TLDR: This research explores using Variational Quantum Circuits (VQCs) to achieve compositional generalization in AI, a human-like ability to understand new situations from known components. By interpreting tensor-based compositional models in Hilbert spaces and training VQCs on an image captioning task, the study shows that quantum models, particularly with multi-hot encodings, outperform classical compositional models in generalizing to unseen data, despite current limitations compared to large pre-trained classical models like CLIP.
Compositional generalization, the remarkable human ability to understand and react to new situations by applying knowledge from previously encountered ones, remains a significant challenge for modern artificial intelligence systems, including advanced vision-language models. Imagine seeing a blue car and a red postbox, and then effortlessly understanding what a red car is, even if you’ve never seen one before. This is the essence of compositional generalization, and it’s a capability current AI often struggles with.
Previous attempts to tackle this problem using classical tensor-based sentence semantics have yielded limited success. However, a new research paper, “Compositional Concept Generalization with Variational Quantum Circuits”, explores a novel approach: leveraging the increased training efficiency of quantum models to improve performance in these complex tasks.
The Quantum Leap for Compositional AI
The core idea behind this research is to interpret the representations of compositional tensor-based models within Hilbert spaces, which are fundamental to quantum mechanics. By doing so, Variational Quantum Circuits (VQCs) can be trained to learn these representations. The researchers applied this concept to an image captioning task that specifically requires compositional generalization, aiming to see if quantum computing could offer a more effective solution.
The study builds upon the Distributional Compositional Categorical semantic model (DisCoCat), a framework that explicitly models language composition by mapping grammatical structure to meanings encoded in vectors and higher-order tensors. While DisCoCat provides a theoretically sound way to model composition, learning and computing with its higher-order tensors on classical computers is computationally expensive. This is where quantum systems offer a significant advantage, as tensors are natural inhabitants of quantum architectures, potentially making their parameters easier to learn and computations less costly.
Inspired by Categorical Quantum Mechanics, DisCoCat has a growing ecosystem of tools that enable its implementation on quantum architectures like VQCs. Previous work has shown VQCs to be efficient for linguistic tasks such as text classification and question answering. This paper extends these methods to multimodal cognitive tasks, hypothesizing that quantum computing’s efficiency will enhance DisCoCat tensors’ training and improve compositional generalization.
Experimental Approach and Findings
To test their hypothesis, the researchers used a spatial visual question answering task, where the system had to identify the spatial relationship between objects in an image. They utilized a dataset consisting of images with two geometric shapes (cube, sphere, cylinder, cone) and captions describing their spatial relations (e.g., ‘cube left sphere’). The task was to match the correct caption to an image, even for unseen combinations of shapes and relations.
Two main image encoding techniques were employed for the quantum models:
- Multi-Hot Encodings (MHE): This method converts image information into a binary vector, focusing on essential data like shape identities and their relative positions. It served as a proof-of-concept for the quantum model’s ability to learn these fundamental relationships.
- CLIP Encodings: Using image vectors from OpenAI’s Transformer-based vision-language model CLIP, which are high-dimensional and capture rich image data. These were reduced in dimension using Principal Component Analysis (PCA) and loaded into quantum circuits using angle and amplitude encoding techniques.
A matching score, based on the inner product of the quantum circuit outputs for images and captions, was used for training. The quantum models were compared against classical DisCoCat implementations and the CLIP model itself.
The results were promising. Quantum models, particularly those using noisy MHE encodings, achieved good proof-of-concept results, outperforming classical compositional models. For instance, Quantum-MHE with noise achieved a 64.06% test accuracy, significantly better than Classical-DisCoCat with MHE, which only managed 30.63% test accuracy. This suggests that quantum models are less prone to overfitting on the training data, a common issue with classical DisCoCat.
While performance on CLIP image vectors was more mixed, quantum models still outperformed classical DisCoCat trained with CLIP vectors, which showed severe overfitting and 0% accuracy on the test set. Although the pre-trained CLIP model itself performed strongly (62.5% test accuracy, improving to 70% after fine-tuning), it has a substantial advantage in terms of pretraining data and model size (tens of millions of parameters compared to hundreds for the quantum models).
Interestingly, the quantum models struggled with recognizing certain shapes, like ‘sphere’, leading to reduced performance when these shapes were involved. This highlights the need for further analysis into training methods and encoding types.
Also Read:
- Quantum Circuits Uncover Optimal Settings for Industrial Systems
- Learning Quantum Environments from Classical Data for Policy Evaluation on Quantum Hardware
Future Outlook
The research concludes that while quantum methods for natural language representations are still in their early stages, they consistently outperform classically trained compositional models, demonstrating a greater ability to generalize to out-of-distribution inputs. The choice of implementation, including encoding and circuit types, significantly impacts performance, indicating fertile ground for future research. This work represents a significant step towards building AI systems that can achieve human-like compositional generalization, potentially unlocking new capabilities in understanding and interacting with the world.
