Quantum Computers Could Redesign AI's Core Learning Bias

Quantum computers have long been touted for their potential in cryptography and simulating quantum systems, but finding other practical applications has proven challenging. Now, researchers propose a compelling new direction: using quantum computers to fundamentally reshape how machine learning models learn from data. The key idea centers on spectral s—techniques that design models by manipulating their properties in Fourier space, which relates to concepts like smoothness and simplicity crucial for generalization. This connection could make quantum computers uniquely suited to implement these s more efficiently than classical approaches.

At the heart of this proposal is the observation that spectral s are fundamental to machine learning, yet often computationally prohibitive on classical hardware. For example, smooth models, which are robust to input perturbations and generalize well, have Fourier spectra that decay super-polynomially—meaning their high-frequency components are small. Designing such decay directly in Fourier space is typically expensive, but quantum computers, through tools like the Quantum Fourier Transform (QFT), can manipulate the Fourier spectrum of a quantum state representing a model efficiently. This could allow for more direct imposition of simplicity biases, unlike the indirect heuristics used in classical s like neural networks.

The researchers illustrate this with a toy example involving generative modeling over binary data. They start with an empirical distribution from training samples, compute its Fourier coefficients using the Walsh-Hadamard transform (a Fourier transform for the group Zn2), apply a low-pass filter to suppress high-order coefficients (associated with noise from finite data), and transform back to direct space to create a smoothed model. This process, which enforces smoothness by decaying higher-order parities, can be implemented on a quantum computer by encoding the data into a quantum state, applying the QFT, manipulating amplitudes in Fourier space, and measuring. However, the paper notes that quantum s manipulate the Fourier spectrum of amplitudes, not probabilities, leading to different effects due to the Born rule, and that some filters may be dequantisable by classical kernel s.

The paper details how spectral s relate to various data types through group theory. For continuous features on RN, smoothness corresponds to a decaying Fourier spectrum. For binary features on Zn2, Fourier coefficients represent expected parities or interaction effects, with low-order coefficients capturing simple correlations. For permutations on the symmetric group Sn, Fourier coefficients relate to patterns like object-position assignments, with low-order coefficients indicating simpler correlations. The researchers argue that these connections show Fourier spectra naturally encode simplicity, making them a prime target for regularization. They also highlight that classical machine learning already uses spectral s implicitly through kernels and convolution, such as in support vector machines, convolutional neural networks, and the neural tangent kernel, which underlies the spectral bias in deep learning.

Quantum computers offer natural advantages for spectral s due to efficient QFTs for many groups, including all finite Abelian groups and some non-Abelian ones like the symmetric group. The QFT allows manipulation of the Fourier spectrum of a quantum state's amplitudes, which relates to the spectrum of a generative model via an autoconvolution formula. Beyond the QFT, quantum neural networks exhibit spectral biases through data encoding strategies, with bandlimited Fourier spectra arising from eigenvalue differences of encoding operators. The paper also reviews existing quantum machine learning research that leverages spectral s, such as probabilistic modeling over permutations using QFTs for inference tasks and resource theories using generalized Fourier analysis to study entanglement.

Despite the potential, significant s remain. Quantum computers impose constraints, such as the inability to directly compute Fourier coefficients—only to manipulate or sample them—and the difficulty of training parametrised circuits. The paper cautions that some spectral manipulations may be classically simulable via kernel s, and that the utility of quantum models will likely depend on accessing medium-order Fourier spectra at the edge of classical tractability. The researchers hope this perspective stimulates further work to explore whether quantum computers can enable more conscious design of simplicity biases without the massive scale of modern neural networks, potentially opening new avenues in quantum machine learning.