Toward reliable signals decoding for electroencephalogram: A benchmark study to EEGNeX

1 Introduction

• Most prior studies target accuracy on a specific dataset or task, making representation-learning generalization across neural networks difficult to understand.Such studies commonly use domain-knowledge preprocessing, regularization, transformation techniques, and fine-tuned hyperparameters.Source: §1 · p. 2
• The study asks how different neural-network designs contribute to efficient representation learning from raw EEG data.EEG is commonly used in BCI because it is non-invasive, low-risk, and affordable.Source: §1 · pp. 1–2
• The study evaluates 16 neural-network models across four EEG-based BCI classification tasks to assess their information representation and inform model design.The authors also made the benchmark models and EEGNeX open source as a research testbed.Source: §1 · p. 2
• EEGNeX achieved higher accuracy with statistical significance than typical MOABB pipelines on 11 diverse motor-imagery datasets.The architecture was developed from literature review findings and benchmark results, with spatial, temporal, and cross-channel representation learning as design targets.Source: §1 · p. 2

2 Literature review: Development of NNs in EEG-based classification tasks

• EEGNet combines depthwise and separable convolution for cross-channel spatial feature extraction, with regularization and pooling choices supporting parameter efficiency and cross-section performance.The review also describes earlier ConvNet and CNN approaches and later EEGNet variants that add multi-head, bandpass-filter, or temporal-convolution components.Source: §2 · p. 3
• RNN applications include LSTM and GRU models for emotion and motor-imagery tasks, while CNN-RNN hybrids assign spatial processing to CNNs and temporal extraction to RNNs.The review reports prior comparisons in which attention-based GRUs outperformed CNNs and LSTMs, and CNN-LSTM hybrids trained faster.Source: §2 · p. 3
• The review frames EEG decoding as requiring both temporal and spatial information, motivating investigation of different networks’ representation processes.It notes that convolutional architectures such as ConvNeXt have pursued accuracy and scalability while retaining convolutional simplicity and efficiency.Source: §2 · p. 3

3.1 Candidate methods

• The benchmark investigates fundamental neural-network layer types and derivatives, comparing CNNs, RNNs, and their variations and combinations across 16 candidate models.Candidates include Conv2D, Conv1D, LSTM, and GRU layers; the benchmark also considers convolutional variations such as depthwise separability, causal padding, and dilation.Source: §3 · p. 3, §3.1 · pp. 3–4

3 Preliminary study: Benchmark test

• The benchmark’s reductionist objective is to quantify how fundamental network layers and their derivatives extract implicit EEG information and inform more efficient model design.The identified design pillars are spatial filtering, cross-channel feature learning, temporal feature learning, and parameter space with performance stability.Source: §3 · p. 3, §3.4 · p. 6

3.2 Datasets

• Four EEG datasets were selected to cover within- and cross-subject settings, different tasks and paradigms, trial lengths, data sizes, and channel counts.Evaluation was within-subject to reduce non-stationarity effects, and accuracy was assessed using five-fold cross-validation.Source: §3.2 · p. 4

3.3 Training strategy

• All candidate models used a shared two-stage training scheme, with validation monitoring followed by test-set evaluation to avoid data leakage.Data were split into training, validation, and test sets at 0.75/0.125/0.125 across ten rounds with random split seeds; training stopped after 20 epochs without validation improvement.Source: §3.3 · p. 5

3.4 Performance evaluation and insights

• The benchmark reports accuracy results for all candidate models across four datasets, with model parameter sizes included to represent network complexity and search-space size.The table caption states that average accuracy and deviation are calculated over ten randomized train-validation rounds.Source: §3.4 · p. 5, Table 2 · p. 5
• CNN-based models averaged higher classification accuracy than RNNs, and Conv2D generally outperformed Conv1D; hybrid models improved accuracy over single models, with Conv2D hybrids outperforming Conv1D hybrids.The findings attribute more efficient extraction to cross-channel spatial information and suggest that additional temporal extraction helps after suitable spatial filtering; the study identifies spatial filtering, cross-channel learning, temporal learning, and parameter-space stability as design pillars.Source: §3.4 · pp. 5–6

4 Roadmap: from EEGNet to EEGNeX

• The roadmap tests successive micro-level changes on an EEGNet backbone, tracking steps validated across ten rounds on the described datasets.The four changes reinforce spatial extraction, replace separable convolution, add an inverted bottleneck, and increase receptive fields with dilation and fewer activations.Source: §4 · pp. 6–7

Step 1. Thicken the first part of spatial information extraction

• Duplicating a Conv2D block in EEGNet’s shallow first part is intended to strengthen spatial information extraction while keeping network depth limited.The authors describe this as an intuitive adjustment and note that a more optimal design may exist.Source: Step 1. Thicken the first part of… · p. 7

Step 2. Replace the separable convolution with two 2D convolutions

• The second part replaces separable convolution with an extra Conv2D block to reinforce temporal learning and expand parameter space after channel compression.The added block uses kernel size 16 and 16 output filters; the authors describe separable convolution as redundant after the channel dimension is compressed.Source: Step 1. Thicken the first part of… · p. 7

Step 3. Inverted bottleneck

• The five CNN blocks use an inverted bottleneck with expansion ratio 4 and output filters 8*32*64*32*8.The 32-to-64 transition uses depthwise convolutions with a depth multiplier ratio of two.Source: Step 3. Inverted bottleneck · pp. 7–8

Step 4. Network dilation with fewer activation functions

• Dilations of 1*2 and 1*4 in the third network part target global temporal features without adding sequential NN blocks; each part uses one ELU activation.The authors describe these choices as avoiding further complexity and reducing activation functions between blocks.Source: Step 4. Network dilation with fewer… · p. 8

Closing remarks

• Deeper stacks, fewer normalization layers, reordered blocks, and added CBAM did not improve performance; deeper models also risk convergence difficulty and earlier overfitting on small EEG datasets.The authors attribute the normalization result tentatively to the shallow network design.Source: Closing remarks · p. 8
• EEGNeX-8,32 significantly improves accuracy over EEGNet-8,2 on four datasets; Figure 3 shows it outperforming EEGNet-8,2 across all datasets.The figure summarizes ten-run average performance at each modification step, with horizontal lines indicating deviation.Source: Figure 3 · p. 7, Closing remarks · p. 8

5.1 Model Complexity

• EEGNeX-8,32 has five units, with 8 temporal filters and a kernel size of 32; depthwise convolution learns frequency-spatial features, while dilation expands the receptive field.Each unit comprises convolution, BatchNormalization, and ELU activation layers.Source: Figure 4 · p. 9
• EEGNeX-8,32 occupies a middle ground at ∼12k parameters and 10 layers, versus EEG-Inception’s ∼15k and 23 layers and EEGNet-8,2’s ∼2k and 7 layers.The authors describe this balance as promoting efficiency and reducing overfitting risk while aiming to avoid excessive computational and memory demands.Source: §5.1 · pp. 9–10

5.2 BCI Competition IV Dataset 2a

• 78.81% accuracy: EEGNeX significantly outperformed EEGNet, EEG-Inception, and EEG-TCNet on BCI Competition IV Dataset 2a.The authors also report comparatively consistent performance based on standard deviation values.Source: §5.2 · p. 10, Table 5 · p. 11

5.3 Mother of All BCI Benchmarks

• Across 11 MOABB MI datasets, EEGNeX outperformed TS+optSVM (p < 0.05) and competed favorably with EEGNet (p < 0.001).The within-subject evaluation used 5-fold cross-validation; some collections had small within-session samples that increased NN training difficulty.Source: §5.3 · p. 11, Figure 5 · p. 12

6 Discussion

• Efficient information extraction matters more than simply expanding model capacity, which must be balanced against overfitting risk.The balance depends on input size and complexity, training-example availability, and computational resources.Source: §6 · p. 13
• Confidence intervals reflect uncertainty and variability across subjects; small samples may widen them, while LOSO cross-validation remains a future cross-subject evaluation.The authors suggest larger samples may narrow intervals and improve estimate precision.Source: §6 · p. 13
• The benchmark aims to cover diverse NN architectures, but the BCI-domain contributions of the four ConvNeXt-inspired design pillars remain to be investigated.The authors refer to ConvNeXt’s claimed computer-vision performance advance over Transformers as motivation, not as an established BCI result.Source: §6 · p. 13
• Hybrid models improved accuracy over single architectures, while EEGNeX targets generalizable decoding without explicit feature extraction and domain knowledge may benefit specific datasets.The authors also identify CNN–RNN combinations, Transformer self-attention, and encoder–decoder dimension reduction as potential approaches.Source: §6 · p. 13
• EEGNeX’s decoding performance could be tested on MEG and fNIRS, and broader biological-signal BCI scenarios may benefit from this work.The benchmark methods and EEGNeX improvements rarely incorporate domain-specific feature-engineering knowledge.Source: §6 · p. 13

7 Conclusion

• Benchmarking large-scale NNs in EEG-based BCI classification revealed patterns in model efficacy and representation, including four pillars for efficient EEG classifiers.The supplied conclusion does not enumerate the four pillars.Source: §7 · p. 13
• EEGNeX is presented as a convolution-centric model with consistent effectiveness across EEG classification challenges and adaptability to knowledge-driven, domain-centric feature engineering.Its standout performance is described as not tethered to context-specific design paradigms.Source: §7 · p. 14
• The accompanying benchmarks aim to streamline advanced neural-model deployment in BCI research and support future representation-centric exploration.Source: §7 · p. 14

9.1 Benchmark model structures

• Basic benchmark models use three blocks to match EEGNet’s structure and reflect EEG-task performance without focusing on model capacity.Source: §9.1 · p. 14
• The common macro design uses an input layer, feature-extracting hidden layers, and a 100-neuron dense layer before output, with dropout for robustness.CNN hidden layers also include pooling; layer variations are compared.Source: §9.1 · p. 14
• The hybrid model places an RNN after CNN blocks to test combined spatial-temporal information extraction.The design aims to capture both spatial and temporal information in EEG tasks.Source: §9.1 · p. 14
• The benchmark design aims to enable comparison with prior research and reflect different NN models’ performance on EEG tasks.It follows a literature-review-based macro design and includes a hybrid model to test combined spatial-temporal extraction.Source: §9.1 · p. 14
• Architecture tables cover GRU/LSTM, 1D CNN, 2D CNN, depthwise/separable 2D CNN, and hybrid models.The GRU/LSTM and hybrid tables share general structures but use different RNN layers; the CNN tables describe differing parameter options or architectures.Source: Table 6 · p. 14, Table 7 · p. 15, Table 8 · p. 15

Event-related potential (ERP) – MindBigData

• The single-subject, 10-class ERP dataset used Emotive EPOC recordings at 128 Hz, with about 6,500 trials per digit across 14 channels and 256 timesteps.Each two-second signal depicts an imagined digit from 0 to 9; -1 denotes noise. The dataset is cited as.Source: Event-related potential (ERP) –… · p. 16
• Preprocessing removed noise events, applied a 63 Hz Butterworth low-pass and 50 Hz notch filter, and trimmed the first 32 timesteps.The 32-timestep trim corresponds to 250 ms and was used to avoid sensor power-on noise.Source: Event-related potential (ERP) –… · p. 16

• The dataset includes 54 healthy subjects performing left- and right-hand motor imagery, with 100 trials per session lasting 4 seconds.Source: Motor Imagery (MI) - OpenMBI Data · p. 17
• Classification uses 20 channels selected from the original 62-electrode recording: FC, C, and CP sites.Source: Motor Imagery (MI) - OpenMBI Data · p. 17
• The recordings were notch-filtered at 1–40 Hz and down-sampled from 1000 Hz to 128 Hz.Source: Motor Imagery (MI) - OpenMBI Data · p. 17

• The task is four-class motor-imagery classification of left hand, right hand, feet, and tongue movements in nine subjects.Source: Sensory Motor Rhythm (SMR) -… · p. 17
• The dataset records 22 EEG and 3 EOG channels at 250 Hz, with a 0.5–100 Hz notch filter.Source: Sensory Motor Rhythm (SMR) -… · p. 17
• The analysis uses all EEG-channel signals from the entire trial.Source: Sensory Motor Rhythm (SMR) -… · p. 17

• The two-class task determines whether P300 feedback is correct; detecting ERN feedback helps improve P300-speller performance in the application.Source: Feedback Error-Related Negativity… · p. 17
• The dataset contains recordings from 26 healthy participants using 56 passive Ag/AgCl EEG sensors.Source: Feedback Error-Related Negativity… · p. 17
• The original 600 Hz data were notch-filtered at 1–40 Hz and down-sampled to 128 Hz for analysis.Source: Feedback Error-Related Negativity… · p. 17