Insider Brief
- A new theoretical quantum machine-learning framework proposes quantum-circuit designs that aim to make large quantum neural networks easier to train while preserving computational tasks that remain difficult for classical computers.
- The framework introduces two scalable circuit architectures and a parallel training method that mathematically avoids exponential “barren plateaus” while reducing the number of quantum-circuit evaluations required for gradient calculations by roughly a factor proportional to the number of qubits.
- The work remains a theoretical proposal without large-scale hardware validation or evidence of outperforming leading classical machine-learning models, leaving open whether its computational advantages translate into practical machine-learning performance.
A proposed quantum machine-learning framework could make larger quantum neural networks easier to train while preserving computations that are difficult for conventional computers to reproduce.
The study, posted on the arXiv preprint server, presents two quantum-circuit designs intended to address several problems that have limited efforts to use quantum computers for machine learning. These include vanishing training signals, the high number of quantum-circuit runs needed to adjust a model and the risk that an easily trained quantum model may also be easy to simulate on a classical computer.
Iordanis Kerenidis, a researcher at IRIF, CNRS and Université Paris Cité and a co-founder of Quantum Signals SAS, developed what the study describes as an end-to-end framework for scalable quantum deep learning.
If the approach proves successful on future quantum hardware, it could help machine-learning systems tackle highly complex problems in areas such as financial risk modeling, portfolio optimization, materials discovery, drug design, and logistics that are difficult to solve efficiently with classical computers.
However, it’s important to recognize that the proposal remains theoretical and the study does not make the claim that the circuits outperform leading classical machine-learning models on practical tasks, nor does it report large-scale tests on quantum hardware. Instead, it however does provide mathematical arguments that the proposed models can be trained without one of the field’s most serious optimization problems while reducing the number of circuit evaluations required during training.
The work also identifies settings in which reproducing the model’s output with a classical computer would be difficult under standard assumptions or beyond the reach of the best currently known algorithms.
That combination is important because trainability and computational power have often worked against each other in quantum machine learning. Adding enough complexity to make a quantum model hard to simulate can also make it nearly impossible to train.
Balancing Trainability and Computational Power
Quantum neural networks use adjustable quantum gates in a role somewhat similar to the adjustable weights in a conventional neural network. A classical computer typically controls the training process, while a quantum processor runs circuits and returns measurements used to update the model.
One major obstacle is known as a barren plateau. In these regions, the mathematical signals used to determine how model parameters should change become extremely small as a circuit grows. Training then becomes impractical because the optimizer cannot reliably tell which changes will improve the model.
Researchers have designed structured quantum circuits that avoid exponentially vanishing training signals. The drawback is that some of those circuits are also simple enough for classical computers to simulate efficiently, reducing the likelihood of a meaningful quantum advantage.
The new study proposes circuits that preserve a fixed number of particle-like excitations as information moves between qubits. The design combines reconfigurable beam-splitter gates, which move an excitation between two qubits without creating or destroying it, with single-qubit phase gates.
According to the study, these phase gates play a central role by enlarging the range of transformations the circuits can perform and prevent the gradients — which are essentially the mathematical signals that tell a machine-learning model how to adjust itself to reduce errors — used in training from shrinking exponentially.
The framework uses two circuit layouts. The first, called a unitary brick-wall circuit, arranges gates in a staggered pattern across neighboring qubits. It is designed for processors in which qubits mainly interact with adjacent qubits, as is common in several hardware platforms.
The second, called a unitary butterfly circuit, connects qubits in a pattern similar to one used in fast Fourier transform algorithms. It assumes hardware with long-range or reconfigurable connections between qubits.
The butterfly design is much shallower. Its depth grows only with the logarithm of the number of qubits, meaning that doubling the size of a large circuit adds relatively few layers. The brick-wall circuit has a depth that grows in direct proportion to the number of qubits but does not require long-range gates.
The study provides an unconditional mathematical result showing that the brick-wall circuit avoids exponential barren plateaus for the measurements considered. For the butterfly circuit, the researcher establishes an unconditional polynomial lower bound on the training signal, which also rules out an exponential barren plateau.
A more precise prediction for the butterfly circuit’s gradient size depends on an unproven mathematical conjecture about how effectively the circuit reproduces certain properties of random quantum operations.
Cutting the Cost of Training
Even a trainable quantum circuit can be impractical if calculating its gradients requires too many trips to a quantum processor.
A commonly used method called the parameter-shift rule evaluates a quantum circuit at slightly different parameter settings to calculate how each gate affects the result. The method can calculate exact gradients, but its cost rises with the number of trainable parameters.
For circuits with thousands or millions of parameters, a separate set of circuit evaluations for each parameter would create a major bottleneck.
The study introduces a multi-layer parallel parameter-shift rule that calculates the effects of many parameters together. It takes advantage of circuit layers containing gates that act on separate pairs of qubits and therefore do not interfere with one another during a single layer.
The method uses random patterns of positive and negative parameter changes to estimate all gradients in a layer at once. The estimator is unbiased, meaning that repeated estimates converge on the exact gradient rather than systematically pushing the model in the wrong direction.
For both circuit layouts, the study reports a reduction in circuit evaluations by a factor equal to the number of qubits divided by twice the number of particle-like excitations used in the model.
In one example involving 1,024 qubits and 10 excitations, the butterfly circuit would require 600 circuit evaluations for a gradient step, compared with about 30,720 under the standard approach. That represents a reduction of roughly 51 times.
The brick-wall circuit at the same size would require about 61,400 evaluations instead of more than 3.1 million, also a reduction of about 51 times.
The absolute cost remains much higher for the brick-wall design because its nearest-neighbor structure requires more layers. The difference illustrates a hardware trade-off rather than a claim that one architecture is universally superior.
The butterfly circuit offers lower depth and fewer circuit evaluations but requires flexible connections between distant qubits. The brick-wall circuit is deeper and more expensive to train but is better suited to processors with local connections.
Encoding Data for Harder Computations
The framework also proposes a method for loading classical data into the quantum circuit.
It begins with specially prepared quantum states known as non-Gaussian magic states. In this context, the term refers to states with a structure that cannot be described by a simpler class of efficiently simulated fermionic states.
The circuit then spreads a small number of excitations across all available qubits. Classical data are encoded through phase rotations applied after this spreading process. Each feature changes the phase of one qubit, allowing all features to influence the entangled quantum state.
The study presents two versions of the input state for different machine-learning tasks.
For applications that use the circuit to produce samples, such as generative modeling or selecting actions in reinforcement learning, the proposal uses paired groups of excitations. The number of excitations grows logarithmically with the number of qubits.
Under standard complexity assumptions, the resulting output distribution would be difficult for a classical computer to sample accurately on average, according to the study.
For supervised tasks such as classification and regression, the framework instead uses groups of three excitations and measures correlations between pairs of qubits. In this setting, the number of excitations grows with the square of the logarithm of the number of qubits.
The study indicates that calculating those two-qubit correlations would require superpolynomial classical time under the best-known simulation methods.
That second claim is more limited than a proof that no efficient classical algorithm exists. It reflects the performance of currently known algorithms and could be weakened by a future classical simulation method.
This is important because quantum-machine-learning proposals can appear difficult to simulate until researchers discover a classical shortcut that exploits their mathematical structure.
The study also acknowledges that claims of classical hardness can change as new algorithms are discovered. It points to a recent example in which researchers found a mathematical shortcut using a structure called a Pfaffian — a specialized calculation used in some quantum simulations — to efficiently estimate results that had previously been thought much harder for classical computers to reproduce.
The new framework is designed to avoid that shortcut. For supervised machine-learning tasks, it uses a different type of input state whose simulation would instead require a more complex mathematical object called a hyper-Pfaffian. No efficient classical algorithm is currently known for performing that calculation, so the study argues these models remain difficult to simulate using today’s best techniques.
The researchers also caution that this is not a proof that an efficient classical algorithm can never be found. Rather, the claim reflects the current state of knowledge that, based on today’s best classical methods, these quantum models appear computationally difficult to reproduce, but future advances in classical algorithms could change that assessment.
Open Questions
The framework could support classification, regression, quantum kernel methods, generative models and reinforcement-learning policies, according to the study. A classical model could also be attached to the quantum circuit’s measured outputs, creating a hybrid system trained across both quantum and classical components.
The work further states that the large quantum state space does not automatically require an equally large training set. The number of samples needed for the model to generalize should remain polynomial in the number of trainable parameters, rather than growing with the full dimension of the quantum state space.
There still remains open questions on whether the functions that are difficult for classical computers to reproduce are useful for real machine-learning problems.
Computational hardness alone does not establish that a quantum model will classify images more accurately, forecast data more effectively or learn better control policies than a classical system. A model can be difficult to simulate without producing useful results.
Testing that question will require quantum processors that can run hundreds of qubits with sufficient control, connectivity and measurement accuracy. The study identifies systems with about 256 or 512 qubits as possible scales for meaningful comparisons, depending on the circuit architecture and hardware connections.
Because noise is another practical issue, the paper focuses primarily on ideal circuit behavior and mathematical scaling. Real processors suffer from gate errors, decoherence and measurement noise, all of which could weaken gradients and increase the number of circuit runs needed to obtain reliable estimates.
Further work could examine whether training one layer or block at a time converges more efficiently than updating the full circuit. Alternative input states, including highly symmetric Dicke states, may also offer different trade-offs between preparation depth, data structure and resistance to classical simulation.
This is a highly technical paper — and this story only scratches the surface of the study. For a deeper, more technical dive, please review the paper on arXiv. It’s also important to note that arXiv is a pre-print server, which allows researchers to receive quick feedback on their work. However, it should not — nor should this article, itself — be considered official peer-review publications. Peer-review is an important step in the scientific process to verify results.
