Source-linked AI summary
The power of quantum neural networks
Amira Abbas, David Sutter, Christa Zoufal, Aurélien Lucchi, Alessio Figalli, Stefan Woerner
TL;DR
Quantum machine learning’s near-term benefits remain unclear, especially for quantum neural-network expressibility and trainability. The paper uses effective dimension and Fisher-information analysis to study these properties, finding higher capacity and faster training for well-designed quantum networks than comparable classical networks. It further links feature-map choice to barren-plateau resilience and supports the trainability result on real quantum hardware.
Problem
Near-term advantages of quantum machine learning remain unclear, motivating further study of quantum neural-network expressibility and trainability.
Method
The paper uses effective dimension and Fisher-information properties to define capacity, prove a generalisation bound, and assess trainability.
Results
Well-designed quantum neural networks achieve higher capacity and faster training than comparable classical feedforward neural networks.
Takeaways & Limitations
Feature-map choice is central: hard-to-simulate encoding is associated with high effective dimension, favourable Fisher spectra, and barren-plateau resilience.
Takeaways & Limitations
The paper calls for studies on multiple datasets, larger models, alternative variational circuits, and noise-induced barren plateaus.
Abstract
from arXiv · showhide
Fault-tolerant quantum computers offer the promise of dramatically improving machine learning through speed-ups in computation or improved model scalability. In the near-term, however, the benefits of quantum machine learning are not so clear. Understanding expressibility and trainability of quantum models-and quantum neural networks in particular-requires further investigation. In this work, we use tools from information geometry to define a notion of expressibility for quantum and classical models. The effective dimension, which depends on the Fisher information, is used to prove a novel generalisation bound and establish a robust measure of expressibility. We show that quantum neural networks are able to achieve a significantly better effective dimension than comparable classical neural networks. To then assess the trainability of quantum models, we connect the Fisher information spectrum to barren plateaus, the problem of vanishing gradients. Importantly, certain quantum neural networks can show resilience to this phenomenon and train faster than classical models due to their favourable optimisation landscapes, captured by a more evenly spread Fisher information spectrum. Our work is the first to demonstrate that well-designed quantum neural networks offer an advantage over classical neural networks through a higher effective dimension and faster training ability, which we verify on real quantum hardware.
1 Introduction
The paper addresses how to measure the expressive power and trainability of quantum neural networks, where conventional capacity measures are difficult to compute or yield weak bounds. It proposes effective dimension and Fisher-information analysis to compare quantum and classical models.
- Motivation: Quantum neural networks may offer speed-ups and faster processing, but their advantages over classical models remain insufficiently quantified.The paper identifies expressibility and trainability as open issues in quantum machine learning.
- Motivation: VC dimension is difficult to compute, relies on unrealistic assumptions, ignores data distribution, and often gives vacuous bounds for overparameterised networks.The cited limitations include assumptions such as access to infinite data and scaling with parameter count.
- Approach: Effective dimension incorporates the data distribution, depends on Fisher information, and is supported here by a novel generalisation bound as a capacity measure.Fisher information describes the geometry of a model’s parameter space.
- Trainability: Barren plateaus make parameter optimisation difficult because quantum-model loss landscapes can become perilously flat.The paper distinguishes noise-induced and circuit-induced barren plateaus.
- Study contribution: The study connects the Fisher-information spectrum to barren plateaus and compares quantum neural networks with classical feedforward networks on capacity and training.The analysis includes numerical experiments and support from an ibmq Montreal 27-qubit device.
2 Quantum neural networks
The study uses parameterised quantum circuits that encode classical data, apply trainable variational gates, and produce outputs through measurement. It benchmarks the quantum model against parameter-matched classical networks and varies the feature map to study its effects.
- Architecture: Quantum neural networks encode data into quantum states with a feature map, then optimise parameterised gates for a task.They are a subclass of variational quantum algorithms.
- Comparison: Faster training is defined as reaching lower training error within a fixed number of iterations for models with matching trainable-parameter and input-output sizes.These conditions define comparability between the models.
- Feature maps: Easy and hard feature maps refer to whether a classical computer can simulate the corresponding data-encoding strategy.The study uses this distinction to examine how encoding affects capacity and trainability.
- Architecture: The model encodes data with RZ and higher-order RZZ gates, applies RY gates with CNOT entangling layers, and classifies outputs by measured parity.Even parity corresponds to class 0 and odd parity to class 1.
- Comparison: The quantum model is benchmarked against fully connected classical feedforward networks across topologies with a fixed number of trainable parameters.The quantum feature map is also replaced with an easy-to-reproduce classical feature map.
3 Information geometry, effective dimension, and trainability of quantum neural networks
The paper uses information geometry to compare model capacity and generalisation across classical and quantum neural networks. Fisher information underpins the effective dimension, a practical capacity measure with a novel generalisation bound.
- Information geometry provides measures applicable to both classical and quantum models for studying capacity and trainability.
- 3.1 The Fisher information: The Fisher information quantifies parameter-space geometry and supports statistical interpretations of model capacity.
- 3.2 The effective dimension: The effective dimension estimates the model-space size occupied by a model using the Fisher information as its metric and data observations as resolution.
- 3.2 The effective dimension: The effective dimension integrates the Fisher information spectrum through its determinant and supports a bounded capacity interpretation.
- 3.3 Generalisation error bounds: Theorem 3.2 gives a generalisation-error bound based on effective dimension under continuity, full-rank, gradient, and loss regularity assumptions.
- 3.3 Generalisation error bounds: Numerical analysis found that effective dimension accurately captures generalisation behaviour under increasing label corruption.
4 Numerical experiments and results
The experiments compare classical feedforward networks with quantum models using Fisher-information spectra, effective dimension, and training performance. The hard-feature-map quantum neural network shows the most uniform spectrum, highest effective dimension, faster training, and resilience on real hardware.
- Experimental setup: The study compares Fisher spectra, effective dimension, and training performance across classical networks, an easy quantum model, and a quantum neural network.Models are matched by trainable-parameter count, input size, and output size.
- Fisher information spectrum: The quantum neural network has the most uniform Fisher-information spectrum, while classical models concentrate eigenvalues near zero with a few very large values.The easy quantum model also has many near-zero eigenvalues, but its large eigenvalues are less extreme than the classical model’s.
- Capacity analysis: The quantum neural network consistently achieves the highest effective dimension across the finite-data ranges studied.Its non-degenerate Fisher matrix supports faster convergence than the highly degenerate classical matrices and the easy quantum model.
- Training performance: Across 100 Iris trials and 100 training iterations, the quantum neural network reaches the lowest average training loss with the smallest spread.On the IBMQ Montreal device, its training advantage persists and performance is even better than in simulation.
- Capacity analysis: The quantum neural network’s average Fisher-Rao norm is roughly 250% higher than the classical neural network’s after fixed-iteration training.The ordering is quantum neural network, easy quantum model, then classical model.
- Trainability: As system size increases, the hard-feature-map quantum neural network retains a stable spectrum with many non-negligible eigenvalues, unlike the easy quantum model.This spectrum is associated with resilience to barren plateaus.
5 Conclusion
The conclusion presents effective dimension and Fisher-information spectra as tools for comparing quantum and classical model capacity and trainability. It attributes the observed advantages to the quantum model’s feature-map choice while identifying broader circuits, noise, datasets, and model sizes as open directions.
- Conclusion: The effective dimension is proposed as a robust capacity measure and supported by a novel generalisation-bound proof.The measure is applied to both quantum and classical models.
- Conclusion: Replacing the conjecturally hard-to-simulate feature map with an easily simulable one impairs the quantum model’s capacity and trainability advantages.The conclusion therefore emphasizes feature-map choice in quantum neural-network design.
- Conclusion: A Fisher spectrum whose eigenvalues do not vanish as qubit count grows indicates lower likelihood of a barren plateau.The hard-feature-map quantum neural network exhibits this resilience, with trainability supported by real-hardware results.
- Open directions: Future work should examine why the higher-order feature map produces its landscape, alternative variational circuits, non-unitary operations, noise-induced plateaus, and larger multi-dataset evaluations.The authors specifically identify generalisation on multiple datasets and larger models as needing further study.
- Conclusion: Quantum neural networks can have higher capacity and faster training than comparable classical feedforward networks when their Fisher spectrum is favourable.Capacity is measured by effective dimension, while trainability is assessed using Fisher-information properties.
B.1 Proof of Theorem 3.2
The proof establishes the effective-dimension generalisation bound by relating Fisher-information covering numbers to concentration of empirical risk. It also records improved scaling for relative-entropy loss and notes that stronger concentration could relax bounded-loss assumptions.
- Covering-number argument: The proof bounds the parameter-space covering number using the Fisher-information metric and then applies concentration inequalities to control generalisation error.The argument uses Hölder and Lipschitz continuity assumptions together with a union bound and Hoeffding’s inequality.
- Fisher-information geometry: The Fisher information matrix is diagonalised through its eigenvalues to estimate the number of metric balls needed to cover the parameter space.The comparison uses determinant monotonicity and bounds relating Fisher information at different parameter values.
- Generalisation bound: The proof combines the covering-number estimate with concentration bounds to derive the effective-dimension generalisation result.The final steps invoke Lemmas B.1 and B.2 after choosing the relevant covering radius.
- Refined scaling: Relative-entropy loss permits improved scaling because it is log-Lipschitz in its first argument.The improvement follows from the behaviour of the inverse of t|log(t)| near the origin.
- Assumptions: A stronger concentration inequality may relax the theorem’s assumption that the loss function is bounded.This is presented as a possible extension rather than an established result.
B.2 Generalisation ability of the effective dimension
This section tests whether effective dimension tracks generalisation and develops a discretised-model extension when Fisher information lacks full rank. The experiments and construction show how data resolution and regularity assumptions affect the measure.
- Generalisation experiment: Increasing label randomisation increases effective dimension until roughly 50% randomised labels, matching the greater expressive power needed to fit noise.The experiment trains a 6-input, 2-output feedforward network with 880 trainable weights on confusion sets.
- Resolution dependence: Effective dimension need not increase monotonically with data count because increasing n changes the resolution at which the model’s geometry is observed.A model can appear two-dimensional at low resolution, one-dimensional at intermediate resolution, and two-dimensional again at high resolution.
- Discretised model: The discretised model partitions parameter space into cubes and simplices, with an affine statistical model whose Fisher information is constant on each simplex.This construction supports generalisation analysis without requiring the original Fisher information to have full rank.
- Generalisation bound: The extension proves a generalisation bound for effective dimension without imposing a full-rank Fisher-information constraint.The theorem applies to a discretised version of the statistical model.
- Discretised model: The discretised model approximates the original generalisation bound when the discretisation parameter κ is much larger than the Fisher-information regularity scale Λ.The approximation relies on the additional condition ∥∇θ ˆF(θ)∥≤Λ.
C.1 Sensitivity analysis for the effective dimension
The sensitivity analysis examines how empirical Fisher-information sampling affects normalised effective dimension. Estimates stabilise with relatively few samples for lower-parameter networks.
- Sampling sensitivity: For networks with fewer trainable parameters, normalised effective-dimension estimates stabilise with as little as 40 data and parameter samples.The analysis varies both sample counts while keeping n fixed.
C.2 The Fisher information spectra for varying model size
The Fisher-information spectra reveal distinct scaling behaviour across classical and quantum models as input size and parameter count increase. The main quantum neural network maintains a more uniform spectrum than the classical and easy quantum models.
- Model-size comparison: The classical model concentrates most Fisher-information eigenvalues near zero while retaining a few very large eigenvalues across input sizes.This spectral concentration is contrasted with the more uniform quantum neural network spectrum.
- Model-size comparison: The easy quantum model has a somewhat uniform spectrum at smaller input size, but this uniformity deteriorates as input size and qubit count increase.Input size also determines the number of qubits in this model.
- Model-size comparison: The quantum neural network maintains a more uniform Fisher-information spectrum as input size and parameter count increase.The passage identifies this as promising for avoiding unfavourable qualities such as barren plateaus.
- Sampling sensitivity: The figure plots normalised effective-dimension sensitivity to the numbers of data and parameter samples used to estimate empirical Fisher information.Those estimates are then used to calculate effective dimension.
C.3 Training the models using a simulator
The models are trained repeatedly on Iris under matched simulator settings, with training loss, Fisher-Rao norm, and Fisher-information spectra reported across model types.
- The simulator experiment uses four input features, two outputs, and eight trainable parameters for each model.The models are trained on 100 points from the first two Iris classes.
- 100 trials and 100 training iterations are used to evaluate average training loss and average Fisher-Rao norm for all three models.The evaluation uses the Iris dataset with standard optimisation settings.
- The initial parameters are sampled from Θ = [−1, 1]^d, a choice that affects the observed effective dimension.This parameter space is used for both initialisation and effective-dimension sampling.
- Figure 9 compares average Fisher information spectra across three model types as input size increases, using rows with sin = 6, 8, and 10 and corresponding d = 60, 80, and 100.All configurations use sout = 2.
C.4 Training the quantum neural network on real hardware
The quantum neural network is trained on a four-qubit real device using a hardware-aware circuit and the Iris training setup. The circuit combines data-dependent rotations, entanglement, and eight trainable variational parameters.
- The hardware experiment trains a four-qubit quantum neural network on the first two Iris classes using the ibmq montreal 27-qubit device.The device connectivity is restricted to a linear arrangement.
- Training stops after roughly 33 steps once the loss stabilises, meaning its iteration-to-iteration change becomes small.The same training specifications as Appendix C.3 are used with random parameter initialisation.
- The hardware circuit uses data-dependent RZ and RZZ rotations, eight trainable RY-gate parameters, and a different entanglement structure to reduce CNOT requirements.The design is adjusted to incorporate current hardware constraints.
- The circuit applies Hadamard gates, encodes features and feature products through rotations, repeats the feature map to depth 2, then applies parameterised RY gates and linear entanglement.The full circuit repeats feature-map encoding once before applying the variational form.