TAHCD: Adaptive Hierarchical Denoising Network

• The paper demonstrates that TAHCD achieves state-of-the-art robustness by hierarchically aligning modality features and mitigating both modality-specific and cross-modality noise.
• TAHCD integrates adaptive strategies (ASSA, SACA, TTCE) to refine feature representations at global and instance levels without relying on labeled data.
• Experimental results show TAHCD outperforms existing methods in noisy settings, maintaining high accuracy with minimal performance degradation.

Test-time Adaptive Hierarchical Co-enhanced Denoising Network (TAHCD) is a multimodal classification method engineered to deliver robust, reliable learning in the presence of heterogeneous and previously unseen data noise. TAHCD jointly removes modality-specific and cross-modality noise via a hierarchical denoising architecture, adaptively refining feature representations at both the global and instance levels, including a test-time label-free cooperative enhancement loop. The approach achieves state-of-the-art generalization and robustness, outperforming prior art under various noisy and mismatched train/test scenarios (Shen et al., 12 Jan 2026).

1. Architectural Overview and Multimodal Integration

TAHCD operates on datasets of $N$ samples, each sample $x_i=\{x_i^1,\dots,x_i^M\}$ comprising $M$ modalities. Each modality is transformed by its encoder $\phi_x^m$ into latent features $z^m=\phi_x^m(x^m)\in\mathbb{R}^{N\times d^m}$, typically corrupted by (i) modality-specific noise (e.g., sensor or sampling errors) and (ii) cross-modality noise (e.g., misalignment, sample shuffling). The TAHCD pipeline is organized into three interactive stages:

• Adaptive Stable Subspace Alignment (ASSA): Suppresses global-level noise and aligns modalities to a common stable subspace via subspace masking and orthogonality/projection alignment losses.
• Sample-Adaptive Confidence Alignment (SACA): Removes residual noise on a per-instance basis using confidence-guided experts and asymmetric slack alignment against learned Gaussian priors.
• Test-Time Cooperative Enhancement (TTCE): At inference, TTCE updates both global and instance denoising components iteratively, adapting the model in a label-free manner to previously unseen noise.

The architecture exploits hierarchical co-enhancement, where global denoising informs fine-grained adjustments and vice versa, facilitating “bootstrap” improvements in representation quality under variable noise.

2. Adaptive Stable Subspace Alignment (ASSA)

ASSA is responsible for global-level feature denoising and multi-modal alignment. For each modality $m$:

• Subspace Construction:
  • Compute per-modality covariance $\Sigma_z^m$ and perform SVD: $\Sigma_z^m=U_z^m\Lambda_z^m(U_z^m)^\top$, where $\mu_z^m$ is the mean feature vector.
  • Learn a soft mask $$w_\lambda^m=\sigma(\phi_\lambda^m(\operatorname{diag}(\Lambda_z^m)))$$ to suppress directions dominated by noise.
  • Reconstruct globally denoised representations $$h^m = z^m U_z^m \operatorname{diag}(w_\lambda^m) (U_z^m)^\top$$.
• Global-Level Losses:
  • Inter-class orthogonality: $$\mathcal{L}_o = \frac{1}{C(C-1)}\sum_{m=1}^M\sum_{c\neq c'}\|\left(\mu^{m,c}\right)^\top \mu^{m,c'}-I\|_F^2$$ penalizes modality-specific spurious patterns.
  • Subspace projection alignment: $$\mathcal{L}_a = \sum_{m=1}^M \sum_{m'\neq m} \|z^m U_z^m \operatorname{diag}(w_\lambda^m) - z^{m'} U_z^{m'} \operatorname{diag}(w_\lambda^{m'})\|_F^2$$, aligning features across modalities.

The total ASSA loss is $$\mathcal{L}_{\text{assa}} = \mathcal{L}_o + \mathcal{L}_a$$, targeting robust multimodal subspaces invariant to both noise types.

3. Sample-Adaptive Confidence Alignment (SACA)

SACA addresses instance-level denoising, leveraging Gaussian priors from globally cleaned features:

• Prior Modeling: Fit priors $p_h^m=\mathcal{N}(\mu_h^m, \Sigma_h^m)$ and cross-modality priors $$p_h^{m-m'}=\mathcal{N}(\mu_h^m-\mu_h^{m'}, \Sigma_h^m+\Sigma_h^{m'})$$.
• Noise Experts: For each sample $i$ and modality $m$, modality-specific experts $E_s^m$ and cross-modality experts $E_c^m$ estimate confidence masks and apply elementwise filtering:
  • $w_{s,i}^m=\sigma(\phi_s^m(h_i^m))$; $\hat{h}_{s,i}^m=h_i^m\odot w_{s,i}^m$; $n_{s,i}^m=h_i^m\odot(1-w_{s,i}^m)$.
  • $w_{c,i}^m=\sigma(\phi_c^m(h_i^m))$; $\hat{h}_{c,i}^m=h_i^m\odot w_{c,i}^m$; $n_{c,i}^m=h_i^m\odot(1-w_{c,i}^m)$.
• Confidence-Aware Asymmetric Slack Alignment: Negative log-likelihood under priors penalizes deviation from clean distributions. Updates are confidence-weighted, with stronger correction applied to lower-confidence modalities:
  • $$\theta_c^m \leftarrow \theta_c^m - \eta\frac{1}{|B|}\sum_{i\in B}\neg \mathrm{conf}_i^m\nabla_{\theta_c^m}\mathcal{L}_{\text{nll}^c}(\hat{h}_{c,i}^m)$$.

This adaptive correction mechanism enables targeted denoising on outlier or contaminated samples, guided by both global and sample-level statistical structure.

4. Test-Time Cooperative Enhancement (TTCE)

TTCE enables dynamic adaptation at inference, crucial for real-world deployments where noise conditions may differ from training:

• Global Reconstruction: Minimize $$\mathcal{L}_{\text{re}} = \frac{1}{M}\sum_{m=1}^M \| \Psi^m (h^m + n_s^m + n_c^m) - x^m \|_F^2$$, where $\Psi^m$ is a modality-m decoder.
• Feature and Prior Updates: Iteratively apply gradient updates to $h^m$, and update Gaussian prior statistics via moving averages:
  • $$\mu^m \leftarrow \alpha \mu^m + (1-\alpha)\Delta \mu^m$$, $$\Sigma^m \leftarrow \beta \Sigma^m + (1-\beta)\Delta \Sigma^m$$.
• Instance Newton-like Updates: For each sample and modality, update $\hat{h}_{s,i}^m$ and $\hat{h}_{c,i}^m$ to minimize $\mathcal{L}_{\text{saca}}$ using the inverse covariance matrices.
• Cooperative Iteration: Repeat the global/instance denoising and prior updates for $E\approx30$ iterations per test batch.

TTCE instantiates a label-free model refinement mechanism whereby global and local denoising reinforce each other, adapting the network to unseen noise distributions without retraining or annotated labels.

5. Hierarchical Co-enhancement and Denoising Bootstrapping

TAHCD’s architecture embodies a hierarchical feedback loop:

• ASSA → SACA: Global denoising outputs a clean subspace and statistical priors; instance-level experts operate on these to suppress residual noise.
• SACA → TTCE → ASSA: Instance-level corrections motivate refined global reconstruction and prior updates during TTCE, sharpening the global representations and priors for subsequent SACA rounds.

This interplay, described as “bootstrap” denoising, ensures progressive improvement of both modal and sample-level representations, enhancing robustness against noise heterogeneity and distributional shifts.

6. Implementation Details

TAHCD deploys established deep learning backbones tailored to each modality:

• Encoders ($\phi_x^m$):
  • Images: ResNet-50 ($d^m=512$)
  • Text: 12-layer Transformer (BERT-base, $d^m=768$)
  • Omics/tabular: 3-layer MLP ($d^m=256$)
• Decoders ($\Psi^m$): Architecturally mirrors encoding networks (e.g., transpose convolution for images).
• Classifier: One-hidden-layer MLP on fused multimodal features.
• Optimization: Adam optimizer ($lr=1e-4$, $weight\_decay=1e-4$, $lr$ decay by $0.2$ per plateau). TTCE loop uses $E=30$ iterations, moving average coefficients $\alpha=0.4$, $\beta=0.3$, mini-batch size 32. The composite objective is $$\mathcal{L}_{\text{tot}}=\mathcal{L}_\text{assa}+\mathcal{L}_\text{saca}+\mathcal{L}_\text{re}+\mathcal{L}_\text{cls}$$.

7. Experimental Results and Analytical Findings

TAHCD was evaluated across omics (BRCA, ROSMAP) and image-text (CUB, FOOD101) benchmarks, using controlled additive Gaussian modality-specific noise ($\epsilon\in\{0,1,5\}$) and cross-modality shuffling ($\eta\in\{0\%,10\%,20\%\}$):

• Robustness: Under matched train/test noise, TAHCD outperformed all baselines (MD, MLCLNet, QMF, PDF, NCR, ALBEF, SMILE, SPS) by 2–8 points (ACC). Competing methods suffered marked degradation under either noise; TAHCD exhibited stable performance.
• Generalization: When trained on clean data but tested under severe noise ($\epsilon=5,\;\eta=10\%$), TAHCD’s accuracy declined only 3–5 points; other methods lost 20–30 points.
• Ablation Study: Isolating components revealed cumulative benefits: ASSA alone outperformed no denoising; adding SACA yielded another 3–5 points; TTCE added 2–4 points improvement.
• Component Analysis:
  • Removing $\mathcal{L}_o$ compromised modality-specific noise handling.
  • Removing $\mathcal{L}_a$ undermined cross-modality robustness.
  • Alignment on latent projections was critical under heterogeneous noise.
  • Slack alignment outperformed strict similarity and mutual information objectives.
  • Confidence weighting accelerated convergence and improved alignment quality.
  • TTCE iteration shifted embeddings towards clean clusters.

A plausible implication is that hierarchical, confidence-weighted, and test-time adaptive strategies are essential for robust multimodal learning in adversarial or real-world settings (Shen et al., 12 Jan 2026).