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

This paper introduces QPredSGG, a hybrid quantum-classical framework that replaces the predicate head of a Causal Feature Enhancement Network with a parameter-efficient Quantum Predicate Head, achieving state-of-the-art performance on long-tail scene graph generation by significantly reducing model complexity while improving mean recall on the Visual Genome 150 dataset.

The Gist

Imagine you are looking at a busy photograph of a park. A computer trying to understand this image needs to do more than just say, "I see a person and a horse." It needs to understand the story: "The person is riding the horse."

This task is called Scene Graph Generation. The computer builds a map of the image where objects are dots and their relationships are lines connecting them.

The Problem: The "Popular Kid" Bias

The paper points out a major flaw in how computers currently learn to do this. They are trained on a massive dataset called Visual Genome. In this dataset, some relationships are super common (like "on," "of," or "in"), while others are very rare but specific (like "wearing," "chasing," or "painting on").

Think of it like a classroom where 90% of the students are named "John." If a teacher asks, "What is the most common name?" the students will almost always guess "John." Even if there is a student named "Zephyr" who is actually the most interesting one in the room, the students ignore them because "John" is so much more frequent.

In the computer world, this means the AI gets really good at guessing common relationships but fails miserably at the rare, specific ones. This is called the Long-Tail Problem.

The Solution: A Quantum "Specialist"

The authors of this paper, Prerana Ramkumar and her team, decided to try something new. Instead of using a giant, heavy computer brain to make the final decision about relationships, they replaced that part with a tiny, Hybrid Quantum Head.

Here is how they did it, using an analogy:

1. The Heavy Lifting (Classical Part): Imagine a very smart, traditional librarian (the "CFEN backbone") who reads the book and summarizes the story. This part stays the same. It takes the visual information and creates a long, detailed summary (4,096 numbers) about the relationship between two objects.
2. The Quantum Specialist (The New Part): Instead of giving that long summary to a giant, expensive decision-maker, they compress it down into a tiny, 16-number summary. They then feed this tiny summary into a Quantum Circuit.
   • Think of the Quantum Circuit as a magic filter or a specialist lens. It doesn't need to be huge to work. It uses the strange rules of quantum physics (like superposition and entanglement) to look at those 16 numbers and decide: "Is this 'riding' or 'wearing'?"
3. The Result: The specialist makes a guess, and a small classical computer checks it.

What They Tested

The researchers treated this like a science experiment to find the perfect "Quantum Specialist." They tested:

• How many "qubits" (quantum bits) to use: They tried 4 and 8.
• How to translate the data: They tried different ways to turn the numbers into quantum states (like "Angle Embedding" vs. "Amplitude Embedding").
• How complex the circuit should be: They tried circuits with different numbers of layers.

The Big Wins

Here is what they found, in plain English:

• Small is Beautiful: The best version used only 4 qubits (the smallest size they tested). It had only 96 trainable parameters. To put that in perspective, the rest of the computer model has millions of parameters. The quantum part is like a tiny, efficient chef in a massive kitchen, doing just the one job of deciding the relationship.
• Better at the Rare Stuff: When they trained the system to pay extra attention to the rare relationships (using a special "weighted" training method), the quantum head got much better at spotting the "Zephyrs" of the world.
  • The old, standard computer model got about 41% of the rare relationships right.
  • Their new 4-qubit quantum model got 57% right.
  • Even the 8-qubit version stayed strong at 55%.
• No Loss in the Common Stuff: While getting better at the rare stuff, the model didn't lose its ability to guess the common stuff (like "on" or "in"). It kept its global accuracy high.
• Real Hardware Test: They didn't just run this on a simulator; they actually ran a tiny version of it on a real quantum computer (an IBM superconducting chip). It worked! It didn't crash or give random answers. It correctly identified 6 out of 9 test cases, proving that this tiny quantum brain can actually run on real, noisy hardware.

The Trade-Off

The paper also noted a catch. If you make the quantum circuit too deep (add too many layers to make it "smarter"), it takes longer to run and uses more computing power. The "sweet spot" was a circuit that was deep enough to be smart, but shallow enough to be fast.

Summary

In short, this paper shows that you don't need a massive quantum computer to improve AI. By swapping out just the final decision-making step with a tiny, efficient quantum module, you can help the AI stop ignoring the rare, specific relationships in images. It's like replacing a loud, biased crowd with a quiet, highly trained specialist who listens to the details everyone else misses.

Technical Summary

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

1. Problem Statement

Scene Graph Generation (SGG) aims to represent images as structured graphs of objects (nodes) and their semantic relationships (edges), typically expressed as triplets $\langle \text{subject}, \text{predicate}, \text{object} \rangle$. A critical bottleneck in current SGG systems is the long-tailed distribution of predicates found in datasets like Visual Genome. Frequent, generic predicates (e.g., "on," "of," "in") dominate annotated relationships, while semantically specific predicates occur rarely.

Consequently, standard models trained with conventional objectives tend to optimize for Global Recall (R@K) by predicting frequent classes, resulting in poor Mean Recall (mR@K) for rare but informative relationships. While existing debiasing strategies (e.g., causal inference, loss reweighting) have improved mean recall, the predicate classification stage in state-of-the-art frameworks like the Causal Feature Enhancement Network (CFEN) still relies on large classical Multi-Layer Perceptrons (MLPs). These decision modules are parameter-heavy, raising the question of whether a more compact decision module could maintain or improve long-tail recognition performance.

2. Methodology

The paper proposes QPredSGG, a hybrid quantum-classical architecture that replaces the classical predicate head in the CFEN backbone with a Quantum Predicate Head (QP-Head). The methodology follows a four-stage experimental pipeline:

• Backbone and Feature Compression: The system utilizes the CFEN backbone, which employs a Bidirectional Tree LSTM (BiTreeLSTM) to extract contextual pair embeddings ($h_{ij} \in \mathbb{R}^{4096}$). Before quantum processing, these high-dimensional features are projected via a classical linear layer into a compressed, quantum-compatible vector (e.g., 16-dimensional for 4 qubits).
• Hybrid Quantum Architecture (QP-Head):
  • Encoding: The compressed features are encoded into a parameterized quantum circuit (PQC) using either Angle Embedding (mapping features to rotation angles) or Amplitude Embedding (normalizing and mapping to state amplitudes).
  • Variational Circuit: The encoded state passes through trainable layers consisting of rotation gates and entangling templates. The study evaluates two templates: Basic Entangling Layers (BEL) and Strongly Entangling Layers (SEL).
  • Readout: The circuit terminates with measurements producing expectation values, which are fed into a lightweight classical readout layer to generate logits for the 51 predicate classes.
• Bias-Aware Training: To address the long-tail imbalance, all models are trained using Weighted Cross-Entropy (WCE) loss. Inverse-frequency weights are applied, capping the ratio of rare-to-frequent class penalties at 46$\times$, ensuring the optimizer prioritizes rare predicates without destabilizing training.
• Evaluation Metrics: Performance is assessed via Global Recall (R@50, R@100) and Mean Recall (mR@50, mR@100). Additionally, quantum-specific metrics are analyzed, including Expressibility (KL divergence from Haar-random fidelity) and Entanglement (Von Neumann entropy).

3. Key Contributions

The paper outlines five primary contributions:

1. Hybrid Quantum Predicate Head: Introduction of the QP-Head, a compact parameterized quantum circuit replacing the classical MLP in an SGG pipeline, preserving the relational feature backbone.
2. Controlled Architecture Study: A systematic evaluation of the QP-Head across qubit counts (4 vs. 8), encoding strategies (Angle vs. Amplitude), entangling templates (BEL vs. SEL), and circuit depths (2, 4, 6 layers).
3. Bias-Aware Evaluation: Analysis of the QP-Head under class-balanced training to determine if it improves rare-predicate recognition (mR) rather than just frequent-class performance.
4. Quantum Quality and Efficiency Analysis: Correlation of semantic performance with circuit-level diagnostics (expressibility, entanglement) and computational overhead (parameter count, runtime).
5. Physical QPU Validation: Execution of the compact 4-qubit QP-Head on a real superconducting quantum processor (IBM ibm_fez) to verify feasibility beyond state-vector simulation.

4. Experimental Results

Experiments were conducted on the Visual Genome 150 (VG-150) dataset under the Predicate Classification (PredCls) setting.

• Training Dynamics: Class-balanced training (WCE) significantly improved mR@50 (from ~0.17 to ~0.26) compared to standard Cross-Entropy, without degrading Global Recall.
• 4-Qubit Search: Among 4-qubit configurations, Amplitude Embedding with Strongly Entangling Layers yielded the best performance, achieving an mR@100 of 57.25% (compared to 41.1% for the classical CFEN reference) with only 96 trainable quantum parameters. This configuration utilized a 16-dimensional compressed representation.
• Scaling to 8 Qubits: Scaling to 8 qubits (256-dimensional state space) maintained strong performance, reaching an mR@100 of 55.38% with 384 quantum parameters. Global recall remained stable (R@100 > 0.90).
• Depth Ablation: Increasing circuit depth from 2 to 6 layers improved expressibility (lower KL divergence) but increased runtime latency significantly (from ~214ms to ~474ms). The 4-layer configuration offered the best trade-off between expressibility and computational cost.
• Parameter Efficiency: The quantum component represented less than 0.001% of the total model parameters, acting as a compact decision layer atop the classical feature extractor.
• Hardware Execution: On the IBM ibm_fez QPU, the 4-qubit model successfully processed 9 validation triplets, achieving a 66.67% batch accuracy. Crucially, the output did not collapse to a single dominant class, preserving class-discriminative structure despite hardware noise.
• Comparison: The QP-Head variants outperformed the classical CFEN reference in mR@100 (57.25% vs. 41.1%) while maintaining competitive Global Recall, using a fraction of the trainable parameters required by the classical head.

5. Significance and Claims

The paper modestly claims that compact hybrid quantum predicate heads can support parameter-efficient long-tail relational classification in complex visual reasoning tasks.

• Not a Claim of Unconditional Advantage: The authors do not claim broad quantum superiority. Instead, they provide controlled evidence that a small, NISQ-era quantum circuit can serve as an effective decision module when integrated into an established classical pipeline.
• Feasibility: The work demonstrates that hybrid quantum models can be trained on simulated environments and executed on physical hardware without collapsing to random or single-class behavior, even with severe dimensional compression.
• Practicality: The results suggest that quantum components can improve mean recall for rare predicates without introducing prohibitive parameter overhead, provided the architecture (encoding and entanglement) is carefully tuned.

The study concludes that while the current evaluation is limited to PredCls and simulation, the QP-Head represents a promising direction for integrating hybrid quantum components into scene graph generation to address long-tail bias efficiently.