potential reduction of costs and time for developing EEG solutions by leveraging existing, large-scale pre-trained TSFMs rather than training massive, domain-specific Foundation Models from scratch. The authors note that while generating initial embeddings requires computational resources, this can be done in advance and in parallel. They conclude that STAMP offers a flexible, parameter-efficient path for modeling EEG data, although they acknowledge that further investigation is required for tasks where current TSFMs struggle, such as fine-grained emotion recognition.