The increasing energy demands of artificial intelligence are driving a search for more efficient computing solutions. While many efforts focus on refining existing hardware and software, a more radical approach lies in the realm of quantum computing. Quantum hardware, in certain respects, aligns more naturally with the mathematical underpinnings of AI than traditional systems. Although current quantum technology is still prone to errors for complex AI models, researchers are actively developing the infrastructure needed to run AI algorithms on quantum processors. Recent work demonstrates progress in transferring classical image data into quantum processors for basic AI image classification, prompting a closer look at the potential of quantum AI. Just as AI encompasses various machine learning techniques, quantum computing offers multiple avenues for enhancing AI algorithms. Some applications leverage the inherent mathematical capabilities of quantum hardware, particularly in performing matrix operations crucial to certain machine learning models. These operations can be executed far more efficiently on quantum systems. The separation of processing and memory in traditional computers creates a bottleneck for AI tasks like neural networks, requiring frequent data retrieval. Quantum computers, however, largely eliminate this separation by housing data directly within qubits, enabling computations via direct operations on these qubits. The potential of quantum systems to outperform classical ones in supervised machine learning has already been demonstrated, even when processing data stored on classical hardware. This approach utilizes variational quantum circuits, involving two-qubit gate operations influenced by a classical factor. This setup mirrors the communication within neural networks, where the two-qubit gate operation represents information transfer between artificial neurons, and the classical factor corresponds to the weight assigned to the signal. A team from the Honda Research Institute, in collaboration with Blue Qubit, has been actively exploring this very system. The primary focus of this recent research was on efficiently transferring data from the classical world into the quantum system for processing. The researchers tested their methods on two different quantum processors, tackling an image classification problem using the Honda Scenes dataset, which contains images from approximately 80 hours of driving in Northern California, tagged with scene information. The specific task was to determine whether it was snowing in a given scene. Since the images resided on classical hardware, they needed to be converted into quantum information for processing on the quantum computers. The team explored three data encoding methods, varying in how the image pixels were divided and distributed across qubits. A classical simulator of a quantum processor was used for training, identifying optimal parameters for the two-qubit gate operations. The trained system was then tested on quantum processors from IBM (156 qubits, higher error rate) and Quantinuum (56 qubits, low error rate). Generally, classification accuracy improved with increased qubit usage and gate operations. While the system demonstrated functionality, achieving accuracy levels significantly above random chance, it still fell short of the performance of standard algorithms on classical hardware. Current quantum hardware still requires improvements in both qubit count and error rates to compete with classical systems. Nevertheless, this work provides tangible evidence that real-world quantum hardware can execute AI algorithms, albeit with the need for further hardware advancements before solving practical, real-world problems. Ultimately, the path to realizing the full potential of quantum AI hinges on continued progress in quantum hardware development. As qubit counts increase and error rates decrease, quantum computers will become increasingly competitive for a wider range of AI tasks, potentially revolutionizing fields that demand immense computational power and efficiency.
