Quantum Models with Multi-Stage Training for Compositional Concept Generalization

Human intelligence possesses a remarkable ability to recombine simple, learned ideas into entirely new situations. A person who understands what a "car" is and what "yellow" means can instantly recognize a yellow car they have never seen before, applying their knowledge of danger to this novel combination. This capacity, known as compositional generalization, allows us to navigate a world where we constantly encounter fresh arrangements of familiar parts. For artificial intelligence, however, this skill remains a stubborn hurdle. While modern systems excel at recognizing patterns they have seen during training, they often stumble when asked to apply those same patterns to new combinations, such as a specific object placed in a new spatial relationship. This limitation suggests that current models are memorizing examples rather than truly understanding the underlying rules of how things fit together.

Researchers at University College London and the University of Amsterdam have explored a different path to solving this problem, turning to the strange and powerful logic of quantum mechanics. In their work, presented for the IEEE Quantum Week conference, they propose that the way quantum computers process information might naturally align with how humans combine concepts. By building a model that separates the learning of objects from the learning of relationships between them, they created a system that can generalize to unseen scenarios far more effectively than traditional methods, all while using a fraction of the computational power.

The team focused on a task where a computer must look at an image of two geometric shapes and choose the correct description of their relationship from two options. For instance, given an image of a cube to the left of a cone, the system must distinguish the correct phrase "cube left cone" from a distractor like "cube right cone." To teach the machine, the researchers used a dataset containing images of shapes like spheres, cylinders, and cones in various positions. The challenge was to train the system on some combinations, such as a cone to the right of a cube, and then test it on combinations it had never seen, like a cone to the left of a cube. This setup forces the model to learn the concept of "left" and "right" as independent rules that can be applied to any shape, rather than just memorizing specific pictures.

Instead of training the entire system at once, the researchers adopted a two-stage approach that mimics a curriculum. In the first stage, the model learned to recognize individual objects. It was shown images of single shapes, like a lone sphere or a single cube, and taught to match them with their names. Once the model had mastered these basic building blocks, the researchers froze its memory of what those shapes looked like. In the second stage, they introduced the relational task. The model was now shown images with two shapes and had to learn only the rules of how they relate to each other, using the stable, pre-learned definitions of the shapes it had already memorized. This design ensured that the system did not try to relearn what a "cube" was every time it saw a new relationship, forcing it to treat the relationship as a separate transformation applied to a fixed object.

To implement this, the team used a framework that translates language and meaning into mathematical structures called tensors, which can be mapped directly onto quantum circuits. They tested different ways of feeding image data into these quantum circuits. One method, called amplitude encoding, tried to preserve the exact geometry of the original image data. Another, called angle encoding, transformed the data in a way that introduced non-linear changes, effectively reshaping the information to make the differences between relationships clearer. They also experimented with a "collage" method, where the quantum circuits for two separate shapes were physically combined to represent a single relational image.

The results of their simulations were revealing. When the researchers compared their multi-stage quantum model to a standard classical system, the difference in efficiency was stark. The classical baseline required over 150 million trainable parameters to attempt the task, yet it failed to generalize, scoring only 50 percent on unseen combinations—essentially guessing. In contrast, the quantum model achieved strong generalization using just 426 trainable parameters. The most successful version of their system, which used the collage method combined with angle encoding, reached an accuracy of over 72 percent on the unseen test cases. This performance was significantly better than a single-stage quantum model that tried to learn everything at once, proving that separating the learning of objects from the learning of relationships was crucial.

The researchers found that the success of their model depended heavily on how the data was encoded. The angle encoding method, which introduced non-linear transformations, proved superior to the amplitude encoding method. This suggests that the original image data, as processed by standard tools, did not clearly separate the spatial relationships needed for the task. The quantum encoding process itself helped restructure the information, making the distinction between "left" and "right" more visible to the model. While the amplitude encoding preserved the original shape of the data, it kept the confusing similarities between different spatial arrangements intact, making them harder to distinguish.

These findings, derived entirely from computer simulations, suggest that structured quantum representations offer a promising avenue for teaching machines to understand compositionality. By enforcing a clear separation between what things are and how they relate, and by using quantum circuits to reshape the data into a more separable form, the model achieved a level of generalization that classical systems struggled to reach with vastly more resources. The work does not claim to have solved the problem of artificial intelligence, but it demonstrates that a specific, structured approach to learning—where primitives are learned first and then combined—can be highly effective when paired with the unique properties of quantum computing. The authors note that future work will need to test these ideas on actual quantum hardware and explore whether these methods can scale to more complex scenes and richer vocabularies, but the current results provide a compelling proof of concept for a new way of thinking about machine learning.