But quantum gates introduce noise, which can hinder the performance of a quantum machine.

Researchers at MIT and elsewhere are working to overcome this problem by developing a technique that makes the quantum circuit itself resistant to noise. (Specifically, these are "parameterized" quantum circuits containing tunable quantum gates.) The team created a framework that can identify the most robust quantum circuit for a given computing task and generate a mapping pattern that fits the qubits of a specific quantum device.

Their framework, called QuantumNAS (noise-adaptive search), is much less computationally intensive than other search methods and can identify quantum circuits that improve the accuracy of machine learning and quantum chemistry tasks. When researchers used their technique to identify quantum circuits for real quantum devices, their circuits outperformed those generated with other methods.

"The key idea here is that, without this technique, we have to sample every single quantum circuit architecture and every mapping scenario in the design space, train them, evaluate them, and if it's not good, we have to discard it and start over. But with this method, we can obtain many different circuits and mapping strategies at once without needing to train them many times," says Song Han, associate professor in the Department of Electrical and Computer Engineering (EECS) and lead author of the paper.

Along with Han, the lead authors of the work are Hanrui Wang and Yujun Lin, both graduate students at EECS; Yongshan Ding, assistant professor of computer science at Yale University; David Z. Pan, holder of the Silicon Labs electrical engineering chair at the University of Texas at Austin, and Jiaqi Gu, a graduate student at the same university; Fred Chong, Seymour Goodman Professor in the Department of Computer Science at the University of Chicago; and Zirui Li, an undergraduate student at Shanghai Jiao Tong University. The research will be presented at the IEEE International Symposium on High-Performance Computer Architecture.

Many design options

Building a parameterized quantum circuit involves selecting a number of quantum gates, which are the physical operations that the qubits will perform. This is not an easy task, as there are many types of gates to choose from. Furthermore, a circuit can have any number of gates, and the positions of those gates—which physical qubits they correspond to—can vary.

"With so many different options, the design space is extremely large. The challenge is how to design a good circuit architecture. With QuantumNAS, we want to design that architecture to be very robust against noise," says Wang.

The researchers focused on variational quantum circuits, which use quantum gates with trainable parameters that can learn a machine learning or quantum chemistry task. To design a variational quantum circuit, a researcher typically has to design the circuit by hand or use rule-based methods to design the circuit for a specific task, and then try to find the ideal set of parameters for each quantum gate through an optimization process.

In the naive search method, where potential circuits are evaluated individually, the parameters of each candidate quantum circuit must be trained, which represents a huge computational overhead. Furthermore, the researcher must first identify the ideal number of parameters and the circuit architecture.

In classical neural networks, including more parameters usually increases the model's accuracy. But in variational quantum computing, more parameters require more quantum gates, which introduce more noise.

With QuantumNAS, researchers are trying to reduce the overall cost of searching and training, while identifying the quantum circuit that contains the ideal number of parameters and the appropriate architecture to maximize accuracy and minimize noise.

To build a "supercircuit
," they first design a "SuperCircuit," which contains all possible parameterized quantum gates in the design space. This SuperCircuit will be used to generate smaller quantum circuits that can be tested.

They train the SuperCircuit once, and then, since all other candidate circuits in the design space are subsets of the SuperCircuit, they inherit the corresponding parameters that have already been trained. This reduces the computational load of the process.

Once the SuperCircuit is trained, it is used to search for circuit architectures that meet a specific objective, in this case, high noise robustness. The process involves searching for quantum circuits and qubit mappings simultaneously using what is known as an evolutionary search algorithm.

This algorithm generates several quantum circuits and candidate qubit mappings, then evaluates their accuracy using a noise model or on a real machine. The results are fed back to the algorithm, which selects the best-performing parts and uses them to restart the process until the ideal candidates are found.

"We know that different qubits have different properties and gate error rates. Since we're only going to use a subset of qubits, why not use the most reliable ones? We can do this by jointly searching for the architecture and mapping the qubits," Wang explains.

Once researchers have arrived at the best quantum circuit, they train its parameters and perform quantum gate pruning, removing gates with values ​​close to zero, as these don't contribute much to overall performance. Removing these gates reduces noise sources and further improves performance on real quantum machines. They then fine-tune the remaining parameters to recover the lost precision.

Once this step is completed, they can deploy the quantum circuit in a real machine.

When the researchers tested their circuits on real quantum devices, they outperformed all baselines, including those designed by hand and others created using different computational methods. In one experiment, they used QuantumNAS to produce a noise-resistant quantum circuit that was used to estimate the ground-state energy of a specific molecule, a significant step in quantum chemistry and drug discovery. Their method provided a more accurate estimate than any of the baselines.

Now that they have demonstrated the effectiveness of QuantumNAS, they want to use these principles to make the parameters of a quantum circuit robust to noise. The researchers also want to improve the scalability of a quantum neural network by training a quantum circuit on a real quantum machine, rather than a classical computer.

To encourage more work in this field, the researchers created an open-source library, called TorchQuantum, which contains information about their projects, tutorials, and tools that can be used by other research groups.

This work was supported by the National Science Foundation, the MIT-IBM Watson AI Lab, the Qualcomm Innovation Fellowship, and the U.S. Department of Energy.

###

Written by Adam Zewe, MIT News Office