bobazzouz

1. Research Vision

My work pioneers novel multimodal fusion architectures that enable cross-domain feature disentanglement—segregating domain-invariant representations from domain-specific noise while preserving semantic coherence. The framework addresses three fundamental challenges:

• Heterogeneous Modality Alignment: Harmonizing vision, language, and sensor data with divergent dimensionalities and temporal scales

• Domain-Agnostic Representation Learning: Isolating transferable features across domains (e.g., medical imaging → satellite data)

• Dynamic Fusion-Forgetting Equilibrium: Adaptive weighting of modality contributions via entropy-constrained attention

Key Insight: "Disentanglement through controlled interference"—strategically introducing cross-modal conflicts to force latent space factorization.

2. Theoretical Innovations

(A) Hypergraph Fusion Layers

• Modality-Aware Hyperedges: Dynamically reconfigurable hypergraphs to model N-way modality interactions (CVPR 2024 Oral)

• Topological Disentanglement Loss: Persistence homology-based constraints to separate features into contractible vs. non-contractible subspaces

(B) Adversarial Disentanglement Gates

• Domain-Contrastive Attention: Dual-path attention with gradient reversal to suppress domain-specific activations

• Quantum-Inspired Fusion: Qubit-like superposition states for probabilistic modality blending (collab. with CQT Singapore)

(C) Self-Supervised Disentanglement

• Cross-Modal Bootstrap Ping-Pong: Iterative refinement between modalities without labeled data (ICML 2025 Spotlight)

• Fractal Regularization: Multi-scale similarity preservation using Hausdorff distance metrics