Dual-Pathway Learning in Modern AI

• Dual-Pathway Learning (DPL) is a design principle that employs two distinct yet complementary pathways to capture richer and robust data representations.
• It integrates different inductive biases using mechanisms like shared encoders, pseudo-labels, or gating to enhance performance across domains.
• Empirical evidence shows that DPL architectures outperform single-path models in tasks such as ECG analysis, semantic segmentation, and image denoising.

Dual-Pathway Learning (DPL) denotes a family of learning architectures in which two explicitly defined pathways are trained jointly to capture complementary structure that a single stream would represent less effectively. In current arXiv usage, the term does not refer to one canonical algorithm. Instead, it appears across several technical lineages: dual-objective self-supervision for ECG analysis, bidirectional state-space modeling for multivariate time series, paired domain-alignment pipelines for unsupervised semantic segmentation, noise-and-context fusion for medical image denoising, multilevel cross-modal discrimination for spatial transcriptomics, polarity-paired rectifier units for image denoising, dyadic activity-difference learning rules in biologically motivated networks, dual-input supervised classifiers, dual progressive branches for DeepFake detection, and dual compute-capacity transformer blocks for LLMs (Manimaran et al., 2024, Du et al., 2024, Cheng et al., 2021, Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Zhang et al., 2016, Høier et al., 2024, Zhang et al., 2018, Zhang et al., 2024, Frey et al., 28 May 2026).

1. Conceptual scope

Within these works, DPL is best understood as a design principle centered on complementarity rather than a fixed mathematical template. The pathways may differ by objective type, as in NERULA’s reconstruction and non-contrastive alignment branches; by temporal direction, as in forward and backward Mamba scans; by domain alignment, as in source-aligned and target-aligned semantic segmentation pipelines; by information source, as in context and noise estimators for denoising; by representational granularity, as in Gene-DML’s instance-level and instance-group discrimination; or by scaling axis, as in deep looped versus wide feed-forward transformer sublayers (Manimaran et al., 2024, Du et al., 2024, Cheng et al., 2021, Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Frey et al., 28 May 2026).

A recurrent property across these instantiations is that the two pathways are not merely duplicated branches. They are assigned different inductive biases and then coupled through shared encoders, shared pseudo-labels, tied weights, gating, clustering-derived prototypes, or explicit fusion modules. In NERULA, complementary masked views ensure that the pair jointly covers the full temporal support of the ECG; in semantic segmentation UDA, the target-side and source-side paths compensate for each other’s translation-induced inconsistencies; in CS-DPMamba, forward and backward selective state evolutions model asymmetric temporal dependencies; and in the adjoint dual-propagation framework, mean activity and compartmental difference play distinct forward and feedback roles (Manimaran et al., 2024, Cheng et al., 2021, Du et al., 2024, Høier et al., 2024).

The same term therefore spans several levels of abstraction. Some papers use DPL to describe a task-level training framework, some an architectural macro-pattern, and some a local activation or neuron-level construction. This heterogeneity is intrinsic to the literature rather than an inconsistency to be resolved away.

2. Recurring architectural motifs

One major motif is complementary view decomposition. NERULA forms two views of the same single-lead ECG, $x_i = m \times x$ and $x_j = (1-m) \times x$, and passes them through a shared Local Lead-Attention encoder; one path reconstructs the original signal from the masked view, while the other aligns latent vectors by cosine similarity without negatives. Medical image denoising adopts an analogous but supervised decomposition: a noise estimator predicts $\eta$, yielding $X'' = X - \eta$, a context estimator predicts $X' \approx Y$, and a third U-Net fuses $[X'; X'']$ into the final output (Manimaran et al., 2024, Fartiyal et al., 25 Jul 2025).

A second motif is directional duality. CS-DPMamba uses a forward selective SSM and a backward selective SSM over the time-reversed sequence, then fuses them as $y_t^{final} = \alpha y_t + \beta y_t^R$. The dual-propagation literature expresses a similar separation at the neuron level: dyadic states $s^+$ and $s^-$ encode mean activity and error differences simultaneously, and local updates take the form $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$ (Du et al., 2024, Høier et al., 2024).

A third motif is role-specialized pathways with shared supervision. In unsupervised domain adaptation for semantic segmentation, path-$x_j = (1-m) \times x$0 is target-aligned and path-$x_j = (1-m) \times x$1 is source-aligned. They interact through Dual Path Image Translation, which uses dual perceptual losses, and through Dual Path Adaptive Segmentation, which fuses target predictions as $x_j = (1-m) \times x$2 before thresholding pseudo-labels. In DeepFake detection, the two progressive branches are not source/target but quality-oriented and identifiability-oriented; both consume features from a shared backbone and a shared recurrent Feature Selection Module, then combine branch representations additively (Cheng et al., 2021, Zhang et al., 2024).

A fourth motif is parameter sharing or constrained coupling. DPLNets use two input branches that share all parameters and fuse features by a random interest value $x_j = (1-m) \times x$3, while the dual-pathway rectifier network constrains paired hidden units to have reversed input and output weights, producing the equivalent odd activation $x_j = (1-m) \times x$4. In the LLM setting, the deep path reuses a shared sublayer $x_j = (1-m) \times x$5 times, whereas the wide path applies a larger FFN once; per-token sigmoidal gates combine them as $x_j = (1-m) \times x$6 (Zhang et al., 2018, Zhang et al., 2016, Frey et al., 28 May 2026).

3. Objective design and optimization

Because DPL is not a single method, its learning objectives vary substantially. In self-supervised time-series settings, one pathway is often generative and the other discriminative. NERULA minimizes a weighted sum of cosine alignment and Huber reconstruction,

$x_j = (1-m) \times x$7

with $x_j = (1-m) \times x$8 and $x_j = (1-m) \times x$9 in reported experiments. The alignment term directly matches encoder outputs of complementary masked views, while the reconstruction branch forces the latent to support recovery of large missing ECG patches (Manimaran et al., 2024).

In graph-structured time-series classification, DPL is embedded inside a larger pipeline. CS-DPMamba first learns temporal representations with a margin-based contrastive loss, then constructs a FastDTW similarity matrix and performs node classification with a KAN-enhanced GIN. The overall objective is

$\eta$0

The dual-pathway component here is not the graph module but the bidirectional Mamba sequence model whose forward and backward hidden dynamics are fused before graph aggregation (Du et al., 2024).

In image translation and domain adaptation, the two pathways are trained by coupled supervised, self-training, and adversarial criteria. The dual-path image translation objective combines GAN losses, cycle-consistency losses, and a dual perceptual term with $\eta$1 and $\eta$2. The dual segmentation objective adds supervised source terms, pseudo-labeled target terms, and feature-level adversarial alignment, with $\eta$3 for DeepLab-V2 and $\eta$4 for FCN-8s. Shared pseudo-labels are selected by max-probability thresholding with $\eta$5 (Cheng et al., 2021).

Purely supervised DPL formulations also occur. Medical image denoising minimizes three MSE terms,

$\eta$6

corresponding to noise-path reconstruction, context-path reconstruction, and fused output reconstruction. Gene-DML combines three scale-wise image-gene discrimination losses, a cross-level instance-group alignment loss, and an MSE regression loss for gene prediction, with the best reported setting using $\eta$7 and $\eta$8. DPLNets minimize a dual prediction risk using the same random interest value for feature fusion and classification loss weighting,

$\eta$9

DeepFake DPL further departs from standard supervised training by using a two-stage procedure: Stage I optimizes classification plus an entropy-based regularizer, and Stage II freezes the detection modules while optimizing the Feature Selection Module with a PPO-style objective (Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Zhang et al., 2018, Zhang et al., 2024).

4. Representative instantiations across domains

Setting	Two pathways	Coupling mechanism
Single-lead ECG SSL (Manimaran et al., 2024)	reconstruction / non-contrastive alignment	shared encoder, complementary masking
MTS node classification (Du et al., 2024)	forward Mamba / backward Mamba	convex fusion, graph aggregation
Semantic segmentation UDA (Cheng et al., 2021)	target-aligned / source-aligned	dual perceptual translation, shared pseudo-labels
Medical image denoising (Fartiyal et al., 25 Jul 2025)	context estimator / noise estimator	concatenation into fusion U-Net
Gene expression prediction (Song et al., 19 Jul 2025)	multi-scale instance discrimination / instance-group discrimination	shared embedding space, cross-modal prototypes
Image denoising activation (Zhang et al., 2016)	rectifier / polarity-reversed companion	reversed weights, odd equivalent activation
Single-phase contrastive Hebbian learning (Høier et al., 2024)	mean activity / error difference	dyadic states, local Hebbian update
Dual-input classification (Zhang et al., 2018)	branch 1 / branch 2	shared weights, $X'' = X - \eta$0-weighted fusion
Cross-quality DeepFake detection (Zhang et al., 2024)	quality-progressive / identifiability-progressive	shared FSM, additive fusion
LLM scaling (Frey et al., 28 May 2026)	deep looped sublayer / wide sublayer	per-token dense gates

These examples illustrate that the pathways can operate over signals, domains, modalities, temporal directions, internal neuron states, or compute budgets. In healthcare-oriented work, DPL often separates complementary evidence sources: masked versus inverse-masked ECG support, context versus noise priors, or image morphology versus transcriptomic group structure. In computer vision adaptation and multimedia forensics, DPL typically manages distinct nuisances such as translation inconsistency, compression severity, or manipulation identifiability. In model-design papers, the term is sometimes used for a lower-level structural constraint, as in polarity-paired rectifiers, or for explicit compute-capacity factorization inside a transformer block (Manimaran et al., 2024, Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Cheng et al., 2021, Zhang et al., 2024, Zhang et al., 2016, Frey et al., 28 May 2026).

The role of inference also varies. Some DPL systems preserve both pathways at deployment, such as the DeepFake detector and DualPath LLM block. Others collapse to one branch at test time: the semantic-segmentation framework deploys only the target-aligned model $X'' = X - \eta$1 in its standard inference mode, and DPLNets can use a single branch with $X'' = X - \eta$2 or both branches with $X'' = X - \eta$3, yielding the same prediction because the parameters are shared (Cheng et al., 2021, Zhang et al., 2018).

5. Empirical evidence and ablation behavior

Reported gains are consistently tied to ablations that isolate the contribution of the second pathway. In NERULA, linear evaluation on PhysioNet 2017 arrhythmia classification showed Random Forest Accuracy $X'' = X - \eta$4 and F1 $X'' = X - \eta$5, compared with BYOL at Accuracy $X'' = X - \eta$6 and F1 $X'' = X - \eta$7, SimCLR at $X'' = X - \eta$8 and $X'' = X - \eta$9, Ti-MAE at $X' \approx Y$0 and $X' \approx Y$1, and CLOCS at $X' \approx Y$2 and $X' \approx Y$3. Its cumulative ablation reported Accuracy/F1 transitions from $X' \approx Y$4 for the BYOL baseline to $X' \approx Y$5 with masking augmentation, $X' \approx Y$6 with feature masking and Local Lead-Attention, and $X' \approx Y$7 after adding reconstruction; no statistical significance tests were reported (Manimaran et al., 2024).

In multivariate time-series node classification, CS-DPMamba achieved average accuracy $X' \approx Y$8 across 10 UEA datasets with 4 dataset wins, and semi-supervised averages of $X' \approx Y$9 and $[X'; X'']$0 using $[X'; X'']$1 and $[X'; X'']$2 labeled data. The reported ablations showed that ContrastFastDTW, DPMamba, and KAN-GIN each improved performance, while the full model performed best, supporting the claim that bidirectional temporal modeling, similarity construction, and graph aggregation are complementary (Du et al., 2024).

In unsupervised semantic segmentation, DPL reached mean IoU $[X'; X'']$3 on GTA5$[X'; X'']$4Cityscapes with ResNet-101 and $[X'; X'']$5/$[X'; X'']$6 mIoU on SYNTHIA$[X'; X'']$7Cityscapes for 16/13 classes, while the optional DPL-Dual inference variant further reached $[X'; X'']$8 and $[X'; X'']$9/$y_t^{final} = \alpha y_t + \beta y_t^R$0. Image-translation ablations improved from $y_t^{final} = \alpha y_t + \beta y_t^R$1 mIoU for naive CycleGAN to $y_t^{final} = \alpha y_t + \beta y_t^R$2 for single-path perceptual translation and $y_t^{final} = \alpha y_t + \beta y_t^R$3 for DPIT on $y_t^{final} = \alpha y_t + \beta y_t^R$4 after one DPAS iteration, and pseudo-label fusion by weighted averaging outperformed DPPLG-Max and DPPLG-Joint (Cheng et al., 2021).

In medical image denoising, the dual-path model achieved PSNR $y_t^{final} = \alpha y_t + \beta y_t^R$5 dB on Gaussian noise for the all-modalities setting, versus $y_t^{final} = \alpha y_t + \beta y_t^R$6 dB for the baseline U-Net, $y_t^{final} = \alpha y_t + \beta y_t^R$7 dB for RED-CNN, and $y_t^{final} = \alpha y_t + \beta y_t^R$8 dB for BM3D; the paper states this as a $y_t^{final} = \alpha y_t + \beta y_t^R$9 improvement over the baseline. Under AWGN and speckle noise, DPL improved over the baseline U-Net but did not surpass RED-CNN, and no statistical significance tests were reported. Gene-DML reported cross-validation results such as HER2ST MSE $s^+$0, PCC(A) $s^+$1, PCC(H) $s^+$2; STNet MSE $s^+$3, PCC(A) $s^+$4, PCC(H) $s^+$5; and skinST MSE $s^+$6, PCC(A) $s^+$7, PCC(H) $s^+$8, with ablations showing that removing either the multi-scale instance pathway or the cross-level instance-group pathway reduced performance (Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025).

In dual-branch supervised classification, DPLNets improved ResNet-152 on CIFAR-10 from test error $s^+$9 to $s^-$0 and on CIFAR-100 from $s^-$1 to $s^-$2, achieved MNIST test error $s^-$3 with a DPL LeNet-5, and reduced Google Commands test error from $s^-$4 to $s^-$5. In the earlier activation-level denoising work, the dual-ReLU model achieved the highest average PSNR across noise levels $s^-$6, including $s^-$7 at $s^-$8 compared with $s^-$9 for tanh and $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$0 for ReLU (Zhang et al., 2018, Zhang et al., 2016).

In cross-quality DeepFake detection, DPL achieved average AUC $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$1 when trained on FF++(c23) and tested on FF++(c23) with random JPEG, and average AUC $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$2 on FF++(c40) with random JPEG, outperforming the cited baselines in both settings. In LLMs, the DualPath block at the $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$3M budget achieved aggregate BPB $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$4 for the best $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$5, compared with $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$6 for PureWide and $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$7 for the best PureLoop, while using $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$8M parameters versus $$\Delta W_k \propto (s_{k+1}^+ - s_{k+1}^-)\bar s_k^\top$$9M for PureWide. At $x_j = (1-m) \times x$00M, the best aggregate BPB was $x_j = (1-m) \times x$01 for DualPath $x_j = (1-m) \times x$02, compared with $x_j = (1-m) \times x$03 for PureWide and $x_j = (1-m) \times x$04 for the best PureLoop (Zhang et al., 2024, Frey et al., 28 May 2026).

6. Limitations, misconceptions, and open directions

A central misconception is that DPL names one standardized architecture. The cited literature shows the opposite: the term ranges from branch-level fusion frameworks to activation-level constructions and neuron-level learning dynamics. It should also not be conflated with Dual Path Networks in the sense of macro-architectural DPNs; the rectifier-denoising paper explicitly distinguishes its activation-level dual-pathway pairing from DPN-style architecture design (Zhang et al., 2016).

A second misconception is that both pathways must remain active at inference. Some systems do so, but others use dual pathways only to improve training. The semantic segmentation framework is explicitly trained with two separate models, generators, and discriminators, yet its standard deployment uses only the target-aligned segmentation model. DPLNets likewise admit an inference reduction to a single branch because parameter sharing makes the two-branch and one-branch test procedures equivalent (Cheng et al., 2021, Zhang et al., 2018).

The main limitations are implementation-specific. NERULA fixed its $x_j = (1-m) \times x$05 masking ratio and $x_j = (1-m) \times x$06 heuristically and did not report ratio ablations. CS-DPMamba is sensitive to FastDTW radius, clustering quality, adjacency scaling, and top-$x_j = (1-m) \times x$07 sparsity. The semantic-segmentation DPL depends on translation quality and incurs the cost of training two segmentation models plus bidirectional translators and adversarial discriminators. Medical image denoising is heavier than a single U-Net because it trains three U-Nets and under AWGN and speckle does not exceed RED-CNN. Gene-DML depends on cluster count $x_j = (1-m) \times x$08 and centroid stability. DeepFake DPL depends on CLIP prompt-based discretization for time-step assignment. The original dual-propagation method is numerically stable only under symmetric nudging, motivating the adjoint DP$x_j = (1-m) \times x$09 variant. The DualPath LLM block does not save inference FLOPs through conditional skipping because both paths are always evaluated; its gates reallocate contributions rather than eliminate computation (Manimaran et al., 2024, Du et al., 2024, Cheng et al., 2021, Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Zhang et al., 2024, Høier et al., 2024, Frey et al., 28 May 2026).

The open directions reported in these papers remain diverse. They include multi-lead ECG extensions and broader architectural searches for ECG self-supervision, other dense-prediction tasks and alternative translation frameworks for dual-path domain adaptation, larger OCT datasets and better noise estimators for medical denoising, extension to other omics and spatial graph constraints for cross-modal pathology, broader forensics applications for dual progressive learning, and combinations of compute-capacity dual paths with MoE or memory modules in LLMs (Manimaran et al., 2024, Cheng et al., 2021, Fartiyal et al., 25 Jul 2025, Song et al., 19 Jul 2025, Zhang et al., 2024, Frey et al., 28 May 2026).

Taken together, these works indicate that DPL is a general strategy for organizing complementary inductive biases into a jointly optimized pair of pathways. What remains stable across usages is not a specific architecture, loss, or domain, but the proposition that explicit dualization—when the two streams are meaningfully distinct and tightly coupled—can yield representations that are more complete, more robust, or more efficiently allocated than those learned by a single path alone.