Inception1D: Multi-Scale Temporal Modeling

Updated 4 July 2026

Inception1D is a family of one-dimensional convolutional architectures that extract temporal features at multiple scales using parallel branches.
It adapts the classic Inception module by employing diverse kernel sizes to simultaneously capture short-, medium-, and longer-range temporal patterns.
Applications include infrasound classification, gait signal analysis, EEG interpretation, and raw audio processing, demonstrating improved accuracy and efficiency over traditional methods.

Searching arXiv for recent and foundational papers on Inception1D-related architectures. Inception1D denotes a class of one-dimensional convolutional architectures that adapt the Inception design pattern to temporal or sequential data. In the cited literature, the term is used in two closely related senses: as a direct reference to the InceptionTime family for time-series classification, including a tsai-based implementation customized for infrasound classification (Klenkert et al., 2024), and as an explicit three-branch 1D convolutional module embedded in hybrid architectures such as InceptoFormer for Parkinson’s disease severity evaluation from gait signals (Naimi et al., 6 Aug 2025). Across these usages, the defining property is multi-branch temporal convolution with different kernel sizes applied in parallel, so that short-, medium-, and longer-range temporal patterns can be extracted simultaneously. Related work in raw audio, EEG, and even image modeling with oriented 1D kernels situates Inception1D within a broader program of replacing handcrafted transforms or conventional 2D convolutional blocks with learned multi-scale 1D feature extractors (Ebrahimpour et al., 2020, Zhang et al., 2021, Kirchmeyer et al., 2023).

1. Terminological scope and conceptual definition

Inception1D is a 1D adaptation of the Inception module originally developed for image processing, reformulated so that convolutions operate along time rather than over two spatial dimensions (Naimi et al., 6 Aug 2025). In one line of work, the designation refers to an InceptionTime-based network applied directly to univariate infrasound time series via the tsai Python library, with only the final output layer customized for the target label space (Klenkert et al., 2024). In another, it refers to a specifically described three-stream convolutional block for vertical ground reaction force (VGRF) gait signals, repeated in cascade and coupled to Transformer encoders (Naimi et al., 6 Aug 2025).

The shared architectural principle is parallel 1D convolutions with heterogeneous kernel sizes. This arrangement is intended to capture temporal patterns at several receptive-field scales without committing the model to a single kernel length. In the InceptoFormer formulation, the authors state that the module “consists of three convolutional streams in parallel in which the previous layer is common to all of them” (Naimi et al., 6 Aug 2025). In the raw-audio “inception nucleus,” the same general idea appears as several 1D temporal convolutions with different kernel sizes whose outputs are concatenated along the channel dimension (Ebrahimpour et al., 2020). EEG-Inception extends the principle to multichannel electroencephalography and combines it with residual connections, again using multiple temporal kernels in parallel (Zhang et al., 2021).

A plausible implication is that “Inception1D” is best understood not as a single canonical architecture, but as a design family defined by multi-scale 1D convolutional branching. The specific choices of kernel sizes, branch count, residual topology, and downstream classifier vary across domains and implementations.

2. Core architectural pattern

At the module level, Inception1D applies several 1D convolutions in parallel to the same input sequence and merges the resulting feature maps by channel-wise concatenation. In the gait-based formulation, the three branches are explicitly defined as Conv1D layers with 32 filters and kernel sizes $K=1$ , $K=3$ , and $K=5$ , respectively; the branch outputs are concatenated, followed by Batch Normalization and an activation function, with SeLU used in the convolutional and fully connected layers (Naimi et al., 6 Aug 2025). The concatenated output is written as

$y_{\text{inception}} = [y_1, y_3, y_5].$

The same source gives the branch-level convolution expression

$y_k(t)=f\left(\sum_{i=0}^{k-1} W_k(i) \cdot x + b_k\right),$

where $x$ is the input segment, $W_k(i)$ the kernel weight at position $i$ , $b_k$ the bias, and $f(\cdot)$ the SeLU activation (Naimi et al., 6 Aug 2025). This formulation emphasizes that the distinction among branches lies in kernel extent rather than in any change of input modality.

The tsai-based InceptionTime implementation used for infrasound classification relies on the standard InceptionTime design rather than re-deriving its internal structure (Klenkert et al., 2024). In the summarized description, a typical InceptionTime module includes a bottleneck 1D convolution of kernel size 1, several parallel 1D convolutions with different kernel sizes, a parallel max-pooling branch, concatenation of branch outputs, and batch normalization plus activation (Klenkert et al., 2024). Residual shortcuts are then used across blocks: $K=3$ 0

This distinguishes two common Inception1D variants. One is the simpler three-convolution-branch module with no explicit pooling branch or bottleneck, as in InceptoFormer (Naimi et al., 6 Aug 2025). The other follows canonical InceptionTime practice and includes bottlenecking, pooling, and residual structure, as implied by tsai defaults and the original InceptionTime reference cited in the infrasound study (Klenkert et al., 2024). A common misconception is that all Inception1D models must include a pooling branch or 1×1 bottleneck; the cited literature shows that some do, while others deliberately omit them.

3. Multi-scale temporal modeling

The principal motivation for Inception1D is multi-scale temporal feature extraction. Different kernel sizes expose the network to patterns with different temporal extents in parallel. In the gait setting, $K=3$ 1 is described as capturing very local, instantaneous variations, $K=3$ 2 as capturing short-range temporal patterns, and $K=3$ 3 as capturing slightly longer local patterns (Naimi et al., 6 Aug 2025). The authors explicitly state that smaller kernels focus on localized patterns whereas larger kernels extract regional patterns, so that “both local and regional information were extracted on the same features map” (Naimi et al., 6 Aug 2025).

In EEG-Inception, kernel sizes are chosen as multiples of 25 samples at a sampling rate of 250 Hz, corresponding to 0.1 s temporal increments over motor-imagery epochs (Zhang et al., 2021). The binary variant uses parallel branches with kernel sizes 25, 75, and 125, plus a max-pooling branch projected with a 1×1 convolution; the four-class variant extends this to 25, 75, 125, 175, and 225 (Zhang et al., 2021). This is an explicit example of tying receptive-field design to domain-specific time scales.

In raw-audio recognition, the inception nucleus is motivated partly by the difficulty of choosing a single appropriate kernel size for waveform modeling. The architecture therefore uses several temporal kernel sizes in parallel and lets end-to-end training determine which branches become useful (Ebrahimpour et al., 2020). The paper describes this as optimizing filter size “on the fly,” but also clarifies that there is no explicit dynamic routing or gating mechanism; the optimization is implicit through backpropagation (Ebrahimpour et al., 2020).

The comparative infrasound study offers a further interpretive point. There, a direct InceptionTime approach on raw 94-sample sequences outperforms a continuous-wavelet-transform-plus-ResNet pipeline, suggesting that learned multi-scale temporal convolution can substitute for handcrafted time-frequency preprocessing in at least some short-sequence classification regimes (Klenkert et al., 2024). This suggests that Inception1D can function as an end-to-end analogue of manual multi-resolution analysis, though the exact equivalence should not be overstated.

4. Representative instantiations across domains

The literature supplied spans several distinct application domains, each using the Inception1D principle differently. The following table organizes the variants that are explicitly described.

Paper	Inception1D formulation	Reported role
(Klenkert et al., 2024)	tsai InceptionTime with custom output layer from 128 features to 8 outputs	Direct classification of 1D infrasound time series
(Naimi et al., 6 Aug 2025)	Three parallel Conv1D branches with $K=3$ 4, 32 filters each, repeated 3 times in cascade	Per-sensor feature extractor in InceptoFormer
(Ebrahimpour et al., 2020)	“Inception nucleus” with parallel 1D temporal convolutions of different kernel sizes	End-to-end raw-waveform feature learning
(Zhang et al., 2021)	Inception-style 1D EEG blocks with multi-scale kernels, pooling branch, and residual modules	End-to-end EEG motor imagery classification

In the infrasound study, the direct approach uses InceptionTime as supported natively by tsai, with the original 128-feature output replaced by a custom layer producing 8 outputs for eight simulated infrasound classes (Klenkert et al., 2024). The internal Inception blocks and residual layout are taken from the library default rather than modified.

In InceptoFormer, 18 foot-mounted sensors are processed independently by 18 parallel Inception1D stacks, each operating on 100-step VGRF segments (Naimi et al., 6 Aug 2025). Each sensor stream uses three cascaded Inception1D modules before a temporal Transformer encoder, and the sensor-wise outputs are later integrated by a spatial Transformer (Naimi et al., 6 Aug 2025).

EEG-Inception adapts InceptionTime to multivariate EEG. It uses one initial Inception module with an inverted bottleneck that expands channel depth, five intermediate Inception modules, and two residual modules inserted after the third and sixth Inception blocks (Zhang et al., 2021). For the binary motor-imagery task, each branch produces 12 channels and the concatenated depth is 48; for the four-class task, each branch produces 48 channels and the concatenated depth is 288 (Zhang et al., 2021).

The raw-audio inception nucleus is conceptually simpler than canonical 2D Inception. It uses multiple temporal Conv1D branches with different kernel sizes and concatenation, but no 1×1 bottlenecks or pooling branches in the core 1D module (Ebrahimpour et al., 2020). The resulting 1D features are subsequently reshaped and processed by a VGG-like 2D CNN backbone (Ebrahimpour et al., 2020).

5. Training configurations, data interfaces, and implementation conventions

Inception1D models are generally trained end to end on raw or minimally processed sequential inputs. The infrasound classification setup uses 2400 simulated signals, each of length 94, split into 1920 training and 480 validation examples with a fixed random seed (Klenkert et al., 2024). Data are loaded through TSDataLoaders, batched in groups of 64, and each batch is standardized according to its own mean and standard deviation (Klenkert et al., 2024). The best direct InceptionTime configuration uses a learning-rate slice $K=3$ 5, trained for 50 epochs but with 33 epochs found to work best based on validation accuracy (Klenkert et al., 2024).

InceptoFormer takes segmented gait windows of 100 time steps with 50% overlap from 18 VGRF channels (Naimi et al., 6 Aug 2025). Class imbalance is addressed by SMOTE-based oversampling, with synthetic samples generated according to

$K=3$ 6

(Naimi et al., 6 Aug 2025). Training uses Nadam with learning rate $K=3$ 7, batch size 64, dropout 0.2, early stopping based on validation loss, and categorical cross-entropy (Naimi et al., 6 Aug 2025).

EEG-Inception operates on 3-second windows sampled at 250 Hz, yielding 750 time steps per trial (Zhang et al., 2021). The binary dataset variant uses 3 channels and the four-class variant 22 channels (Zhang et al., 2021). Training uses Adam with learning rate $K=3$ 8, batch size 32, and 100 iterations in the reported configuration (Zhang et al., 2021). The architecture is explicitly designed to support subject-dependent and preliminary subject-independent evaluation.

A recurring implementation convention is shape preservation through padding so that branch outputs can be concatenated along channels. Although some papers do not state padding explicitly, their reported concatenated output dimensions imply a length-preserving configuration in practice (Naimi et al., 6 Aug 2025, Zhang et al., 2021). Another common convention is to end with global average pooling and a simple linear or softmax classification head rather than large dense blocks (Klenkert et al., 2024, Ebrahimpour et al., 2020, Zhang et al., 2021).

6. Empirical performance and comparative significance

The cited studies report strong results for Inception1D-style models, though direct comparison across domains is not meaningful because the tasks and datasets differ substantially.

For simulated infrasound classification into eight classes, the direct InceptionTime approach reaches a final validation accuracy of 95.2%, while the wavelet-plus-ResNet50 approach reaches 90.2% under the best reported settings (Klenkert et al., 2024). The same source states that both approaches achieve above 90% accuracy, but that the direct method “significantly outperformed the wavelet approach on training speed” because it processes only 94 points per signal rather than $K=3$ 9 values after time-frequency transformation (Klenkert et al., 2024).

For Parkinson’s disease severity staging, InceptoFormer achieves 96.6% accuracy, 93.97% precision, 94.15% recall, and 93.6% F1-score (Naimi et al., 6 Aug 2025). Its ablation study is particularly informative for the role of Inception1D. An Inception1D-only model yields 78.34% accuracy, whereas a Transformers-only model without Inception1D yields 89.17%; the full hybrid model reaches 96.6% (Naimi et al., 6 Aug 2025). This indicates that the convolutional front end is neither sufficient by itself nor dispensable in the presence of attention.

For environmental sound recognition on UrbanSound8K, the Inception Nucleus model reaches 88.4% accuracy with 289K parameters (Ebrahimpour et al., 2020). The paper states that this bests EnvNet-v2 at 78.0% by 10.4 percentage points (Ebrahimpour et al., 2020). Variants that move large kernels to later layers or add batch normalization perform worse, with Inception Nucleus-FI at 75.3%, Inception Nucleus-FA at 70.9%, and Inception Nucleus-BN at 83.2% (Ebrahimpour et al., 2020).

For EEG motor imagery, EEG-Inception reports average accuracy of 88.58% with standard deviation 5.50 on BCI Competition IV 2b, and average accuracy of 88.39% with standard deviation 7.06 on 2a (Zhang et al., 2021). The paper also reports inference times of approximately 0.0187 s per sample for 2b and 0.0215 s per sample for 2a, supporting the claim that the method is suitable for real-time processing (Zhang et al., 2021).

Taken together, these results support a narrow but consistent conclusion: multi-scale 1D convolution is an effective default for time-series classification, especially when the alternative is either a handcrafted transform pipeline or a single-scale temporal CNN. A broader claim that Inception1D is universally superior would go beyond the evidence, since the studies do not provide exhaustive cross-family benchmarking across all modern sequence models.

7. Relation to adjacent research and open directions

Inception1D sits at the intersection of several research trajectories: InceptionTime-style time-series classification, end-to-end raw-signal learning, hybrid CNN-Transformer architectures, and efficient replacements for conventional 2D convolution.

The infrasound study places InceptionTime alongside MiniRocket and a wavelet-plus-ResNet pipeline, choosing InceptionTime for its performance and direct support in tsai (Klenkert et al., 2024). EEG-Inception explicitly builds on the backbone of InceptionTime and adds domain-specific modifications such as the inverted bottleneck and noise-based augmentation (Zhang et al., 2021). The raw-audio inception nucleus paper shows that early 1D filters learned on waveform data resemble wavelet-like transforms, suggesting that end-to-end Inception1D modules can learn filterbanks that previously had to be engineered (Ebrahimpour et al., 2020).

A more distant but conceptually related development appears in work on oriented 1D kernels for image models. That paper argues that ConvNets consisting entirely of 1D convolutions can match 2D ConvNets on ImageNet classification if the kernels are oriented at multiple angles, and it describes block designs in which additional large-kernel oriented 1D branches augment standard architectures (Kirchmeyer et al., 2023). While this is not a time-series paper, it expands the meaning of “Inception1D” toward a general principle of replacing or augmenting standard convolution blocks with parallel branches of 1D convolutions of different extents or orientations (Kirchmeyer et al., 2023). This suggests that the Inception1D idea is not confined to temporal signals, though such an extension should be treated as an architectural analogy rather than a terminological identity.

Several limitations recur across the literature. The infrasound study uses simulated rather than real infrasound data and does not include detailed architecture ablations, cross-validation, or metrics beyond accuracy (Klenkert et al., 2024). InceptoFormer is demonstrated on Physionet gait data with fixed 100-step windows and does not analyze alternative kernel configurations inside Inception1D (Naimi et al., 6 Aug 2025). EEG-Inception shows that subject-independent performance remains substantially lower than subject-dependent performance, despite promising robustness indicators (Zhang et al., 2021). The raw-audio study evaluates only UrbanSound8K, so claims about transfer to speech, music, or biosignals remain prospective (Ebrahimpour et al., 2020).

A plausible implication is that future work on Inception1D will focus less on establishing the utility of multi-scale temporal branching in isolation, and more on how to integrate it with other inductive biases: residual routing, attention, sensor topology, learned or fixed frequency structure, and computational constraints. The existing literature already points in that direction, especially in hybrid designs where Inception1D supplies local multi-scale feature extraction and Transformers model longer-range dependencies (Naimi et al., 6 Aug 2025).