| name | parameter-efficient-quantum-mtl |
| description | Parameter-efficient Quantum Multi-task Learning (QMTL) methodology. Replaces conventional task-specific linear heads with fully quantum prediction heads in hybrid architectures. Quantum head parameters scale linearly with task count vs quadratic for classical heads. Use when designing multi-task learning systems for medical imaging, NLP, or multimodal tasks with constrained parameter budgets. |
Parameter-Efficient Quantum Multi-task Learning
Description
Framework for parameter-efficient quantum multi-task learning (QMTL) that replaces conventional task-specific linear heads with fully quantum prediction heads in a hybrid architecture. Demonstrates that quantum head parameter cost scales linearly with task count, while classical head cost scales quadratically. Evaluated on medical imaging, NLP, and multimodal benchmarks (arXiv:2604.13560).
Activation Keywords
- quantum multi-task learning
- 量子多任务学习
- parameter-efficient quantum
- QMTL
- quantum prediction head
- quantum MTL
- variational quantum multi-task
Core Concepts
The Problem: Parameter Explosion in Multi-task Learning
In standard hard-parameter-sharing MTL:
- Shared backbone: processes input → shared representation
- Task-specific heads: map shared representation → task predictions
- Problem: Task-specific parameters grow quadratically O(d × T) with tasks T and representation dimension d
The QMTL Solution
Replace task-specific linear heads with a hybrid quantum-classical architecture:
- Shared VQC encoding: maps classical data to Hilbert space (task-independent)
- Task-specific ansatz blocks: lightweight quantum blocks for localized adaptation
- Linear scaling: quantum head parameters scale as O(T), not O(d × T)
