🤖 machine learning

Spatiotemporal Convolutions on EEG signal -- A Representation Learning Perspective on Efficient and Explainable EEG Classification with Convolutional Neural Nets

This paper demonstrates that replacing independent 1D convolutions with bi-dimensional spatiotemporal convolutions in shallow CNNs significantly accelerates training and inference for high-dimensional EEG classification tasks while preserving performance, despite revealing distinct internal representational geometries.

Original authors: Laurits Dixen, Stefan Heinrich, Paolo Burelli

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Laurits Dixen, Stefan Heinrich, Paolo Burelli

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 Idea: One Big Filter vs. Two Small Filters

Imagine you are trying to understand a complex song. The song has two main parts: the rhythm (time) and the instruments (space/channels). In the world of Brain-Computer Interfaces (BCI), scientists use computer programs called "Neural Networks" to listen to brain signals (EEG) and figure out what a person is thinking (like imagining moving their left hand vs. their right hand).

For a long time, the standard way to build these programs was to use two separate filters:

Filter A listens only to the rhythm (time).
Filter B listens only to the instruments (space).

The paper asks a simple question: What if we combined these into one giant filter that listens to both rhythm and instruments at the exact same time?

The authors call this a 2D Spatiotemporal Convolution. Think of it like this:

The Old Way (1D): You have a chef who first chops all the vegetables (time), and then a second chef who mixes them with spices (space). They work in a line.
The New Way (2D): You have one chef who chops and mixes everything in a single, smooth motion.

The Experiment: Testing the New Chef

The researchers tested this idea on two different "kitchens" (datasets):

The Small Kitchen: Only 3 brain sensors (channels).
The Big Kitchen: 22 brain sensors (channels).

They compared the old "two-chef" method against the new "one-chef" method using two types of computer models (a simple one and a complex one).

What They Found

1. The Result is the Same (Accuracy)
Whether they used the two separate filters or the single combined filter, the computer got the right answer just as often. It's like saying, "Whether you chop then mix, or mix-while-chopping, the soup tastes the same." The new method didn't make the computer smarter; it just didn't make it dumber.

2. The New Way is Much Faster
This is the big win. In the Big Kitchen (22 sensors), the new "one-chef" method was significantly faster to train.

Why? It's like taking a shortcut. Instead of walking through two doors to get to the kitchen, you walk through one.
The Catch: The computer needed to practice (run through the data) slightly more times to get perfect, but because each practice run was so fast, the total time spent was much less.
Real-world impact: In the world of Brain-Computer Interfaces, speed is everything. If a person is trying to control a cursor with their mind, they need instant feedback. A faster computer means a smoother, more usable experience.

3. The "Secret Sauce" is Different (Representational Geometry)
This is the most fascinating part. Even though the final soup tasted the same (accuracy), the way the chefs organized the ingredients inside the pot was totally different.

The researchers looked at the "internal thoughts" of the computer models.
They found that the Old Method and the New Method organized the information in completely different shapes and patterns.
The Analogy: Imagine two people solving a maze. They both reach the exit at the same time. But Person A took a path that went up, down, and left. Person B went right, up, and left. They got the same result, but their "mental map" of the maze was totally different.

The paper found that this difference in "mental maps" was much more obvious in the Big Kitchen (22 sensors) than in the Small one.

Why Does This Matter?

The paper concludes with two main takeaways for scientists and engineers:

Efficiency: If you are building a brain-reading computer, you should probably use the 2D filter (the one-chef method). It saves time and energy, especially when you have lots of data (many sensors).
Understanding the Brain: If scientists want to use these computers to understand how the brain works (explainability), they need to be careful. Just because two models get the same score doesn't mean they are "thinking" the same way. The architecture (the design of the filter) changes the internal geometry of the data.

Summary

The paper proves that combining time and space filters into a single step is a smart move. It doesn't change the final answer, but it makes the process much faster and creates a different internal structure for how the computer learns. For anyone building brain-computer interfaces, this is a win for speed and a new clue for understanding how these digital brains actually work.

Technical Summary: Spatiotemporal Convolutions on EEG Signals

Problem Statement

Classification of Electroencephalography (EEG) signals using shallow Convolutional Neural Networks (CNNs) is a prevalent approach in fields ranging from medical diagnosis to Brain-Computer Interfaces (BCI). The standard architecture for these models typically employs independent one-dimensional (1D) convolutional layers to process the temporal and spatial dimensions separately, often concatenating them without a non-linear activation layer in between. While numerically equivalent to a single operation under specific conditions (identical weights, no intermediate non-linearity), the impact of replacing these sequential 1D layers with a single bi-dimensional (2D) spatiotemporal convolutional layer on learning dynamics, representational geometry, and efficiency remains uncertain. This paper investigates whether this architectural shift offers tangible benefits in training efficiency and how it alters the internal representations of EEG decoder models.

Methodology

The authors conducted empirical tests comparing 1D and 2D convolutional architectures across two state-of-the-art model families: a shallow CNN based on ShallowFBCSPNet and a hybrid CNN-Transformer model based on the EEG Conformer.

Architectures

1D Baseline Models: These models utilize separate temporal and spatial 1D convolutional layers. The temporal layer convolves over time points, while the spatial layer operates across channels.
2D Variants: The independent variable in this study. These models replace the sequential 1D layers with a single 2D spatiotemporal convolutional layer ( $k_{st} \in \mathbb{R}^{C \times m}$ ), which simultaneously captures temporal and spatial information in one operation. Crucially, no non-linear activation is applied between the temporal and spatial dimensions in either architecture, ensuring the 2D operation is theoretically equivalent in representational power to the two 1D operations.

Datasets and Tasks

Experiments were performed on two motor imagery (MI) datasets from the BNCI Competition IV, processed via the MOABB framework:

BNCI2014-001 (Dataset 1): A high-dimensional task with 22 channels, 4 classes, and 62,208 trials.
BNCI2014-004 (Dataset 2): A low-dimensional task with 3 channels, binary classification, and 32,400 trials.
Both tasks utilized a mixed-subject generalization setting. Preprocessing included bandpass filtering (8–32 Hz) and standard scaling.

Evaluation Metrics

Performance: Classification accuracy and training time (epochs and total duration).
Feature Analysis: Linear regression was used to reconstruct band-power features (5 bands within 8–32 Hz) from model activations to assess spectral feature learning.
Representational Similarity Analysis (RSA): Representational Dissimilarity Matrices (RDMs) were constructed using activations after the pooling layer to compare the geometric structure of representations across different model types and folds.

Key Results

1. Training Efficiency

The most significant finding was a substantial reduction in training time for 2D models, particularly in high-dimensional settings.

Dataset 1 (22 channels): 2D models trained significantly faster than their 1D counterparts. While 2D models required more epochs to reach convergence, the time-per-epoch was drastically lower, resulting in a net reduction in total training time.
Dataset 2 (3 channels): 2D models were similarly or slightly faster, though the effect was less pronounced than in the high-dimensional task.
Inference: The authors note that 2D convolutions also yield faster inference times, a critical factor for online BCI applications.

2. Classification Performance

Accuracy remained statistically unchanged between 1D and 2D variants for both model families (Shallow CNN and Conformer) across both datasets. The models demonstrated comparable reliability and variance across cross-validation folds.

3. Feature Learning and Reconstruction

Analysis of band-power reconstruction revealed that both 1D and 2D models learned essentially identical spectral features.

There were no systematic differences in the ability to reconstruct alpha and beta band powers from kernel activations between the two architectures.
Both architectures successfully identified channel-specific spectral importance (e.g., C3 vs. Cz differences in specific frequency bands), indicating that the shift to 2D did not alter what spectral features were learned.

4. Representational Geometry

Despite identical performance and feature reconstruction, RSA revealed a fundamental difference in the internal structure of the models:

High-Dimensional Divergence: In Dataset 1 (22 channels), the representational dissimilarity between 1D and 2D models was significant. The 2D models formed a distinctly different representational geometry compared to the 1D models.
Low-Dimensional Convergence: In Dataset 2 (3 channels), the dissimilarity between 1D and 2D models was lower, likely because the limited channel information was spread across a larger number of kernels, diluting the observable effects of the architectural bias.
Stability: Models were generally more similar to different folds of their own architecture than to different architectures, confirming that the architectural inductive bias (1D vs. 2D) strongly dictates the representational space.

Significance and Claims

The paper claims that while 2D spatiotemporal convolutions are numerically equivalent to sequential 1D convolutions in terms of raw feature extraction capability, they are not equivalent in terms of learning dynamics and representational geometry.

Efficiency: The primary practical contribution is the demonstration that 2D convolutions offer a computationally efficient alternative for EEG decoding, significantly reducing training and inference times, especially for high-channel-count datasets. This efficiency allows for more extensive hyperparameter tuning and model exploration without additional computational cost.
Architectural Inductive Bias: The study highlights that the choice of encoding layer (1D vs. 2D) imposes a specific inductive bias that reshapes the internal representation space. This suggests that researchers utilizing explainable AI (XAI) techniques must account for architectural differences, as the "how" of representation differs even when the "what" (performance and spectral features) remains the same.
Recommendation: The authors propose adopting 2D spatiotemporal convolutional layers as a standard for EEG decoder models to achieve faster training and inference while maintaining state-of-the-art performance, particularly in complex, multivariate signal processing tasks.