Decoupled Multimodal Distillation

• DMD is a framework that decouples shared and modality-specific components to achieve robust multimodal learning.
• Dynamic graph distillation employs sample-adaptive graphs to regulate crossmodal knowledge transfer effectively.
• Hierarchical and fine-grained alignment techniques enhance feature interpretability and improve empirical performance on benchmarks.

Decoupled Multimodal Distillation (DMD) refers to a class of frameworks and algorithms that leverage explicit separation of shared (modality-irrelevant) and private (modality-exclusive) components in multimodal signals, combined with specialized distillation mechanisms. These approaches aim to address the persistent heterogeneity and unequal contribution inherent in multimodal learning tasks such as emotion recognition, audio-visual dataset distillation, and related applications. DMD incorporates dynamic, sample-adaptive crossmodal distillation, usually instantiated through learnable graph-based modules, to enhance the discriminativeness and interpretability of the resulting multimodal feature spaces (Li et al., 2023, Li et al., 22 Nov 2025, Li et al., 4 Feb 2026).

1. Decoupled Representation Learning

Central to DMD is the decoupling of each modality’s representation into two parts: (1) a modality-irrelevant (homogeneous) component, modeled to contain information shared across modalities, and (2) a modality-exclusive (heterogeneous) component, representing strictly modality-specific information. Given low-level temporal features $$\widetilde{\mathbf X}_m \in \mathbb{R}^{T_m \times d}$$ for modality $m$, encoding proceeds as: $$\mathbf X_m^{\rm com} = \mathcal E^{\rm com}(\widetilde{\mathbf X}_m), \quad \mathbf X_m^{\rm prt} = \mathcal E_m^{\rm prt}(\widetilde{\mathbf X}_m)$$ where $\mathcal E^{\rm com}$ is shared and $\mathcal E_m^{\rm prt}$ is modality-specific (Li et al., 2023).

Decoupling constraints include:

• Self-regression loss: joint reconstruction via private decoder $\mathcal D_m$,

$$\mathcal L_{\rm rec} = \|\widetilde{\mathbf X}_m - \mathcal D_m([\mathbf X_m^{\rm com}, \mathbf X_m^{\rm prt}])\|^2_F$$

• Cycle-consistency: private part recovery after reconstruction,

$$\mathcal L_{\rm cyc} = \|\mathbf X_m^{\rm prt} - \mathcal E_m^{\rm prt}(\mathcal D_m([\mathbf X_m^{\rm com}, \mathbf X_m^{\rm prt}]))\|^2_F$$

• Margin-based loss: pulls together same-class, different-modality homogeneous pairs and separates same-modality, different-class pairs by a margin $\alpha$,

$$\mathcal L_{\rm mar} = \frac{1}{|S|} \sum_{(i,j,k)\in S} \max\left(0,\, \alpha - \cos(\mathbf x^{\rm com}_{m[i]},\mathbf x^{\rm com}_{m[j]}) + \cos(\mathbf x^{\rm com}_{m[i]},\mathbf x^{\rm com}_{m[k]})\right)$$

• Orthogonality penalty: discourages redundancy between $\mathbf X_m^{\rm com}$ and $\mathbf X_m^{\rm prt}$,

$$\mathcal L_{\rm ort} = \sum_{m} \cos(\mathbf X_m^{\rm com}, \mathbf X_m^{\rm prt})$$

• The decoupling loss aggregates these components:

$$\mathcal L_{\rm dec} = \mathcal L_{\rm rec} + \mathcal L_{\rm cyc} + \gamma (\mathcal L_{\rm mar} + \mathcal L_{\rm ort})$$

This paradigm is observed in various DMD instantiations, including audio-visual dataset distillation frameworks, where decoupling is performed via shallow MLP-based decouplers attached to frozen pretrained encoders (Li et al., 22 Nov 2025), and hierarchical models that support additional fine-grained alignment (Li et al., 4 Feb 2026).

2. Dynamic Graph Distillation Mechanisms

DMD utilizes Graph Distillation Units (GD-Units)—parameterized, directed graphs whose vertices correspond to modalities and whose edges encode dynamic, learned crossmodal distillation weights. For representations $\mathbf H_i$ (either homogeneous or heterogeneous),

• Node logits: $\ell_i = f(\mathbf H_i; \theta_1)$
• Edge weights:

$$w_{i \to j} = {\rm softmax}_i\, g \left( [f(\mathbf H_i), \mathbf H_i] \,\Vert\, [f(\mathbf H_j), \mathbf H_j];\, \theta_2 \right)$$

Edge weights $w_{i\to j}$ specify "how strongly $i$ should teach $j$"; a softmax normalization ensures their sum over all $i$ targeting $j$ is 1.

• Distillation error:

$\epsilon_{i\to j} = \|\ell_i - \ell_j\|^2_2$

• Graph-distillation loss:

$$\mathcal L_{\rm dtl} = \sum_{i,j} w_{i\to j} \,\epsilon_{i\to j} = \| W \odot E \|_1$$

DMD typically deploys two such units:

• HomoGD: operates in the homogeneous (shared) feature space, directly on $\{\mathbf X_m^{\rm com}\}$.
• HeteroGD: operates in the heterogeneous space, usually after cross-modal alignment (e.g., MulT-style cross-attention transformer operations).

In DHMD, coarse-grained distillation is implemented with GD-Units complemented by finer-grained mechanisms (see Section 4) (Li et al., 2023, Li et al., 4 Feb 2026).

3. Sample- and Distribution-Level Alignment

Beyond pairwise distillation, certain DMD variants explicitly enforce sample- and distribution-level alignment in the shared (common) subspace. For instance, DAVDD incorporates:

• Sample-level contrastive (InfoNCE) loss between crossmodal pairs $c_a$ (audio common) and $c_v$ (visual common),

$L_{A\to V},\,L_{V\to A}$

• Distribution-level alignment via exponential moving average class prototypes $P^A_c$, $P^V_c$ and per-batch class means:

$$L_{\rm align} = \frac{1}{|C_b|} \sum_{c\in C_b} [d_{\rm cos}(\mu_c^A, P^V_c) + d_{\rm cos}(\mu_c^V, P^A_c)]$$

Combined (weighted) alignment loss:

$$L_{sd} = \lambda_{\rm intra} L_{\rm intra} + \lambda_{\rm align} L_{\rm align}$$

Private (modality-specific) representations are optimized via moment matching only, with no crossmodal interactions:

$$L_{\rm pr}^A = \| \mu^A_{\rm real} - \mu^A_{\rm synth}\|^2, \quad L_{\rm pr}^V = \| \mu^V_{\rm real} - \mu^V_{\rm synth}\|^2$$

This separation ensures that private information is preserved for each modality, addressing a common limitation in prior distribution matching approaches (Li et al., 22 Nov 2025).

4. Hierarchical and Fine-grained Distillation Extensions

The DHMD framework introduces decoupled hierarchical multimodal distillation, extending DMD with two-stage knowledge transfer:

• Coarse-grained distillation: as in prior DMD, with GD-Units on both homogeneous and heterogeneous spaces.
• Fine-grained stage: implements crossmodal dictionary matching via a learnable dictionary $D \in \mathbb{R}^{K \times C}$, representing $K$ "atoms." For a feature $Z\in\mathbb{R}^{T\times C}$,
  • Compute attention $A = Z D^\top$,
  • Aggregate via columnwise max, followed by softmax normalization,
  • Aggregate features via weighted sum over atoms.

Contrastive losses enforce semantic alignment on both homogeneous and heterogeneous spaces at this fine-grained level.

The addition of dictionary matching not only augments class cluster separability in feature space but also strengthens previously weak crossmodal graph edges in the GD-Unit, facilitating more robust multi-way knowledge transfer across modalities (Li et al., 4 Feb 2026).

5. Objective Functions and Optimization

Across DMD frameworks, the full training objective comprises: $$\mathcal L_{\rm total} = \mathcal L_{\rm task} + \lambda_1 \mathcal L_{\rm dec} + \lambda_2 (\mathcal L_{\rm dtl}^{\rm homo} + \mathcal L_{\rm dtl}^{\rm hetero}) + \lambda_3 \mathcal L_{\rm dic}$$ where $\mathcal L_{\rm task}$ is the task-specific loss (e.g., mean absolute error for emotion regression), and the remaining terms govern representation decoupling, graph-based distillation, and (for DHMD) dictionary matching. Parameters $\lambda_1, \lambda_2, \lambda_3$ respectively control the importance of each component in the loss function (Li et al., 2023, Li et al., 4 Feb 2026).

Training is generally end-to-end and does not require separate pretrain-finetune stages for decoupling or distillation. In audio-visual distillation settings, optimization alternates between decoupling head updates on real data and distillation loss updates on synthetic data (Li et al., 22 Nov 2025).

6. Empirical Performance and Visualization Insights

DMD consistently achieves superior performance over prior state-of-the-art multimodal learning frameworks (e.g., MulT, MISA, FDMER, AVDD, DM) on multiple benchmarks:

• On CMU-MOSI and CMU-MOSEI, DMD and DHMD achieve 1–2% absolute improvement in ACC$_7$, ACC$_2$, and F1-score compared to prior arts. DHMD additionally provides gains on UR-FUNNY and MUStARD (Li et al., 4 Feb 2026).
• In dataset distillation contexts, DMD-based methods (e.g., DAVDD) set state-of-the-art performance on VGGS-10K, MUSIC-21, and AVE, with particular gains at low instances-per-class (IPC) regimes (Li et al., 22 Nov 2025).

Visualization analyses reveal that:

• Homogeneous (shared) decoupled spaces form emotion-centric clusters; heterogeneous (private) spaces form modality-centric clusters.
• GD-Unit learned edge weights display interpretable, dynamic teaching patterns. For example, in emotion recognition, language often teaches vision and audio in the homogeneous graph, whereas vision substantially influences audio after cross-modal alignment in the heterogeneous graph.
• Dictionary matching further sharpens class-specific clusters and induces denser crossmodal bridges.
• In sarcasm (MUStARD), visual features dominate teaching, reflecting task-specific modality salience.

7. Broader Impact and Related Methodologies

DMD advances multimodal learning by providing granularity in information sharing and preserving, enabling both flexible, adaptive knowledge transfer and robust feature disentanglement. Unlike direct feature concatenation or non-adaptive knowledge distillation, DMD decouples modality contributions, regulates crossmodal transfer with graph-structured, sample-adaptive weights, and optionally enforces semantic crossmodal alignment at both coarse and fine granularity.

The core DMD concept underpins varied settings—emotion recognition, dataset distillation, audio-visual alignment—and has generalized through hierarchical (DHMD) and sample-distribution joint matching (DAVDD) variants. These developments mark a technical progression over static, hand-designed fusion and transfer mechanisms, supporting more interpretable and effective multimodal integration (Li et al., 2023, Li et al., 22 Nov 2025, Li et al., 4 Feb 2026).