qpredsgg-hybrid-quantum-predicate

name	qpredsgg-hybrid-quantum-predicate
category	quantum
description	Hybrid quantum predicate classifier for long-tailed scene graph generation. Replaces classical predicate head with QP-Head using amplitude embedding + strongly entangling layers.
arxiv	2606.04689
published	2026-06-03T00:00:00.000Z
categories	quant-ph, cs.LG
activation	qpredsgg, quantum-predicate, scene-graph, long-tail, hybrid-quantum, q-head, amplitude-embedding, strongly-entangling

QPredSGG: Hybrid Quantum Predicate Learning for Long-Tailed Scene Graph Generation

Overview

QPredSGG introduces a hybrid quantum predicate classifier for scene graph generation that replaces the classical predicate head in Causal Feature Enhancement Network (CFEN) with a Quantum Predicate Head (QP-Head). This achieves superior long-tail performance with dramatically fewer parameters.

Core Methodology

Architecture

• Classical backbone: CFEN extracts object features
• Quantum head: QP-Head replaces classical predicate classifier
• Feature compression: 4096D → 16D quantum-compatible (256× reduction)
• Training: Weighted cross-entropy loss for long-tail imbalance

Quantum Circuit Design

• Encoding: Amplitude Embedding for compact state preparation
• Layers: Strongly Entangling Layers for relational reasoning
• Measurement: Pauli-Z observables for predicate classification

Performance Results

Configuration	mR@100	Parameters	Notes
Classical CFEN	41.1%	-	Baseline
4-qubit QP-Head	57.25%	96	Best result
8-qubit QP-Head	55.38%	384	Strong long-tail

Key Findings

1. Parameter efficiency: 96 quantum params outperform classical reference
2. Compression: 256× feature reduction maintains relational accuracy
3. Depth trade-off: Expressibility vs runtime overhead analysis
4. Long-tail improvement: Significant gains on rare predicates

Implementation Patterns

Hybrid Quantum-Classical Pipeline

Object Features (4096D) 
    ↓
Amplitude Embedding
    ↓
Strongly Entangling Layers
    ↓
Pauli-Z Measurement
    ↓
Predicate Classification

When to Use

• Scene graph generation with long-tail predicate distributions
• Parameter-efficient relational reasoning
• Hybrid quantum-classical computer vision pipelines
• Visual reasoning tasks requiring fine-grained semantic classification

Pitfalls

• Qubit count doesn't monotonically improve performance (4 > 8 qubits)
• Runtime overhead increases with circuit depth
• Amplitude embedding requires careful normalization
• Strongly entangling layers may overfit on small predicate sets

Verification Steps

1. Validate amplitude embedding normalization (unit vector constraint)
2. Check entangling layer depth vs expressibility trade-off
3. Evaluate long-tail predicate recall specifically
4. Compare parameter efficiency against classical baselines

Research Directions

• Hybrid architectures for other relational reasoning tasks
• Quantum predicate heads for video scene understanding
• Integration with larger vision-language models
• Transfer learning across visual reasoning domains