University of Toronto Researchers Demonstrate First Trainable Quantum Neural Network on 100-Qubit System
University of Toronto researchers demonstrated the first fully trainable quantum neural network on a 100-qubit system, solving the barren plateau problem with a new parameterization scheme and successfully classifying quantum states with no classical analog.
Researchers at the University of Toronto's Centre for Quantum Information and Quantum Control demonstrated this month the first fully trainable quantum neural network operating on a 100-qubit system. The network, which uses a novel architecture that avoids the barren plateau problem that has plagued previous quantum neural network attempts, successfully learned to classify quantum states and perform quantum data compression tasks that classical neural networks cannot efficiently represent.
The barren plateau problem has been the major obstacle to quantum neural network training. When quantum neural networks exceed a certain size, the gradient—the signal that guides learning—becomes exponentially small, making training effectively impossible. The Toronto team solved this problem by developing a new parameterization scheme for quantum neural networks that ensures gradients remain large enough to support learning even as the network scales.
Dr. Juan Miguel Arrazola, who led the research, explained the breakthrough: "For years, people assumed that quantum neural networks would suffer from barren plateaus—that they simply couldn't be trained at scale. What we've shown is that barren plateaus are a feature of certain architectures, not a fundamental limitation. By designing the architecture carefully, we can create quantum neural networks that train reliably even with 100 qubits."
The demonstration used Quantinuum's trapped-ion quantum computer, which provides the high gate fidelities and long coherence times needed for the deep circuits required for quantum neural networks. The team trained their network to perform two tasks: classifying quantum states that differ by subtle quantum correlations (a task that classical networks cannot perform efficiently), and compressing quantum data into more compact representations while preserving quantum information.
The quantum classification task is particularly significant because it demonstrates a capability that is impossible for classical neural networks. The network learned to distinguish between quantum states that have the same classical description but different quantum entanglement structures—a task that requires a quantum processor to even represent the inputs. This represents a genuine quantum advantage: a quantum neural network performing a task that has no classical analog.
The University of Toronto team has released their architecture specifications and training code under an open-source license, allowing other researchers to build on their work. They have also established a collaboration with IBM to implement the architecture on IBM's quantum hardware, which would make it accessible to a broader research community.
Potential applications for trainable quantum neural networks include quantum error correction, quantum state preparation, and quantum control—all tasks that currently require extensive manual tuning. If quantum neural networks can learn to perform these tasks automatically, it could accelerate progress in quantum computing by automating tasks that currently require expert human intervention.