InceptoFormer: Multi-Branch Transformer Models

• InceptoFormer is an architectural motif that fuses inception-style multi-branch local processing with Transformer-based global context modeling across various domains.
• It encompasses distinct variants, including vision backbones with frequency-aware token mixing, segmentation models with pyramid encoders, and gait analysis frameworks for Parkinson’s staging.
• Key challenges include nomenclature ambiguity, parameter tuning such as high/low frequency split ratios, and adaptations specific to visual versus biomedical applications.

InceptoFormer is a non-unique label used for several architectures that combine an inception-style multi-branch inductive bias with Transformer components. In the vision-backbone literature, the nickname “InceptoFormer” refers to the Inception-style token mixer in the Inception Transformer, or iFormer, which explicitly fuses high-frequency local processing and low-frequency global self-attention within each block (Si et al., 2022). In semantic segmentation, the spelling “IncepFormer” denotes a pyramid-structured Transformer encoder with Inception Multi-Head Self-Attention and a lightweight decoder (Fu et al., 2022). In gait analysis, “InceptoFormer” names a multi-signal neural framework for Parkinson’s Disease severity staging that combines Inception1D, temporal transformers, and a spatial transformer over 18 VGRF channels (Naimi et al., 6 Aug 2025). A common source of confusion is that these systems share a family resemblance but are architecturally and application-wise distinct.

1. Nomenclature and referential ambiguity

The term has been applied to multiple models rather than a single canonical architecture.

Designated name	Paper	Domain
Inception Transformer / iFormer	(Si et al., 2022)	Vision backbone for classification, detection, and segmentation
IncepFormer	(Fu et al., 2022)	Fully supervised semantic segmentation
InceptoFormer	(Naimi et al., 6 Aug 2025)	Parkinson’s Disease severity evaluation from gait

In the iFormer paper, the nickname “InceptoFormer” refers to the Inception-style token mixer that sits inside each Transformer block. In the segmentation paper, the correct name is IncepFormer, and the “InceptoFormer” spelling is noted as an alternative rendering of the same model. In the gait paper, InceptoFormer is the formal model name. This naming overlap has created a terminological collision across subfields (Si et al., 2022, Fu et al., 2022, Naimi et al., 6 Aug 2025).

2. iFormer as a frequency-aware hierarchical vision backbone

The Inception Transformer addresses a specific diagnosis of ViTs: strong long-range dependency modeling but weak capture of high-frequency local information. Its central claim is that global token-to-token self-attention over non-overlapping patches behaves like a low-pass mechanism, over-emphasizing low-frequency global structure while under-capturing local edges and textures. iFormer therefore performs explicit per-block frequency mixing. Each pre-norm block follows the sequence LayerNorm $\rightarrow$ Inception Token Mixer $\rightarrow$ residual add $\rightarrow$ LayerNorm $\rightarrow$ FFN $\rightarrow$ residual add. The Inception Token Mixer splits input channels into a high-frequency subset $X_{HF}$ and a low-frequency subset $X_{LF}$, processes them in parallel, concatenates the outputs, and fuses them through

$$Y = \mathrm{FC}(Y_c + \mathrm{DwConv3\times3}(Y_c)).$$

The high-frequency path itself has two branches: a max-pooling branch, $Y_{h1}=\mathrm{FC}(\mathrm{MaxPool}(X_{h1}))$, and a convolution branch, $$Y_{h2}=\mathrm{DwConv3\times3}(\mathrm{FC}(X_{h2}))$$. The low-frequency path applies vanilla MSA to the pooled low-frequency subset, with

$\rightarrow$0

Average pooling with stride $\rightarrow$1 is used only in stages 1 and 2, and nearest-neighbor upsampling restores the original stage resolution afterward.

A distinctive element is the frequency ramp. The model uses four stages at progressively reduced resolutions, and the split ratios satisfy $\rightarrow$2 and $\rightarrow$3, with $\rightarrow$4 monotonically decreasing over depth. Stage 1 uses $\rightarrow$5, stage 2 uses $\rightarrow$6, stage 3 uses a within-stage linear ramp, and stage 4 assigns only a small residual fraction to the high-frequency mixer: $\rightarrow$7 for iFormer-S, $\rightarrow$8 for iFormer-B, and $\rightarrow$9 for iFormer-L. The resulting per-block complexity is

$\rightarrow$0

with $\rightarrow$1, which is explicitly positioned as cheaper than applying attention to all channels at full resolution. The backbone is hierarchical, uses two $\rightarrow$2 stride-2 convolutions for stage-1 patch embedding to reach $\rightarrow$3 resolution, and uses $\rightarrow$4 stride-2 patch embeddings between later stages. Main variants are iFormer-S at 20M parameters and 4.8G FLOPs, iFormer-B at 48M and 9.4G, and iFormer-L at 87M and 14.0G. On ImageNet-1K, the reported top-1 accuracies are 83.4%, 84.6%, and 84.8%, respectively; after fine-tuning at $\rightarrow$5, they become 84.6%, 85.7%, and 85.8%. On COCO detection and instance segmentation with Mask R-CNN $\rightarrow$6, iFormer-S reports $\rightarrow$7 and $\rightarrow$8, while iFormer-B reports 48.3 and 43.4. On ADE20K with Semantic FPN, iFormer-S reaches 48.6 mIoU. Ablations further show that the full Inception mixer improves over an attention-only mixer from 80.8% to 81.2% top-1 in a 100-epoch ImageNet-1K setting, and that the proper ramp outperforms an inverse ramp and a flat split at 81.2%, 80.5%, and 80.7%, respectively (Si et al., 2022).

3. IncepFormer as a semantic segmentation architecture

IncepFormer is a segmentation-specific Transformer whose objective is to capture global context, fine localisation information, and multi-scale features in a single hierarchical encoder-decoder pipeline. The input image $\rightarrow$9 is split into $\rightarrow$0 patches, and a four-stage pyramid encoder produces feature maps at $\rightarrow$1 resolution. Each stage contains Inception Transformer Blocks. Inter-stage downsampling follows the patch- or feature-merging idea of PVT and SegFormer, halving spatial size and increasing channel width. The decoder is an Upsample-Concat head: all stage features are upsampled to $\rightarrow$2, concatenated channel-wise, fused by a $\rightarrow$3 convolution, and mapped to per-pixel logits by a second $\rightarrow$4 convolution.

The block-level design differs materially from iFormer. Pre-normalization uses BatchNorm rather than LayerNorm. Each block is

$\rightarrow$5

No positional encodings are used. Incep-MHSA reduces the token count for keys and values through three depth-wise branches: a strip-kernel path using $\rightarrow$6 followed by $\rightarrow$7, a $\rightarrow$8 local-kernel path with stride $\rightarrow$9, and a pooling branch $\rightarrow$0. Their outputs are flattened, concatenated, normalized, and used as reduced $\rightarrow$1 inputs while $\rightarrow$2 remains at full resolution:

$\rightarrow$3

$\rightarrow$4

The reduction ratios are stage-specific: $\rightarrow$5 for stages 1 through 4. The E-FFN replaces the standard token MLP with a 2D convolutional sequence $\rightarrow$6 and residual addition, preserving locality and avoiding repeated sequence-to-image reshaping within the block.

Three model scales are reported. IncepFormer-T uses decoder fusion width $\rightarrow$7 and has approximately 14.0M parameters; IncepFormer-S uses $\rightarrow$8 and approximately 24.6M parameters; IncepFormer-B also uses $\rightarrow$9 and approximately 39.6M parameters. On ADE20K, the single-scale and multi-scale mIoU values are 44.8/46.3 for Tiny, 47.7/49.0 for Small, and 49.4/50.2 for Base, with 21.2, 38.5, and 54.6 GFLOPs, respectively, under $X_{HF}$0 FLOPs accounting. On Cityscapes at $X_{HF}$1, the reported single-scale and multi-scale mIoU values are 80.5/81.8, 81.6/82.6, and 82.0/82.7. On COCO-Stuff, Pascal Context, and Pascal VOC, the model is also evaluated under the same general training setup. Ablations show that decoder fusion width exhibits diminishing returns beyond $X_{HF}$2, and that the Upsample-Concat decoder outperforms the tested SegFormer-like and SETR-like alternatives for the Tiny variant; removing stage-1 features from the decoder slightly reduces performance by 0.2 mIoU (Fu et al., 2022).

4. InceptoFormer for Parkinson’s Disease severity staging from gait

In the gait-analysis literature, InceptoFormer is a multi-signal neural framework for Parkinson’s Disease severity evaluation from Vertical Ground Reaction Force signals measured by 18 foot-mounted sensors. The target labels are four Hoehn and Yahr levels available in the PhysioNet gait database: Stage 0 for healthy controls, Stage 1 for mild with H&Y=2, Stage 2 for moderate with H&Y=2.5, and Stage 3 for severe with H&Y=3. The dataset contains 93 PD patients and 73 healthy controls, and the recorded walking conditions come from three teams: Yogev et al. for normal walking on level surface, Hausdorff et al. for casual-speed walking with rhythmic auditory stimulation, and Toledo et al. for treadmill walking. Walks are divided into segments of length 100 time steps with 50% overlap, and each segment inherits the subject label.

The architecture has five sequential stages: 18 parallel branches, one per VGRF channel; an Inception1D component repeated three times per branch; a temporal transformer encoder per signal; per-branch dimensionality reduction; concatenation across signals followed by a spatial transformer and a classifier head. Each Inception1D module has three parallel 1D convolution streams with kernel sizes $X_{HF}$3, each using 32 filters. Their outputs are concatenated,

$X_{HF}$4

with SeLU activation and Batch Normalization. The convolutional branches model multi-scale local temporal patterns, while the transformer components model long-range dependencies. The temporal transformer uses multi-head attention with two heads and a feed-forward network for each signal independently. The spatial transformer, also with two heads and an FFN, operates after concatenation of the reduced per-signal features to model inter-sensor dependencies. The classifier consists of two fully connected layers with SeLU activations, followed by a four-class softmax. The attention, softmax, and cross-entropy definitions follow standard forms:

$X_{HF}$5

Class imbalance is addressed before training through SMOTE applied only to the minority classes 0 and 3. For minority samples $X_{HF}$6 and $X_{HF}$7 and $X_{HF}$8, synthetic data are generated as

$X_{HF}$9

and the number of new samples is chosen so that classes 0 and 3 match the class-1 count. Training uses Nadam with learning rate $X_{LF}$0, batch size 64, dropout 0.2, and early stopping on validation loss. Evaluation is performed with 10-fold cross-validation, preserving a 70% PD / 30% control ratio per fold. The reported headline metrics are 96.6% accuracy, 93.97% precision, 94.15% recall, and 93.60% F1. Averaged confusion values are 97% correct for Stage 0, 97% for Stage 1, 98% for Stage 2, and 94% for Stage 3. Compared baselines include OF-DDNet at 84.00% accuracy, a 1D-ConvNet at 85.30%, a feedforward neural network at 86.50%, a 1D-Convolutional Transformer at 88.00%, and the same baseline retrained with SMOTE at 89.12%. Ablations isolate the complementary roles of local and global modeling: an Inception1D-only model reaches 78.34% accuracy, a transformer-only model 89.17%, and the full InceptoFormer 96.60% (Naimi et al., 6 Aug 2025).

5. Shared design motifs and decisive differences

All three architectures combine local multi-scale operators with Transformer-style attention, but the placement of those operators is different. iFormer performs split-channel frequency mixing inside each vision block, assigning convolution and max-pooling to a high-frequency subset and self-attention to a low-frequency subset. IncepFormer for segmentation integrates its multi-branch structure directly into self-attention by reducing keys and values through strip-kernel, $X_{LF}$1, and pooling branches. The PD-staging InceptoFormer separates local and global modeling into distinct modules: Inception1D for per-signal temporal feature extraction, then temporal and spatial transformers for long-range dependencies. These are not minor implementation variants of one shared architecture; they instantiate different decompositions of the same broad design idea (Si et al., 2022, Fu et al., 2022, Naimi et al., 6 Aug 2025).

The hierarchy of representation also differs. iFormer and IncepFormer are four-stage pyramidal visual encoders operating at spatial resolutions down to $X_{LF}$2 of the input. The gait model is not pyramidal in that sense; it operates on fixed-length 100-step windows and 18 sensor streams. The target outputs are likewise different: iFormer is a general-purpose visual backbone evaluated on image classification, COCO detection, and ADE20K segmentation; IncepFormer is a fully supervised per-pixel segmentation system; the gait InceptoFormer is a four-class clinical staging classifier. This suggests that “InceptoFormer” functions less as the name of a single established model family than as a recurrent architectural trope: inception-like local feature extraction plus Transformer-based global context.

6. Limitations, misconceptions, and research directions

Several misconceptions arise from the name overlap. The first is that InceptoFormer denotes one standardized architecture; the literature instead uses the term for at least three distinct systems. The second is that all such models target vision. The 2025 gait paper demonstrates a separate usage in biomedical time series. A third misconception is that the inception component plays the same role in each case. In iFormer it is a frequency-aware token mixer, in IncepFormer it is a token-reduction mechanism inside MHSA plus a convolutional FFN, and in the gait model it is a 1D temporal feature extractor preceding transformers (Si et al., 2022, Fu et al., 2022, Naimi et al., 6 Aug 2025).

The reported limitations are also domain-specific. For iFormer, the split ratios $X_{LF}$3 and $X_{LF}$4 are manually specified and performance can be sensitive to this schedule; large-scale pretraining beyond ImageNet-1K was not explored; and transfer to non-visual domains would require reassessment of the mixers and splits. For IncepFormer, the paper identifies scaling to very large models and transfer beyond segmentation as future work, and notes that aggressive downsampling can still challenge extremely small objects if the fusion width is too narrow or shallow features are underused. For the gait InceptoFormer, cross-cohort generalization is not demonstrated, subject independence of cross-validation folds is not stated, sensor noise and protocol variability are not analyzed explicitly, and only four H&Y levels are available because of dataset constraints. A plausible implication is that the shared multi-branch-plus-attention pattern is robust across modalities, but reproducibility and external validation remain highly application-dependent.