parameter-efficient-quantum-mtl

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:

1. Shared VQC encoding: maps classical data to Hilbert space (task-independent)
2. Task-specific ansatz blocks: lightweight quantum blocks for localized adaptation
3. Linear scaling: quantum head parameters scale as O(T), not O(d × T)

Architecture

Input → Classical/Shared Backbone → Shared VQC Encoding
                                        ├── Task 1 Ansatz Block → Prediction 1
                                        ├── Task 2 Ansatz Block → Prediction 2
                                        └── Task T Ansatz Block → Prediction T

Implementation Workflow

Step 1: Shared Representation

- Classical backbone (e.g., ResNet for images, Transformer for text)
- OR shared VQC encoding stage for quantum-native representation
- Output: shared feature vector of dimension d

Step 2: Task-Specific Quantum Heads

For each task t:
  - Lightweight ansatz block (2-3 parameterized gates)
  - Maps shared representation → task-specific output
  - Parameters per task: O(1), not O(d)

Step 3: Training

- Joint loss: L = Σ_t L_t(prediction_t, label_t)
- Backprop through shared backbone + task ansatz blocks
- Quantum simulation or real hardware execution

Usage Patterns

Pattern 1: Medical Multi-task Diagnosis

Input: Medical image (e.g., fundus photo)
Tasks: 
  - Disease classification (diabetic retinopathy grade)
  - Severity assessment
  - Risk prediction
Architecture: Shared CNN → Shared VQC → 3 task-specific ansatz blocks

Pattern 2: NLP Multi-task

Input: Text sequence
Tasks:
  - Sentiment analysis
  - Named entity recognition
  - Text classification
Architecture: Shared Transformer → Shared VQC → task ansatz blocks

Pattern 3: Parameter-Constrained Edge Deployment

When classical MTL exceeds memory/compute budget:
  - Replace heavy task heads with lightweight quantum heads
  - Maintain shared representation quality
  - Quantum simulation provides compact parameterization

Parameter Scaling Analysis

Architecture	Parameters vs Tasks	Space Complexity
Classical MTL head	O(d × T)	Quadratic
QMTL (proposed)	O(T)	Linear
Hybrid quantum MTL (prior)	O(d × T)	Quadratic

Error Handling

Quantum Simulation Bottleneck

• Limit qubit count to 4-8 for practical simulation
• Use statevector simulator for training, shot-based for evaluation
• Consider hardware execution for final deployment

Task Interference

• If tasks conflict, increase ansatz block expressibility
• Monitor per-task loss curves for negative transfer
• Consider task grouping before quantum head assignment

Gradient Flow Through Quantum Layers

• Use parameter-shift rule for gradient computation
• Monitor barren plateau indicators
• Shallow circuits (2-3 layers) recommended for NISQ era

Resources

• arXiv:2604.13560 — Parameter-efficient Quantum Multi-task Learning
• PennyLane for VQC implementation
• Medical imaging benchmarks (e.g., BreastMNIST, PathMNIST)
• NLP benchmarks for multi-task evaluation