---
title: 'EEG-Inception: End-to-End EEG CNN'
url: https://www.emergentmind.com/topics/eeg-inception
type: topic
---

# EEG-Inception: End-to-End EEG CNN

EEG-Inception is an EEG-adapted Inception-style convolutional neural network architecture for end-to-end EEG classification. In the motor-imagery literature, the name most specifically denotes the 2021 model built on the backbone of InceptionTime and designed to classify raw EEG-based motor imagery without a complex handcrafted preprocessing chain [2101.10932]. At the same time, later papers use “EEG-Inception” more broadly as a reference point for multibranch, multiscale EEG CNN design; this broader usage is evident in comparative and extension papers that treat it as either a baseline architecture or a precursor for later designs emphasizing interpretability, temporal modeling, robustness, and cross-subject transfer [2204.06947][2412.19725].

## 1. Terminological scope and historical placement

The nomenclature around EEG-Inception is not fully uniform across the literature. In one line of work, EEG-Inception is the motor-imagery classifier introduced as “EEG-Inception: An Accurate and Robust End-to-End Neural Network for EEG-based Motor Imagery Classification,” where the network is explicitly presented as a modification of InceptionTime for EEG signals [2101.10932]. In another line of work, later comparative research describes EEG-Inception as a recent end-to-end EEG architecture “utilis[ing] inception modules originally proposed by Szegedy et al. in computer vision,” cites Santamaría-Vázquez et al., and summarizes it as “a CNN architecture that comprises two inception modules and an output module” [2204.06947].

This terminological split matters because later studies sometimes distinguish explicitly between “EEG-Inception (Motor Imagery)” and other inception-based EEG models. EEG-Reptile, for example, states that it used “EEG-Inception (MI)” in its original form, sourced from the Braindecode Python library, and treats it as one of several standard EEG backbones rather than as a new architecture introduced in that paper [2412.19725]. A plausible implication is that “EEG-Inception” functions both as the name of a specific MI classifier and as a broader architectural label for EEG CNNs that rely on parallel, multi-scale convolutional branches.

Historically, the motor-imagery EEG-Inception was proposed to address three persistent MI-BCI problems: limited accuracy, overfitting caused by small EEG datasets, and poor robustness across subjects [2101.10932]. Its central design choice was to import the InceptionTime idea of parallel temporal convolutions with different receptive-field lengths into EEG classification. Later works inherit, critique, or repurpose that design logic rather than treating the original network as closed or definitive [2204.06947][2407.04738].

## 2. Canonical motor-imagery architecture

In its motor-imagery form, EEG-Inception takes raw EEG time series as input, represented as tensors of shape \( \mathrm{RE}[N \times C \times 750] \), where \(N\) is the number of trials, \(C\) the number of channels, and 750 corresponds to 3 seconds at 250 Hz [2101.10932]. For dataset 2b, a trial therefore has 3 channels and 750 time points; for dataset 2a, it has 22 channels and 750 time points.

The network contains six inception modules and two residual modules. More specifically, it comprises one initial inception module, five intermediate inception modules, and residual modules inserted after every three inception modules. The high-level sequence is:

raw EEG input \(\rightarrow\) initial inception module \(\rightarrow\) intermediate inception module \(\rightarrow\) intermediate inception module \(\rightarrow\) residual module \(\rightarrow\) intermediate inception module \(\rightarrow\) intermediate inception module \(\rightarrow\) intermediate inception module \(\rightarrow\) residual module \(\rightarrow\) average pooling + flatten \(\rightarrow\) fully connected layer \(\rightarrow\) classification output [2101.10932].

The inception module has two variants. In the initial inception module, EEG-Inception uses an “inverse bottleneck”: instead of reducing channel depth with a \(1\times1\) bottleneck, it first increases channel depth because the raw EEG input has only 3 or 22 channels. For dataset 2b, the initial bottleneck increases depth from 3 to 12; for dataset 2a, from 22 to 48. The stated motivation is that otherwise the feature maps are “too small for effective feature extraction” [2101.10932].

Temporal feature extraction is performed with parallel 1D convolutions on the time axis. For the binary dataset 2b, the three convolution branches use kernel sizes \([25 \times 1]\), \([75 \times 1]\), and \([125 \times 1]\). For the four-class dataset 2a, five branches are used with kernel sizes \([25 \times 1]\), \([75 \times 1]\), \([125 \times 1]\), \([175 \times 1]\), and \([225 \times 1]\). Because the sampling rate is 250 Hz, these correspond to temporal windows of 0.1 s, 0.3 s, 0.5 s, 0.7 s, and 0.9 s. Each inception module also includes a max-pooling branch with kernel size 25, followed by a \(1\times1\) 1D convolution; the outputs of the convolution branches and the pooled branch are concatenated, batch normalized, and passed through ReLU [2101.10932].

The intermediate inception modules revert to the usual bottleneck role of reducing depth. In the binary setting, the concatenated depth is \(4 \times M\), and the bottleneck reduces it back to \(M\); in the four-class setting, concatenation yields \(6 \times M\). The paper gives \(M=12\) for the binary model and \(M=48\) for the four-class model [2101.10932].

Residual stabilization is inserted after every three inception modules. The residual module is a \(1\times1\) convolutional layer added to the output of the surrounding inception stack, with the residual equation
\[
y = F(x_0,\{W_i\}) + x .
\]
Its stated purpose is to mitigate degradation and vanishing gradients in deeper networks [2101.10932].

After the feature-extraction stack, the classifier head performs average pooling and flattening, followed by a fully connected layer. ReLU and batch normalization are used in the feature extractor. The paper explicitly writes binary cross-entropy as
\[
H_p(q) = -\sum_{i=1}^{N} y_i \log(p(y_i)) + (1-y_i)\log(1-p(y_i)),
\]
while also referring to SoftMax for multiclass behavior [2101.10932].

A notable architectural asymmetry appears in parameter count. The selected binary model has 204,002 trainable parameters and size 10.83 MB, whereas the selected four-class model has 8,917,348 parameters and size 34.10 MB. This large expansion is tied to the 22-channel, five-branch 2a configuration [2101.10932].

## 3. Data representation, preprocessing, and augmentation

The canonical evaluation uses BCI Competition IV datasets 2a and 2b, both sampled at 250 Hz. Dataset 2b contains 9 subjects, 2 MI classes, and 3 EEG channels. Dataset 2a contains 9 subjects, 4 MI classes, and 22 EEG channels [2101.10932].

Although EEG-Inception is framed as an end-to-end raw-EEG model, “raw” does not mean an absence of all signal handling. The operational pipeline extracts 3-second imagery segments, removes trials rejected by the dataset labels, and trains on those segments directly without CSP, FBCSP, ICA, EOG regression, channel selection beyond the benchmark channel sets, or handcrafted time-frequency transforms [2101.10932]. This is narrower than a classical MI preprocessing chain, but it is not a claim of completely untouched sensor data.

The training protocol for the main results is subject-dependent, with each subject trained and tested separately using an approximate 3:1 train:test ratio after removing rejected trials [2101.10932]. The paper also includes a preliminary subject-independent examination, but that is not its primary evaluation mode.

A distinctive contribution of the original work is a task-specific EEG augmentation method intended to diversify high-frequency artifact structure while preserving MI-related content. The procedure applies an 8th-order Butterworth high-pass filter with cutoff 100 Hz to extract a “noise candidate,” subtracts that component from the original signal, and adds the extracted high-frequency component from another randomly chosen trial. The resulting augmented trial is
\[
S_{\text{aug}(i)} = S_o(i) - S_n(i) + S_n(k),
\]
where \(S_o(i)\) is the original signal of trial \(i\), \(S_n(i)\) its extracted noise candidate, and \(k\) a randomly selected trial index [2101.10932].

The augmentation is applied only to the training set. In the reported experiments, it enlarges the training set to 3× its original size for the binary dataset and 6× for the four-class dataset. The paper reports two effects: faster convergence, from roughly 40–60 iterations without augmentation to roughly 10–20 with augmentation, and higher average accuracy. On dataset 2b, average accuracy rises from 85.77% to 88.58%, a gain of 2.81%. On dataset 2a, it rises from 84.82% to 88.39%, a gain of 3.57% [2101.10932].

## 4. Empirical performance and its limits

On dataset 2b, EEG-Inception reports an average subject-dependent accuracy of 88.58%, standard deviation 5.50, and inference time 0.0187 s per sample [2101.10932]. On dataset 2a, it reports an average accuracy of 88.39%, standard deviation 7.06, and inference time 0.0215 s per sample [2101.10932]. The original paper treats the low cross-subject standard deviations as evidence of robustness and highlights difficult subjects in 2b for which the method remains around 80%+, especially subjects 2 and 3.

The strongest empirical case is on dataset 2b. There, EEG-Inception exceeds the compared methods listed in the paper, including HS-CNN at 87.64%, and also achieves the lowest reported standard deviation among those methods [2101.10932]. The dataset 2a result is more qualified. EEG-Inception reaches 88.39%, but the same comparison table reports HS-CNN at 91.57%. Accordingly, the broad claim that EEG-Inception “outperforms all the state-of-the-art methods” is fully supported for 2b but not strictly by the 2a table [2101.10932].

The paper also presents preliminary subject-independent results. On dataset 2b, it reports average accuracy 77.44–77.5%, F1-score 0.55, kappa 0.737, recall 0.628, and AUC 0.8852 with augmentation versus 0.8540 without. On dataset 2a, it reports average accuracy 65.88%, F1-score 0.655, kappa 0.544, and recall 0.657 [2101.10932]. These figures are well below the subject-dependent results. A fair reading is therefore that EEG-Inception shows subject-independent potential, but does not establish strong subject-independent performance in the modern cross-subject sense.

This caution is reinforced by later cross-subject work. EEG-Reptile uses EEG-Inception as one of three backbones for Reptile-based meta-learning on BCI IV 2a and Lee2019 MI, but reports that for “other models” including EEG-Inception, positive average differences over transfer learning were smaller than the 95% confidence interval, and on Lee2019 MI the meta-learned EEG-Inception was comparable to baseline [2412.19725].

## 5. Later comparative uses and architectural descendants

Later work rarely treats EEG-Inception as an immutable endpoint. More often, it is used as a baseline, a template, or a design ancestor.

| Work | Task | Relation to EEG-Inception |
|---|---|---|
| “EEG-Reptile” [2412.19725] | Cross-subject MI meta-learning | Uses EEG-Inception (MI) in original Braindecode form; gains over transfer learning are smaller and statistically inconclusive, with Lee2019 MI reported as comparable to baseline |
| “EEG-ITNet” [2204.06947] | MI classification and explainability | Frames EEG-Inception as a strong but opaque baseline; characterizes it as using inception, no dilation, no visualization, no interpretability, high depth (23 layers), and very high parametric complexity (~15k parameters) |
| “A Contrastive Learning Based Convolutional Neural Network for ERP Brain-Computer Interfaces” [2407.04738] | Cross-subject P300/ERP | Uses an Inception-style encoder with three branches and outperforms EEG-Inception in AUC, \(0.7233 \pm 0.0750\) versus \(0.7013 \pm 0.0681\) |
| “Robust EEG-based Emotion Recognition Using an Inception and Two-sided Perturbation Model” [2404.15373] | Subject-independent emotion recognition | Uses an Inception feature generator on DE tensors inside a two-sided perturbation framework; under PGD-10, reports clean accuracy \(0.93 \pm 0.03\) and robust accuracy \(0.91 \pm 0.04\) |
| “Inception networks, Data Augmentation and Transfer Learning in EEG-based photosensitivity diagnosis” [2502.12021] | PPR/photosensitivity detection | Applies InceptionTime to raw multichannel EEG with transfer learning and augmentation; IT+DA reports ACC \(0.9868\), SENS \(0.8730\), SPEC \(0.9944\) |
| “IncepFormerNet” [2502.13972] | SSVEP classification | Extends Inception-style temporal multiscale CNN design with Transformer attention; reports 87.41% on Benchmark at 1.0 s, while the BETA results contain an abstract/body inconsistency |

Two patterns recur in these later papers. First, the multibranch, multiscale temporal-convolution idea remains intact across tasks. EEG-ITNet explicitly states that inception modules “eliminate the need for a fixed-length kernel” and allow learning filters representing various frequency sub-bands, while IncepFormerNet uses parallel Conv1D kernels of sizes 1, 3, 5, and 8 before Transformer encoding [2204.06947][2502.13972]. Second, later authors often identify limitations in the older baseline formulation: insufficient temporal dependency modeling, lack of visualization, or limited cross-subject generalization [2204.06947][2412.19725].

EEG-ITNet is especially important because it turns EEG-Inception into an explicit contrast case. On BCI Competition IV 2a, it reports EEG-Inception averages of 73.50 within-subject, 63.54 cross-subject, and 75.04 cross-subject with fine-tuning, all below the corresponding EEG-ITNet averages; on OpenBMI, it reports EEG-Inception means of 69.3, 71.15, and 75.11 in the same three scenarios [2204.06947]. That paper repeatedly frames EEG-Inception as strong but relatively opaque and computationally heavier.

## 6. Conceptual significance and recurring misconceptions

A common misconception is that EEG-Inception denotes one universally agreed architecture. The literature summarized here does not support that reading. The term refers unambiguously to a 2021 MI architecture in one context, but later papers also use it as a broader label for inception-based EEG backbones or for older baseline models associated with related EEG tasks [2101.10932][2204.06947][2412.19725].

A second misconception is that EEG-Inception’s “raw EEG” formulation implies the total elimination of data handling. The original MI paper does avoid a classical handcrafted MI pipeline, but it still performs segment extraction, rejected-trial removal, and training-time augmentation via high-pass filtering and cross-trial noise recombination [2101.10932]. The raw-input claim is therefore best understood as the absence of conventional expert-designed feature engineering, not the absence of all preprocessing operations.

A third misconception is that the original work conclusively solved subject-independent EEG classification. The evidence is weaker. The original paper’s subject-independent results are explicitly preliminary and markedly lower than its subject-dependent results, while later cross-subject meta-learning work finds that EEG-Inception’s gains can be small or comparable to baseline transfer learning [2101.10932][2412.19725].

The enduring significance of EEG-Inception lies elsewhere. It established a durable design pattern for EEG deep learning: multi-scale temporal convolutions in parallel branches, EEG-specific bottleneck logic, residual stabilization, and task-aware augmentation. Later architectures often preserve the multiscale Inception front end while adding new components—dilated temporal blocks, contrastive projectors, adversarial perturbation objectives, transfer-learning schemes, or Transformer encoders—rather than discarding the Inception principle itself [2204.06947][2407.04738][2404.15373][2502.13972]. This suggests that EEG-Inception is best understood less as a single frozen network than as a technically influential architectural paradigm within EEG representation learning.

Source: https://www.emergentmind.com/topics/eeg-inception