Our collaboration with Honda Research Institute (HRI) demonstrates quantum image classification at scale, pushing the boundaries of what's possible with current quantum hardware.
Our quantum image classification pipeline transforms classical images into quantum states, processes them with variational circuits, and produces classifications through selective measurements.
The process starts by encoding Honda's image data using quantum gates (green), transforms this quantum information through learnable circuits (blue), and extracts classification predictions through strategic measurements (purple). We train this pipeline in two phases: first, optimizing the data encoding, then training the classification components.
This architecture enables us to process complex image data while keeping circuits practical for today's quantum computers.
We developed and implemented three approaches for encoding image data into quantum states:
Divides the image into independent blocks and encodes each onto separate sets of qubits, trading higher qubit count for better accuracy.Our best-performing method partitions images into blocks, each encoded independently. This approach achieved 94% accuracy using 224 qubits across 16 blocks.
This method encodes the full image by gradually incorporating qubits, training from most to least significant bits to mitigate vanishing gradients.We encoded full images using a hierarchical circuit structure using 19 qubits, achieving 81% accuracy. This method proved effective for smaller quantum systems.
Directly encodes image data into qubit rotation angles, interleaving data loading with processing layers.This rotation-based encoding demonstrated 88% accuracy using just 20 qubits, offering an efficient alternative for limited qubit counts.
For each encoding method, we design quantum circuits that transform the input data through trainable gates and measurements:
The image shows the hierarchical learning circuit used in AAE. Starting with two qubits, we train variational parameters (shown in blue boxes) while gradually adding new qubits initialized in |+⟩ state through Hadamard gates (H). Red lines indicate two-qubit entangling operations. The bottom diagram shows how active qubits and connectivity expand through the training stages. You can find more details in our post on quantum data loading.
Key achievements:
Testing on H1 (20 qubits) and H2 (56 qubits) systems showed:
We tested three circuit depths on Quantinuum's H2 system (244, 376, and 664 RZZ gates). The comparison shows strong alignment between simulation (blue) and hardware (red) results, with all circuits maintaining correct classification between clear and snowy conditions.
We quantified hardware performance using L1 distance from the ideal simulation on IBM's processors. Results showed meaningful signal retention up to 500 RZZs, staying well below the theoretical noise threshold (red line) and 99th percentile bound (yellow line).
Using Brisbane (127 qubits) and Fez (156 qubits) processors, we demonstrated:
IBM's 156-qubit Fez processor uses a heavy-hex layout, shown with qubit readout errors (darker circles indicate lower error) and two-qubit gate fidelities (edge colors). We leverage this topology to map our 12-qubit blocks using the ring connectivity pattern.
Our best results compared favorably with classical methods:
This research enables:
Using Honda's Scenes Dataset, we demonstrated quantum classification of road conditions between clear and snowy environments. This capability has direct applications in autonomous driving systems, where rapid and accurate scene understanding is crucial for safety. While current quantum computers don't yet outperform classical methods, our results show that complex visual tasks can be executed on quantum hardware, laying the groundwork for future advantages as quantum systems scale. The methods we developed for handling high-dimensional visual data could extend beyond automotive applications to robotics, surveillance systems, and other computer vision tasks requiring real-time environmental assessment.
For those interested in further details:
We're focusing on:
Contact our team to learn more about implementing quantum solutions for machine learning-based tasks.