Deep Convolutional Architectures for EEG Classification: A Comparative Study with Temporal Augmentation and Confidence-Based Voting

Aryan Patodiya; Hubert Cecotti

cs.AIFeb 25, 2026

Deep Convolutional Architectures for EEG Classification: A Comparative Study with Temporal Augmentation and Confidence-Based Voting

Authors: Aryan Patodiya, Hubert Cecotti

Organizations: Department of Computer Science California State University, Fresno Fresno, CA, USA

Abstract

Electroencephalography (EEG) classification plays a key role in brain-computer interface (BCI) systems, yet it remains challenging due to the low signal-to-noise ratio, temporal variability of neural responses, and limited data availability. In this paper, we present a comparative study of deep learning architectures for classifying event-related potentials (ERPs) in EEG signals. The preprocessing pipeline includes bandpass filtering, spatial filtering, and normalization. We design and compare three main pipelines: a 2D convolutional neural network (CNN) using Common Spatial Pattern (CSP), a second 2D CNN trained directly on raw data for a fair comparison, and a 3D CNN that jointly models spatiotemporal representations. To address ERP latency variations, we introduce a temporal shift augmentation strategy during training. At inference time, we employ a confidence-based test-time voting mechanism to improve prediction stability across shifted trials. An experimental evaluation on a stratified five-fold cross-validation protocol demonstrates that while CSP provides a benefit to the 2D architecture, the proposed 3D CNN significantly outperforms both 2D variants in terms of AUC and balanced accuracy. These findings highlight the effectiveness of temporal-aware architectures and augmentation strategies for robust EEG signal classification.

Figures & tables

Panel of Figure 1 — Figure 1: Images of targets (top) and non-target (bottom).

F7	F3	Fz	F4	F8
FC5	FC1	-	FC2	FC6
T7	C3	Cz	C4	T8
CP5	CP1	-	CP2	CP6
P7	P3	Pz	P4	P8
-	PO3	-	PO4	-
-	O1	Oz	O2	-

Table 1: 2D sensor layout.

Layer	Filter/Units	Output Shape	Remarks
Input	–	$C\times T\times 1$	CSP/Non-CSP input
Conv2D	$1\times 10$ , $F=8$	$C\times T^{\prime}\times 8$	Temporal
ReLU + BN	–	same	–
Conv2D	$C\times 1$ , $F=16$	$1\times T^{\prime}\times 16$	Spatial
ReLU + BN	–	same	–
Conv2D	$1\times 10$ , $F=32$	$1\times T^{\prime\prime}\times 32$	Full 2D

Table 2: Baseline 2D CNN (ReLU + Cross-Entropy)

Layer	Filter/Units	Output Shape	Remarks
Input	–	$C\times T\times 1$	CSP/Non-CSP input
Conv2D	$1\times 10$ , $F=8$	$C\times T^{\prime}\times 8$	Temporal
GELU + BN	–	same	–
Conv2D	$C\times 1$ , $F=16$	$1\times T^{\prime}\times 16$	Spatial
GELU + BN	–	same	–
Conv2D	$1\times 10$ , $F=32$	$1\times T^{\prime\prime}\times 32$	Full 2D

Table 3: Improved 2D CNN (GELU + Focal Loss)

Layer	Filter/Units	Output Shape	Remarks
Input	–	$C\times T\times 1$	CSP/Non-CSP input
Conv2D	$1\times 10$ , $F=8$	$C\times T^{\prime}\times 8$	Temporal
GELU + BN	–	same	–
Conv2D	$C\times 1$ , $F=16$	$1\times T^{\prime}\times 16$	Spatial
GELU + BN	–	same	–
Conv2D	$1\times 10$ , $F=32$	$1\times T^{\prime\prime}\times 32$	Full 2D

Table 4: 2D CNN with Global Average Pooling

Layer	Details	Output Shape	Remarks
Input	–	$C\times T\times 1$	CSP/Non-CSP input matrix
Conv2D	$1\times 10$ , $F=8$	$C\times T^{\prime}\times 8$	Temporal filtering
GELU + BN	–	same	Nonlinear activation and normalization
Conv2D	$C\times 1$ , $F=16$	$1\times T^{\prime}\times 16$	Spatial filtering
GELU + BN	–	same	–
Conv2D	$1\times 10$ , $F=32$	$1\times T^{\prime\prime}\times 32$	Full feature integration

Table 5: 2D CNN with Squeeze-and-Excitation (SE) Block

Layer	Filter/Units	Output Shape	Remarks
Input	–	$C\times T\times 1$	CSP/Non-CSP input
Conv2D	$1\times 10$ , $F=8$	$C\times T^{\prime}\times 8$	Temporal
Swish + BN	–	same	–
Conv2D	$C\times 1$ , $F=16$	$1\times T^{\prime}\times 16$	Spatial
Swish + BN	–	same	–
Conv2D	$1\times 10$ , $F=32$	$1\times T^{\prime\prime}\times 32$	Full 2D

Table 6: 2D CNN with Swish Activation

Layer	Filter/Units	Output Shape	Remarks
Same as GELU model	–	same	Only loss function changed
Loss	WCE	–	Uses inverse class frequency

Table 7: 2D CNN with Weighted Cross-Entropy Loss

Layer	Filter Size / Units	Output Shape	Remarks
Input	–	$32\times 64\times 5\times 1$	Raw EEG tensor with shift augmentation
Conv3D	$1\times 7\times 3$ , $F=16$	same	Temporal filtering across $T$ and $D$
GELU + BN	–	same	Non-linearity + normalization
Conv3D	$32\times 1\times 1$ , $F=32$	$1\times 64\times 5\times 32$	Spatial filter across channels
GELU + BN	–	same	–
Dropout	$p=0.3$	same	Regularization

Table 8: Baseline 3D CNN Architecture

Layer	Filter Size / Units	Output Shape	Remarks
Input	–	$32\times 64\times 5\times 1$	Augmented EEG input
Conv3D	$1\times 7\times 3$ , $F=16$	same	Temporal convolution
GELU + BN	–	same	–
Conv3D	$32\times 1\times 1$ , $F=32$	$1\times 64\times 5\times 32$	Spatial filtering
GELU + BN	–	same	–
Dropout	$p=0.3$	same	Regularization

Table 9: Deep 3D CNN with Additional Convolution

Layer	Filter Size / Units	Output Shape	Remarks
Input	–	$32\times 64\times 5\times 1$	Augmented EEG input
Conv3D	$1\times 7\times 3$ , $F=16$	same	Temporal filtering
GELU + BN	–	same	–
Conv3D	$32\times 1\times 1$ , $F=32$	$1\times 64\times 5\times 32$	Spatial filtering
GELU + BN	–	same	–
Dropout	$p=0.3$	same	–

Table 10: Best Performing 3D CNN (with GAP)

Participant	Baseline	WCE	Swish	GELU	GAP	SE Block (Best)
s1	0.529	0.612	0.709	0.845	0.855	0.868
s2	0.516	0.595	0.691	0.829	0.839	0.861
s3	0.523	0.604	0.700	0.836	0.846	0.835
s4	0.533	0.617	0.714	0.851	0.861	0.837
s5	0.519	0.598	0.695	0.832	0.842	0.874
s6	0.527	0.609	0.706	0.843	0.853	0.849

Table 11: Mean AUC for All 2D CNN Architectures (Without CSP Preprocessing)

Participant	Baseline	WCE	Swish	GELU	GAP	SE Block
s1	0.531	0.615	0.713	0.868	0.870	0.873
s2	0.518	0.598	0.695	0.853	0.855	0.858
s3	0.525	0.607	0.704	0.850	0.860	0.862
s4	0.535	0.619	0.718	0.869	0.872	0.875
s5	0.521	0.601	0.699	0.857	0.859	0.861
s6	0.529	0.612	0.710	0.861	0.863	0.866

Table 12: Mean AUC for All 2D CNN Architectures (With CSP Preprocessing)

Participant	Baseline	Deep Stack	Best Model (GAP)
s1	0.761	0.935	0.995
s2	0.730	0.918	0.991
s3	0.775	0.941	0.999
s4	0.748	0.929	0.993
s5	0.781	0.945	0.994
s6	0.735	0.921	0.992

Table 13: Comparative Per-Participant Mean AUC for All 3D CNN Architectures

EEG Decoding

Electroencephalography (EEG) decoding maps brain-signal patterns to mental states, stimuli, speech, or intended actions.

See all 208 papers

Spatiotemporal Convolutions on EEG signal -- A Representation Learning Perspective on Efficient and Explainable EEG Classification with Convolutional Neural Nets
May 5, 2026Laurits Dixen, Stefan Heinrich, Paolo BurelliEfficient Neural Network Inference MI Classification
IT University of Copenhagen Denmark
The Deep-Match Framework for Event-Related Potential Detection in EEG
Mar 11, 2026Marek Zylinski, Bartosz Tomasz Smigielski, Gerard CybulskiElectroencephalography EEG Decoding
Institute of Metrology and Biomedical Engineering, Faculty of Mechatronics, Warsaw University of Technology, Warsaw, Poland
Transformer Based Model for Spatiotemporal Feature Learning in EEG Emotion Recognition
Jun 9, 2026Xinglong Cui, Dian GuElectroencephalography Temporal Attention
Beijing Neurodeep Technology Co., Ltd, Beijing, 100176, China · University of Pennsylvania, Seattle, 98121, USA

May 5, 2026J/K move · Enter open · S save

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

Laurits Dixen, Stefan Heinrich, Paolo Burelli

IT University of Copenhagen Denmark

Efficient Neural Network Inference MI Classification EEG Decoding Convolutional Neural Networks

Open paper AskPDF arXiv

Classification of EEG signals using shallow Convolutional Neural Networks (CNNs) is a prevalent and successful approach across a variety of fields. Most of these models use independent one-dimensional (1D) convolutional layers along the spatial and temporal dimensions, which are concatenated without a non-linear activation layer between. In this paper, we investigate an alternative encoding that operates a bi-dimensional (2D) spatiotemporal convolution. While 2D convolutions are numerically identical to two concatenated 1D convolutions along the two dimensions, the impact on learning is still uncertain. We test 1D and 2D CNNs and a CNN+transformer hybrid model in a low-dimensional (3-channel) and a high-dimensional (22-channel) BCI motor imagery classification task. We observe that 2D convolutions significantly reduce training time in high-dimensional tasks while maintaining performance. We investigate the root of this improvement and find no difference in spectral feature importance. However, a clear pattern emerges in representational similarity across models: 1D and 2D models yield vastly different representational geometries. Overall, we suggest an improved model with a 2D convolutional layer for faster training and inference. We also highlight the importance of architecturally-driven encoding when processing complex multivariate signals, as reflected in internal representations rather than purely in performance metrics.

F7	F3	Fz	F4	F8
FC5	FC1	-	FC2	FC6
T7	C3	Cz	C4	T8
CP5	CP1	-	CP2	CP6
P7	P3	Pz	P4	P8
-	PO3	-	PO4	-
-	O1	Oz	O2	-

F7	F3	Fz	F4	F8
FC5	FC1	-	FC2	FC6
T7	C3	Cz	C4	T8
CP5	CP1	-	CP2	CP6
P7	P3	Pz	P4	P8
-	PO3	-	PO4	-
-	O1	Oz	O2	-

Abstract

EEG Decoding

Explore similar work

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

The Deep-Match Framework for Event-Related Potential Detection in EEG

Transformer Based Model for Spatiotemporal Feature Learning in EEG Emotion Recognition

F7	F3	Fz	F4	F8
FC5	FC1	-	FC2	FC6
T7	C3	Cz	C4	T8
CP5	CP1	-	CP2	CP6
P7	P3	Pz	P4	P8
-	PO3	-	PO4	-
-	O1	Oz	O2	-